Obsidian MCP Server

A memory system for your AI agents — stored as plain markdown you can open in Obsidian.

A self-hosted Model Context Protocol server that gives every agent you connect a durable, shared place to remember things. The storage isn't a vector database you can't see into: it's a folder of markdown files in your Obsidian vault, backed by full-text and semantic search and by your own wikilink graph. Obsidian is the human window onto it — open a note, read exactly what an agent wrote about you, correct it, delete it, or take the whole folder somewhere else. Self-describing, too — agents read what you read, link what you link, and pick up your folder layout, frontmatter schema, and tag conventions on the first call instead of being briefed from scratch every session.

To be precise about the scope: what the server supplies is MCP-accessible storage, keyword and semantic search, and graph operations over markdown notes, for whatever MCP clients you connect. The agents direct their own reads and writes. There is no automatic extraction, consolidation, or decay pipeline running behind them — an agent remembers something because it wrote a note, and forgets it because someone deleted one.

Stack: Python 3.12, FastAPI, PostgreSQL with pgvector. Pluggable embeddings (Ollama bge-m3, or OpenAI text-embedding-3-{small,large}).

Contents

• Why this exists
• A session at the keyboard
• A session away from the keyboard
• What's in the box
• vs. other Obsidian MCP servers
  • vs. hosted memory systems
  • vs. an agent with raw file access
• Who this is for
• Control panel
• Quick start
• Cost expectations
• The self-describing vault
• Multi-user mode
• Configuration
• Architecture
• Project layout
• Development
• Security notes

Why this exists

There are three things going on here, and they're more interesting together than apart.

1. Agent memory that you can actually read

If you let an agent run for a while, it needs memory. Most setups solve this with an opaque vector store, a SQLite blob, or a managed "memory" service that you can't see into. That works until you want to know what the agent thinks it knows about you, or you need to correct something, or you want to understand why it just made a weird suggestion.

This server gives you a different deal. Agent memory lives as markdown files in your vault. Folder structure, file names, frontmatter, all visible. You can open the file in Obsidian and read it. You can edit it. You can delete it. You can grep it. The agent's "memory" is a human-auditable artifact that sits in the same place as your own notes, with the same tools available.

The home lab is the use case that sold me on this. My vault has notes on the rack, the network, and every Home Assistant integration. I can say "set up a night-light mode in the master bathroom, 1% after 11pm" and a sysadmin agent finds the right config, makes the change, and updates the doc in the same pass. Six months later when I've forgotten how it works, the answer is in the vault, not buried in some chat history I can't search.

The semantic search and wikilink graph still work over that material, so retrieval is fast and conceptual. But the substrate is files you own, not a black box.

2. A shared memory layer between you and your agents

The other half runs the other way: the vault isn't only the agents' memory, it's mine. I think of my Obsidian vault as my exocortex. The "big me" that includes notes, calendars, scripts, search, and AI assistants is substantially more capable than the "small me" of the biological brain alone. It's also where I do most of my thinking, because writing something down is itself a form of thought.

The problem is that until recently, the vault was passive. I had to go find things. Agents that wanted to help me had to be briefed from scratch every session, and they had no way to see what I'd already written about a topic.

This server fixes that. Now the same vault feeds my own daily writing and any agent I plug into it. The agent reads what I read, links what I link, follows the same wikilinks, sees the same frontmatter. When I write a project note on Sunday, my Monday-morning briefing agent already knows about it. When the agent leaves notes from a research session, they show up in my normal Obsidian search.

A concrete version of this: I'll spend a session in Claude Code on a project, wrap up, push the commits, and then just say "update Obsidian." The agent reads the vault guide, figures out where project notes live in my structure, picks the right format and frontmatter, and leaves a session log I can later roll into a status report. No path-passing, no telling it what to write — the conventions are already in the vault, and it follows them.

That's the exocortex idea made concrete: one place that holds context, and both the human and the agents reading and writing into it on the same terms.

3. The vault follows you

The thing that still surprises me is that this is internet-facing. Same vault, same notes, reachable from anywhere I happen to be talking to an agent — Claude in a browser tab, Claude on my phone on the train, an n8n workflow firing on a schedule, a Claude Code session on whatever laptop is in front of me. All behind OAuth or a bearer key, all hitting the same source of truth.

So when I hear an interesting podcast on a commute and want to think out loud about it with Claude on my phone, the agent isn't starting cold. It can pull up what I've already written on adjacent topics, surface a note I half-finished six months ago, and at the end of the conversation suggest updates and write them in. The vault doesn't have to be near me to be the thing I'm working in. (A session away from the keyboard walks through exactly one of these.)

A session at the keyboard

To make the abstract concrete, a short transcript from a real session. Tool calls inline; the user types in plain English and the agent picks the right tool.

you  > find anything I've written about why retrieval-augmented
       chat is hard

claude
  → semantic_search(query="why retrieval-augmented chat is hard")
    • 2025-12-03 - RAG failure modes.md
    • 2025-10-18 - Chunking tradeoffs.md
    • 2024-06-22 - Vector recall vs precision.md

  the closest hit is your December note on RAG failure modes. let
  me see what it's connected to.

  → get_neighborhood(path="2025-12-03 - RAG failure modes.md",
                     depth=1)
    8 connected notes. notable ones:
    • "Hybrid search" — a re-ranking design you sketched but
      didn't ship
    • "Out-of-distribution queries" — a half-finished draft from
      October
    • "Embedding model choice" — links to bge-m3, your current
      model

  you have three connected pieces here: a failure-modes catalog,
  an unshipped hybrid-search sketch, and an abandoned O-O-D draft.
  want me to summarize all three together, or pull up the draft
  so you can decide whether to finish it?

you  > pull up the draft, and tag it #revisit-2026 in frontmatter
       while you're there

claude
  → read_note(path="2025-10-18 - OOD queries (draft).md")
  → set_frontmatter(
      path="2025-10-18 - OOD queries (draft).md",
      updates={"tags": ["draft", "rag", "revisit-2026"]}
    )

  here's the draft, and the tag is on. you stopped right before
  the section on confidence thresholds; the open question you
  left yourself was…

Two things to notice. First, the agent didn't need to be told what folder to look in or what tools to use — it picked them. Second, the write at the end is structured (set_frontmatter mutating YAML, not a regex over the file body), so the note round-trips cleanly. The self-describing vault and the wikilink graph are doing the work that makes this feel natural.

A session away from the keyboard

The transcript above is the easy case: I'm at a desk, I can see what the agent is doing, and Obsidian is one alt-tab away. The session that actually changed how I think about this server had none of that.

I was out walking with a health podcast in my ears — a long one, two people who clearly disagreed with each other, an hour of it. I had my phone and no intention of going home to a laptop. So I pulled the episode's transcript, handed it to Claude on my phone, and we talked it through while I kept walking: what the actual claim was, which parts I already had notes on, where it cut against something I'd decided months ago and written down at the time.

The agent had the vault the whole way. It surfaced what I'd already written on the topic, flagged that two dates in an older note were wrong, and asked whether a decision I'd recorded last year still stood given what the episode argued. By the time I got back it had written all of it in: the health-related decisions I'd actually landed on during the walk, the date corrections in the old note, a couple of new notes on the episode itself — and, because the conversation kept circling back to it, a durable note on how I decide which experts to trust on medical questions in the first place. That last one is the artifact I keep returning to. It wasn't about the episode at all; it was the reasoning underneath a whole class of decisions, and it now sits in the vault where the next agent will find it.

I never opened Obsidian. Not on the walk, not when I got home. The whole session — retrieval, argument, correction, and the writing that came out of it — went through an agent, and the vault is simply where it landed. Obsidian is how I check the work afterwards, not how the work gets done. That inversion is most of the reason this project looks the way it does.

What's in the box

The server exposes 25 MCP tools across five families, plus the auth and ops layer around them.

Search and discovery

• keyword_search(query, folder?, tags?, frontmatter?, limit=20), full-text via PostgreSQL tsvector; the text-search config(s) are configurable via FTS_CONFIGS (see Full-text search language(s))
• semantic_search(query, folder?, tags?, frontmatter?, limit=15), vector similarity via pgvector, one preview chunk per note
• list_notes(folder?, limit=50), sorted by modified time
• get_recent(folder?, limit=20), recently changed
• get_tags(limit=50), tag and count
• get_vault_guide(), the Obsidian primer plus this vault's CLAUDE.md, served live

Read and write

• read_note(path, section?, offset=0, limit?) returns a structured result — path, title, tags, frontmatter_yaml and a JSON frontmatter view, heading (section reads), content, and truncation as data (truncated, offset, next_offset, total_chars, outline, notice). Bounded by MAX_READ_RESPONSE_CHARS (default 40,000) — see Response size limits. section=<heading> returns one section's body instead of the whole note; offset continues a truncated read.
• create_note(path, content), atomic write, refuses overwrite
• edit_note(path, …) with four mutually exclusive modes: full replace (default), append=True, find=… (with optional replace_all), or section=<heading> (ATX headings, supports Parent/Child path-style and #N ordinal disambiguation). dry_run=True returns a unified diff without writing. Legacy clients may use operation="append"; operation="replace" explicitly selects full replace.
• move_note(from_path, to_path, rewrite_links=False), relocates and optionally rewrites incoming [[Old]], [[Old|alias]], [[Old#anchor]], ![[Old]], and [[folder/Old]] references in source notes
• delete_note(path, permanent=False), soft-delete to .trash/<YYYYMMDD-HHMMSS>-<basename>-<8 hex> by default, via a single non-replacing rename, so it never overwrites an existing trash entry (a filesystem that cannot do that rename makes the soft delete refuse with a named error rather than fall back). permanent=True unlinks.
• set_frontmatter(path, updates, remove?), structured YAML mutation. Body is byte-identical when only frontmatter changes.

File access (non-markdown)

Raw read/write/browse of arbitrary vault files (PDFs, images, skill assets, data files) — distinct peers to the note tools, which stay markdown-only. Pure byte transport: no server-side PDF/text extraction, no embedding or indexing of non-markdown files.

• read_file(path, encoding="auto", offset=0, limit?), returns text-like files as text, images as an inline image block that renders in-client, and other binaries as a base64 string. text/base64 force the form. Refuses files over MAX_FILE_READ_BYTES (default 10 MB); text results are additionally bounded by MAX_READ_RESPONSE_CHARS and continue via offset. hash_only=True returns the whole file's content_hash without content; base64 results include that hash in their header. Text results remain plain text.
• write_file(path, content, encoding="base64", overwrite=False), lands a file in the vault; base64 for binary, text for UTF-8. No-clobber by default, auto-creates parent dirs, atomic write. Capped at MAX_FILE_WRITE_BYTES (default 25 MB).
• list_files(folder=".", pattern="*", recursive=False, limit=200), ls-style browse of files and subdirectories with size and mtime, glob-filterable and result-capped.
• delete_file(path, permanent=False), soft-deletes a non-markdown file to .trash/<YYYYMMDD-HHMMSS>-<basename>-<8 hex> with a single atomic rename. Refuses markdown (that is delete_note), directories, and symlinks.

All four reuse the path-traversal guard and exclude any path with a component starting with . (dot-directories and dot-files) (.obsidian, .git, .trash, …), matching the indexer's visibility rule.

Guarding edits against stale reads

Pass a read's content_hash as expected_hash when editing, updating frontmatter, moving or deleting a note, overwriting a raw file, or deleting a raw file. The canonical token is sha256:<64 lowercase hex>, computed over the complete raw file bytes; a section or truncated read_note still returns the whole-file hash. For raw files, use read_file(hash_only=True) or the base64 header. Do not hash the returned text yourself.

A stale token refuses the operation before mutation, with a final MCP-REFUSAL JSON line naming stale_precondition and the current hash. Re-read and reconsider the edit before retrying. Moves bind the source note only; moves and deletes still allow an in-place edit after their preflight comparison. Overwrites retain their separate in-call byte check. Successful publishing writes report a new hash when available.

The argument is optional by default. WRITE_PRECONDITION_REQUIRED=true requires it on the supported destructive calls; enable this only after clients supply tokens. Creation is exempt and refuses a supplied token as no_incumbent. Files above their read cap cannot be guarded.

File transfer

No MCP client can hand a tool the bytes of a file the user is looking at, so write_file is only usable when the agent already has the content. These tools close that gap with short-lived capability links, redeemed over the public /transfer/* routes.

• request_upload(path, overwrite=False, expires_in?), mints a single-use link bound to exactly one destination path. The human opens it, picks a file, and it lands at path — nothing else can be written with it.
• check_upload(upload_id), reports pending / uploading / completed (with path, size, sha256 and MIME) / unknown (a stream started and the server never recorded how it ended — read the path before re-minting) / revoked (the credential or vault root changed under the link) / expired, scoped to the identity that minted it.
• request_download(path, expires_in?), mints a link the human can save one vault file from. Usable more than once until it expires, and bound to the file's exact bytes at mint time.
• import_from_url(url, path, overwrite=False), fetches a public https asset straight into the vault under an explicit outbound deny policy (no private, loopback, link-local, metadata or tunnelled addresses, in any spelling, re-checked at every redirect).

The token travels in the URL fragment, which browsers never send, so no server-generated request target or access log contains it. Uploads are claimed before a body byte is read, published atomically with no-clobber semantics, and bound at mint time to the file state they were minted against — a link cannot silently undo an edit made while it was waiting. MCP_HOSTNAME or BASE_URL must be set; without a public origin the mint tools refuse rather than emit a localhost link.

Wikilink graph

• get_backlinks(path, limit=50), notes linking TO path
• get_links(path), outgoing links, both resolved and dangling
• get_neighborhood(path, depth=1, limit=50), undirected BFS over the resolved-link graph, capped at depth ≤ 5 and limit ≤ 200
• find_related(path, limit=10), semantic neighbors via averaged chunk embeddings and pgvector cosine distance, deduped per note
• find_orphans(folder?, limit=50), notes with zero in or out resolved links

Auth and ops

• API keys with the omcp_ prefix, stored as SHA-256 hashes, with read and readwrite permission scopes. Write tools refuse on read-only keys.
• OAuth 2.0 PKCE (S256) flow for public and confidential clients, including ChatGPT, Claude Desktop, and claude.ai. Dynamic registration defaults to both vault permission levels; the user chooses the actual grant on the consent screen.
• Control panel (Jinja2, htmx, Tailwind) for keys, usage logs, indexer status, embedding-provider info, and a danger-zone reset.
• Every tool call is logged to usage_logs with name, params (truncated to 200 chars), duration, response size, and the calling credential's name — recorded at call time, so the audit trail survives deleting the key or OAuth client it describes.
• /health is unauthenticated and returns status plus two capability fields: transfer_mount_check_available (the kernel supports the mount check transfer writes need) and vault_named_staging_fallback_active (a write has actually staged under a name on this process).

Every write — note tools, write_file, uploads and imports — stages the new bytes in a temporary inode, fsyncs them, and only then publishes. Creation publishes with a kernel-atomic hard link that refuses to clobber; move_note and the soft delete publish with a single non-replacing rename; an overwrite is a same-directory rename onto the destination. The destination directory (and any directory the call created) is fsynced afterwards, so a crash mid-write can neither truncate a note nor lose one the server reported as written.

Staging happens in an unnamed inode wherever the filesystem supports one, so no temporary name is ever visible in the vault. On a mount that refuses that (some NFS exports do), those writes refuse with an error naming VAULT_ALLOW_NAMED_STAGING_FALLBACK; setting that flag takes named staging back on both write paths as a declared, weaker guarantee. See System requirements.

vs. other Obsidian MCP servers

There are several existing MCP servers for Obsidian, and most of them solve a different problem than this one. The lightweight ones are glue over Obsidian's Local REST API plugin or the filesystem: they let an agent reach the files, but don't build any infrastructure of their own. They're great if "I just want Claude to read my notes" is the goal and you keep Obsidian running locally.

This server is on the other end of the spectrum: a real backend with a persistent index, semantic retrieval, a wikilink graph, OAuth, and an admin UI. The cost is Postgres and Docker. The benefit is everything you can build on top of that.

Comparison reflects each project's documented features at time of writing; verify the specifics before betting on them.

vs. hosted memory systems

The comparison that matters more, now that most of my vault traffic is agents rather than me, is against memory as a service: your agent calls an API, the service stores what it's told, and it hands back what it judges relevant later. mem0, Zep and Letta are the names people usually reach for. What follows is about that architecture — memory behind a service boundary — not about any one product's current feature list, which moves faster than a README can track.

The difference is where the memory lives and who can open it.

• Readability. When memory sits behind a service API, reading it means whatever endpoint or console the service exposes, in whatever shape it stores. Here the memory is the artifact: Health/2026-08 - Trusting expertise.md, in a folder, in your editor, in grep. There's no gap between what the agent stored and what you can look at.
• Shared with you, and between agents. A memory service is generally scoped to an application and its users; the human's own writing is a different system. Here it's one corpus. I write into it by hand, and every connected client — Claude Desktop, Claude Code, Claude on the phone, an n8n workflow — reads and writes the same files on the same terms. A note I type on Sunday is context for an agent on Monday with no import step.
• Portability. The exit path from a folder of markdown is cp -r. No export format, no migration script, no question about what you'd be left holding if a project stopped being maintained. That's a property of files, not something this server does for you.
• Self-description. The rules live in the corpus rather than in client config. CLAUDE.md at the vault root tells every agent, on its first call, where things go and what frontmatter they carry, so conventions are versioned next to the notes they govern.

What the hosted shape buys you in exchange is real, and worth saying plainly. There's no Postgres to run, no pgvector version to keep current, no container to babysit — you get a memory layer by adding a dependency, which is a genuinely better trade for most people. And systems in that class typically do work this server deliberately doesn't attempt: pulling facts out of a conversation automatically, reconciling ones that contradict each other, and scoring relevance or decaying old memories so they stop crowding out new ones. Here an agent remembers something because it decided to write a note, and the judgment about what's worth keeping is the agent's, not the server's. If you want memory that curates itself, that's a fair reason to pick the other shape.

vs. an agent with raw file access

The other baseline isn't an MCP server at all: point Claude Code, a generic filesystem MCP, or any agent with file tools straight at the vault folder. That works — until a write goes wrong. An agent rewriting a whole file from its memory of an earlier read will eventually clobber a note, follow a symlink somewhere it shouldn't, or "tidy up" your .obsidian config. Nothing in a raw file API pushes back. This server's write path is shaped by exactly that kind of incident, and it assumes the caller will eventually do something wrong:

• Targeted edits instead of rewrites. edit_note can address a find-string or a single section rather than replacing the file, and dry_run=True returns the unified diff before anything lands. set_frontmatter mutates YAML structurally and leaves the body byte-identical.
• No-clobber defaults. create_note and write_file refuse to overwrite an existing file; replacing one is an explicit opt-in.
• Atomic writes. Content is staged and renamed into place against a descriptor opened at validation time — a note is never left half-written, and the file that gets replaced is the file that was checked.
• Reversible deletes. delete_note and delete_file soft-delete into .trash/ with a non-replacing rename; permanent=True is the explicit escape hatch, not the default.
• Kernel-proved containment. Paths resolve under the vault root via openat2(RESOLVE_BENEATH | RESOLVE_NO_SYMLINKS | RESOLVE_NO_MAGICLINKS), writes refuse a symlink as the final component, and dot-directories (.obsidian, .git, .trash) are out of reach of every tool.
• Bounded responses. Reads are capped and truncation is data (truncated, next_offset, an outline) rather than silent loss, so one huge note can't flood an agent's context into a bad edit.
• An audit trail. Every call is attributed to a key and logged; the control panel shows who touched what, and when.

When an agent misbehaves through this server you get a refused call, a diff, a trash entry, and a usage-log line. When it misbehaves with raw file access you get whatever git diff can recover — if the vault was in git at all.

Who this is for

• Homelab folks who already run Postgres and Docker, or are happy to spin them up. The setup tax is the price of admission for the semantic and graph layers.
• People who keep an opinionated vault — task placement logic, frontmatter schemas, tag taxonomy — and want agents to follow those conventions on the first call instead of being briefed every session.
• Anyone running more than one MCP client (Claude Desktop, Claude Code, Claude in a browser, n8n) against the same notes and tired of re-explaining the vault to each.
• Folks who want agent memory to live as plain markdown files they can read, edit, grep, and version-control, not in an opaque vector store or a managed memory service.

Who this isn't for

• "I just want Claude to read my notes" with the lightest possible setup. Use one of the filesystem-glue projects above; you don't need this.
• Anyone unwilling to run a database. There is no SQLite fallback; pgvector is doing real work, and a managed Postgres with pgvector support is part of the stack.
• People who want a turnkey hosted product. This is a self-hosted server you run yourself.

Control panel

The server ships with a built-in admin UI for the parts of operations that are easier to look at than to query: minting keys, watching the indexer, eyeballing tool-call traffic, and resetting embeddings when you switch providers.

Usage

Per-tool-call audit log with a 14-day request histogram. Every MCP call is recorded with the calling key, tool name, duration, and response size — useful for noticing a misbehaving agent burning tokens on something it shouldn't.

API keys and OAuth clients

Bearer keys with read / readwrite scopes for API clients, and a separate OAuth 2.0 PKCE flow for clients like ChatGPT, Claude Desktop, and claude.ai that expect a proper authorization-code dance. The OAuth server supports public (none) and confidential (client_secret_post) token-endpoint authentication plus refresh tokens.

Each client's page lists its grants — one row per /authorize approval, not per token — with a Revoke control and a permission select per grant, so revoking really ends the session instead of leaving a refresh token to mint a replacement. Revoked and expired rows stay listed, dimmed, for a week.

Vault browser

A read-only file tree of the mounted vault, mostly for sanity-checking that the container sees what you think it sees.

Settings

Indexer status, current embedding provider and model, vault path, and the danger zone: Reset embeddings (drops and recreates the embeddings column at the configured dimension — use it when switching providers) and Force re-embed (keeps the column, clears every note's embedded-content hash so the next pass re-embeds the vault). Both pause the indexer while they run.

The dashboard separates two things that used to be conflated: Last run is the indexer's own heartbeat — the last pass that completed, whether or not anything had changed — and Last change detected is the newest indexed_at on any note. A quiet vault makes the second one old while the indexer is perfectly healthy.

Quick start

Deploying on a VPS from scratch? See DEPLOYMENT.md for the full walkthrough: Postgres setup, Caddy and TLS, vault sync via Nextcloud, and the gotchas that bite first-time deploys.

The bundled Caddy configuration fails closed on /admin, /api, and /authorize; replace its placeholder basic-auth hash before starting it.

Prerequisites

• Docker and Docker Compose
• A PostgreSQL 16 instance reachable from the container, with pgvector 0.8.0 or newer installed
• Either an Ollama instance running bge-m3, or an OpenAI API key. Anything that speaks the OpenAI embeddings protocol works (Azure OpenAI, OpenRouter, Together, etc.).
• Linux, kernel 5.6 or newer (see below)

System requirements

The server checks these at startup and tells you which one failed rather than misbehaving later.

Linux kernel ≥ 5.6. Every directory below the vault root is opened with a single openat2(RESOLVE_BENEATH | RESOLVE_NO_SYMLINKS | RESOLVE_NO_MAGICLINKS), which is what makes the kernel — not the application — prove that a write stayed inside the vault. There is no fallback: on an older kernel, or under a container seccomp profile that blocks openat2, the server logs the reason and exits non-zero.

Kernel ≥ 5.8 for file transfer. statx()'s STATX_MNT_ID is how a publication refuses a destination that sits on a different mount than the staging directory (a nested bind mount under the vault root would otherwise fail only after a whole upload body had streamed). Below 5.8 the server logs one warning and starts: request_upload, import_from_url and PUT /transfer/upload refuse, and everything else — reads, note writes, search, downloads, the panel, OAuth — is unaffected. /health reports it as transfer_mount_check_available.

pgvector ≥ 0.8.0. Filtered semantic search needs hnsw.iterative_scan, which landed in 0.8.0. An older extension accepts the setting as an unknown placeholder and silently runs a plan that drops post-filter candidates — silently worse search results — so the server exits instead. Fix with ALTER EXTENSION vector UPDATE or a newer database image.

Filesystem. Case-sensitive and non-normalising (ext4, xfs, and the usual bind mounts). It must support hard links within the vault root and renameat2(RENAME_NOREPLACE); without those, note creation, move_note and the soft delete refuse with a named error rather than degrading to a publish that can clobber. O_TMPFILE is wanted but optional: where it is unavailable, set VAULT_ALLOW_NAMED_STAGING_FALLBACK=true to accept named staging instead (see Configuration). macOS and Windows hosts are out of scope; run the container on a Linux VM.

1. Clone, configure, point at your vault

git clone https://github.com/maxkuminov/obsidian-mcp.git
cd obsidian-mcp
cp .env.example .env
$EDITOR .env

In docker-compose.yml, point the /obsidian volume at your vault:

volumes:
  - /path/to/your/vault:/obsidian

2. Pick an embedding backend

Option A, OpenAI (zero local infra):

EMBEDDING_PROVIDER=openai
OPENAI_API_KEY=sk-...
EMBEDDING_DIMENSIONS=1024
OPENAI_EMBEDDING_MODEL=text-embedding-3-small

The server validates OPENAI_API_KEY at startup and refuses to boot if it's missing.

Option B, Ollama (self-hosted, GPU recommended):

EMBEDDING_PROVIDER=ollama
OLLAMA_URL=http://your-ollama-host:11434
EMBEDDING_MODEL=bge-m3
EMBEDDING_DIMENSIONS=1024

This is the default. Omitting EMBEDDING_PROVIDER falls back to Ollama.

3. Deploy

make init       # data dirs and .env from template (skip if you've already edited)
make db-init    # create database, user, and pgvector extension
make deploy     # build, push to local registry, run migrations, recreate container

The first deploy backfills the index, the wikilink graph, and the embeddings. For a 2 to 3k-note vault on Ollama with a GPU this takes a few minutes. On text-embedding-3-small it's seconds.

4. Connect a client

Mint an API key in the control panel, then point your MCP client at:

URL:  https://obsidian-mcp.<your-domain>/mcp
Auth: Bearer omcp_...

For Claude Desktop, add to claude_desktop_config.json:

{
  "mcpServers": {
    "obsidian": {
      "url": "https://obsidian-mcp.<your-domain>/mcp",
      "headers": { "Authorization": "Bearer omcp_..." }
    }
  }
}

For Claude Code:

claude mcp add obsidian --transport http \
  --url "https://obsidian-mcp.<your-domain>/mcp" \
  --header "Authorization: Bearer omcp_..."

The first thing any agent should do in a new session is call get_vault_guide(). That's how it learns your folder structure, naming conventions, and YAML schema before it writes anything.

Cost expectations

If you go the OpenAI route (the realistic path on a CPU-only VPS), the first-index spend is small and the steady state is nearly free. Rough numbers assuming an average note around 1,500 tokens (three 512-token chunks), at OpenAI's published rate at time of writing:

After the first index, only changed notes are re-embedded. Ongoing cost is proportional to edits — pennies a month for a typical vault.

If you self-host Ollama with a GPU, embedding cost is whatever your power bill is. Ollama on CPU works but is too slow to be usable on a vault of more than a few hundred notes.

The self-describing vault

This is the part most "MCP for Obsidian" projects miss. They stop at read, write, and list. The interesting question isn't "can the agent reach the files," it's "does the agent know the rules?"

If you have an opinionated vault — task placement logic, folder conventions, required frontmatter, tag taxonomy — an agent with write access can do real damage without that context. Tasks land in the wrong folder. Bare-date filenames collide with templates. Wrong tags break Dataview queries. The data layer works fine; the context layer is where the failures show up.

The fix is small. Keep a machine-readable instruction file (CLAUDE.md at the vault root) that describes the system's own rules. Expose it as a dedicated tool. Every connecting agent calls it once at the start of a session and immediately knows how the vault works. Update the file, every agent sees the change on the next call. No client-side config. No system-prompt injection. The vault is authoritative about its own rules.

get_vault_guide() does exactly this. It returns a generic Obsidian primer (wikilink syntax, embed syntax, tag conventions, common plugin literals) plus the vault's CLAUDE.md live. The hint to call it first is baked into the write-tool descriptions so the agent gets pulled into the right behavior even without prompting.

Multi-user mode

Single-user mode is the default and works exactly as described above — one vault, one set of API keys, no in-app user concept. Multi-user mode is an opt-in flag that turns the same container into a small multi-tenant deployment: in-app username/password login, per-user vault scoping, an admin role for troubleshooting, and a regular-user role that sees only its own keys/OAuth clients/usage. One container, one Postgres, strict isolation between users.

Enable it on an existing deployment with no data loss — your current vault and keys carry over to the bootstrap admin.

Enabling

1. Set MULTI_USER_MODE=true and a strong SECRET_KEY in .env (openssl rand -hex 32 is fine). The app refuses to start with a placeholder SECRET_KEY unconditionally — single-user mode included — so this is not something the flag turns on.
2. make deploy (or docker compose up -d --force-recreate).
3. Visit the panel. Because the users table is empty, you're routed to /admin/register — the one-time bootstrap form. It's still behind Traefik's chain-oauth@file middleware, so only people Traefik already trusts can claim admin.
4. Register with a chosen username and password. The bootstrap form pre-fills vault_path with whatever VAULT_PATH was set to, so your existing notes immediately belong to this new admin. No re-index, no re-embed, no data loss — every previously indexed note, API key, OAuth client, and usage log row gets backfilled to the bootstrap user in a single transaction.

Inviting users

1. Edit docker-compose.yml to add a volume mount for the new user's vault under /vaults/<username>. Host paths with spaces must be quoted as a single YAML string:

   yaml

   Copy

   volumes:
     - "/storage/vaults/alice:/vaults/alice"
     - "/storage/shared/bob/Obsidian:/vaults/bob"
   

   make deploy to apply.

2. In the panel, /admin/users/create — pick a username and set an initial password.

3. /admin/users/{id}/edit — set the user's vault_path to the container path you just mounted (e.g. /vaults/bob). The form shows a dropdown of unassigned /vaults/* directories that exist on disk.

4. Share the credentials out-of-band. The user logs in at /admin/auth/login, gets their own keys/OAuth/usage views, and cannot see other users' notes.

What admins see

Admins see API keys, OAuth clients, and usage logs for all users; they own the Settings page (embedding provider, indexer trigger, danger zone) and the Users page. Admins do not browse other users' vault contents through the panel — that's intentional. Troubleshooting another user's vault means either inspecting it via docker exec or temporarily reassigning their vault_path, not UI snooping.

Rolling back

Set MULTI_USER_MODE=false, restart. Existing API keys keep working (per-user filters skip when no user context is set), the login UI and session cookies disappear, and the panel falls back to its Traefik-OAuth-only mode. The schema stays in place, so flipping back to multi-user later resumes where you left off without re-bootstrapping (the users table is non-empty, so /admin/register is closed).

Constraints and known limits

• The indexer iterates active users sequentially each cycle. Fine for tens of users; hundreds would need parallelization.
• Password recovery is admin-driven — there's no email-based reset. A signed-in user can rotate their own password at /admin/account (current password, new password, confirmation; minimum 12 characters), which signs their other browsers out and keeps the one they changed it from signed in. The admin reset stays the recovery path for somebody who cannot sign in at all, and it also ends every live session of the account it resets.
• /admin/auth/login and /admin/account/password are rate-limited at 5 requests per minute; the login limit is keyed on the client address, and the password change carries two independent limits — one per account, one per address. The limiter's storage is in-memory and per-process, so counters reset on restart. The Traefik OAuth gate in front of the panel is still the main brute-force defense; if you expose /admin/auth/login to the open internet, put a rate-limit middleware in front of it as well.
• Panel sessions are server-side rows (user_sessions), so logging out, changing a password, a deactivation or a delete really ends them. The trade-off: the first deploy of the build that introduced the registry signs every live panel session out once, because a cookie issued before it carries no session id and is refused rather than grandfathered. Everyone signs in again; nothing else changes.
• The vault_path validator does not resolve symlinks, so an admin can technically point a user at host files via a symlinked /vaults/<name>. Treat /vaults/ as an admin-trust boundary. What is checked, since the vault-root overlap guard: two active users' roots may not name overlapping directories. Each root is opened once and compared by inode identity — (st_dev, st_ino), which catches a symlink alias or a bind mount naming one directory twice — and by a component-wise containment test over the two canonical real paths in both directions, which catches an ancestor/descendant pair like /vaults/team and /vaults/team/private. A conflicting assignment is refused in the panel naming the other user, and the same checks re-run before every index pass, so an alias created after the assignment quarantines both accounts: their MCP tools, index passes and transfer redemptions are refused until an administrator corrects it, and no index rows are deleted. A root that cannot be opened at all quarantines only its own account. What is still not detected, and the consequence: a bind mount that grafts one user's vault — or any mount nested inside it — to a path inside another user's root. mount --bind /vaults/b /vaults/a/inner leaves both root inodes distinct and both canonical paths outside each other, so neither check sees it, and user A can then read, overwrite and delete every note in user B's vault through the ordinary write tools, while A's index pass files B's notes under A's account so A's searches return B's content. The same gap covers an accessible alias of a root that could not be examined: that peer keeps serving. Neither condition is reported anywhere. Both require an administrator to write a bind mount into the deploy configuration — which is why /vaults/ and the compose file's mounts are the admin-trust boundary, not just the path strings.

Configuration

See .env.example for the full set with comments. For first-index spend on OpenAI, see Cost expectations above.

The MCP transport's request-body limit is derived, not configured: max(2 × MAX_FILE_WRITE_BYTES, 6 × 10 MB) + 1 MiB, which is 61 MiB with the defaults. It has to track the write caps so that every supported write is refused by the tool — with an actionable message — rather than by the transport with a bare HTTP 413. Raise MAX_FILE_WRITE_BYTES and the transport limit follows.

Switching providers or models

Different models produce vectors in different spaces, and cosine distance between two spaces is meaningless. So any change to what produced the stored vectors requires a full re-embed — not only a provider switch. That is every one of:

• EMBEDDING_PROVIDER
• EMBEDDING_MODEL (Ollama) or OPENAI_EMBEDDING_MODEL (OpenAI) — including a swap between two models of the same dimension, which the dimension guard cannot see
• EMBEDDING_DIMENSIONS
• CHUNK_SIZE and CHUNK_OVERLAP

The server stores a fingerprint of that configuration and compares it at startup. On a mismatch it logs both fingerprints and the fields that differ, names the repair, and exits non-zero — so a model swap that used to mix two vector spaces in one column silently, for ever, now stops the process instead.

The steps, in this order:

1. Update .env.
2. make deploy (or docker compose up -d --force-recreate). The new container will refuse to start — at the fingerprint guard, or at the dimension guard if the width changed — and that refusal is the point: a container that will not start embeds nothing while the reset runs.
3. make reset-embeddings while it is down. The target is docker compose run --rm, so it starts a one-off container that reads your edited .env: it recreates the column at the new dimension, clears every embedded_content_hash, and records the new fingerprint in the same transaction.
4. Restart the service. It starts silently, because the stored rows really were produced under the configuration it is now running, and the next indexer pass re-embeds the vault.

This inverts the older reset-before-recreate advice. That ordering was safe only while nothing depended on a stored claim about the configuration; now the reset is what writes that claim, so it has to run with the new .env in place and with no old-configuration container able to embed against it. Skipping a step costs time rather than correctness — a database-level generation lock makes an old-configuration container's certifications refuse rather than land — but the ordering above is the one that never has to rely on it.

Maintenance waits for an in-flight index pass. That same generation lock is taken at the head of the index pass's transaction and held until it commits, so make reset-embeddings and make rebuild-tsvectors block until the pass finishes — up to a few minutes on a large vault — rather than interleaving with it. That wait is the required behaviour, not a stall to work around: a reset that landed mid-pass is precisely the interleaving that stores vectors from one configuration under a fingerprint naming another. Neither command sets a short lock timeout, and neither should be given one — and because the server sets a 60-second statement_timeout on every connection, both commands (and the panel's Danger-zone resets) lift that timeout for the acquisition itself and restore it once the lock is theirs. Without that, a command started against a live service was cancelled after a minute rather than waiting, which reads as a broken command instead of a busy index.

You can also use Settings → Danger zone → Reset embeddings in the control panel, which performs the same SQL — including the fingerprint record — while the server is running (pauses the indexer, runs the SQL, resumes).

The fingerprint records the configuration, not the model artifact. bge-m3 is a mutable Ollama tag, so ollama pull can replace the weights behind it, and OLLAMA_URL / OPENAI_BASE_URL are deliberately excluded from the fingerprint — repointing at another host or proxy is usually an infrastructure move that serves the identical artifact, and including it would demand a full re-embed for one. The consequence is an accepted limitation: replacing the artifact behind an unchanged model name — re-pulling a tag, or pointing at a host serving different weights under the same name — mixes vector spaces undetected. It requires make reset-embeddings, and no startup check will catch it if you skip that. No value available to the server distinguishes the two cases, and a probe would have to trust the endpoint it is checking.

Full-text search language(s)

keyword_search runs over a PostgreSQL tsvector. The text-search configuration it uses — the stemmer and stop-word dictionary — is controlled by FTS_CONFIGS. It defaults to english, which reproduces the historical behavior exactly, so existing deployments need no action.

FTS_CONFIGS is a list, settable as JSON (FTS_CONFIGS=["simple","norwegian"]) or comma-separated (FTS_CONFIGS=simple,norwegian). Each note is indexed under every listed config, and a query matches if any listed config's parse hits. This is what makes a mixed-language vault work:

The setting is global — applied to every vault (consistent with EMBEDDING_MODEL, CHUNK_SIZE, etc., which are global too). For a mixed-language multi-user instance, set a superset (e.g. ["english","norwegian"], or ["simple"]). Per-user FTS config is a clean future extension but is not implemented.

A typo'd or uninstalled config name fails fast at startup with a message listing the configs available in your Postgres instance, rather than producing silent zero-result searches.

Changing FTS_CONFIGS requires a rebuild, and the server refuses to start until it has run. Stored tsvectors are computed at index time, so they go stale when the config list changes — and a stale stemmer is not merely incomplete. Under english the token running is stored as the lexeme run, so a query under simple for run matches a note that does not contain the word — a false positive, indistinguishable from a real hit. Keyword vectors therefore fail closed exactly as embeddings do: the server stores a fingerprint of FTS_CONFIGS, compares it at startup, and on a membership change logs both lists and the differing entries, names the rebuild, and exits non-zero. (Reordering the same names is not a change: a note is indexed under every config and a query matches if any hits, so order changes nothing and is not compared.)

The runbook:

1. Edit FTS_CONFIGS in .env.
2. make deploy. The new container refuses at the keyword fingerprint guard and stays down.
3. make rebuild-tsvectors. It rebuilds every scope that holds rows — every owner, including rows with no owner in single-user mode — in one transaction, and records the new fingerprint only if every one of them reported a completed rebuild. It is all-or-nothing: one scope it cannot rebuild rolls the whole thing back, names the scope and the reason, and writes no fingerprint, because the fingerprint is a single claim about every retained row.
4. Restart. It starts silently.

If step 3 names a scope it could not rebuild — a user whose vault is not assigned, a tenant still re-deriving its provenance, or ownerless rows under multi-user mode — there are three recourses, in order of preference:

• Settle the scope: assign or delete the user, or let the re-derive finish, then re-run the rebuild.
• Delete or reassign the ownerless rows, then re-run the rebuild.
• Put FTS_CONFIGS back to its previous value. That clears the refusal immediately, with no rebuild at all — a configuration edit is always reversible, which is what keeps this refusal from being an outage.

The rebuild re-reads each note and recomputes its content_tsvector under the new config(s). It rebuilds the keyword index only — it does not touch embeddings/vectors and makes no API calls, so it finishes in seconds for a few thousand notes. (Do not confuse it with the expensive make reset-embeddings flow.)

Tokenization caveat: the tsvector parser still splits on punctuation and hyphens regardless of config, so bge-m3 tokenizes to bge + m3. simple preserves word forms, not punctuation-bearing strings; exact-string-with-punctuation matching would need a trigram index and is out of scope.

Response size limits

A tool result is model input. Whatever read_note returns is fed straight back into the caller's next request, so an unbounded read is an unbounded prompt — and the caller usually finds out only when its inference provider rejects the request.

MAX_READ_RESPONSE_CHARS (default 40,000, roughly 10K tokens) bounds what read_note and the text results of read_file return. It is a different limit from MAX_FILE_READ_BYTES, which bounds what the server reads off disk. A 3 MB note is comfortably within the 10 MB read cap and will still destroy a context window; both caps are needed and they have different correct values.

It applies per component, not once to the whole response: the content window gets the cap, the heading outline gets it independently, and the metadata fields (title, tags, frontmatter_yaml and its JSON view, heading) share a third. A truncated read can carry all three, so budget for a worst case of roughly 3 × MAX_READ_RESPONSE_CHARS plus fixed prose — doubled again because the MCP result carries both structured content and a JSON text block, and multiplied by JSON escaping for content that is mostly control characters.

When a note exceeds the cap you get the first window plus truncation as data — truncated, the next_offset to continue from, total_chars — and, for a whole-note read, an outline of the note's sections:

{"entries": [
  {"ordinal": 1, "depth": 1, "text": "Client Records",
   "size": 2855343, "exceeds_cap": true,  "duplicate": false},
  {"ordinal": 2, "depth": 2, "text": "Balance Sheet.xlsx",
   "size": 391199,  "exceeds_cap": true,  "duplicate": false},
  {"ordinal": 3, "depth": 2, "text": "Lease Agreement.pdf",
   "size": 464,     "exceeds_cap": false, "duplicate": false},
  {"ordinal": 4, "depth": 2, "text": "Invoice 2025-044.pdf",
   "size": 1075,    "exceeds_cap": false, "duplicate": true}
 ], "truncated": false}

Paging a multi-megabyte note 40K at a time is technically possible and practically useless, so prefer the outline: read the one section you want with read_note(path, section="Lease Agreement.pdf"). Sections are addressable three ways — the #N ordinal shown in the outline, the Parent/Child path-style form, and exact heading text. The ordinal is the only form that separates duplicate sibling headings, which share every ancestor and so cannot be disambiguated by path; notes generated by bulk extraction tend to be full of them.

A bare #N always selects by position, so an ordinal we hand you in an outline can never be shadowed by a heading that happens to be titled #2. Such a heading stays reachable via the path form (Parent/#2) or via its own ordinal.

The outline is itself bounded by the cap: a note with thousands of headings gets a truncated listing that reports how many sections were omitted (omitted) and the full ordinal range (first_ordinal, last_ordinal), rather than an outline larger than the content window it accompanies. Metadata that does not fit its budget is dropped whole and reported in metadata_omissions — never cut short and never marked inside the field itself, so nothing in a note-controlled field is ever a prefix or server prose. frontmatter_yaml is the frontmatter block's YAML source with the fence lines removed, LF-normalized (the same declared terminator residual content carries); it is the authoritative copy, and the frontmatter JSON view beside it is a convenience that is omitted, with a reason, when YAML holds something JSON cannot say.

limit can lower the cap for a single call but never raise it. If your clients genuinely want larger reads, raise MAX_READ_RESPONSE_CHARS — that is an operator decision, made once, by someone who knows the deployment.

Upgrading: three visible contract changes.

read_note on a large note used to return the whole thing; it now truncates. The response is self-describing, so an agent needs no prior knowledge to continue, but a script that assumed whole-note reads should either pass section= or raise the cap.

And read_note used to return one rendered string — a # <title> / **Path:** header, a \n---\n separator, then the content. It now returns fields, because every component of that header was note-controlled: a note could forge the separator, so an agent recovering the section body by splitting the response could recover a crafted string and write it back over the section. A client that parsed the old envelope must read content (and, for section reads, heading) instead; clients that ignore structuredContent still get an unambiguous JSON text block.

Panel sessions are now server-side rows, so everyone is signed out once at that upgrade. A cookie issued before it carries no session identifier, and such a cookie is refused rather than grandfathered — accepting it would keep the old replay window open for another seven days after the fix shipped. Sign in again; there is nothing to migrate.

Rate limits

The consumer of this server is an agent, and a retry-storming or prompt-injected agent is an ordinary input. Three controls bound how fast one credential can create work.

• A general bucket — MCP_RATE_LIMIT_PER_MINUTE (120) sustained, MCP_RATE_LIMIT_BURST (30) capacity — on every tool call.
• A write bucket — 60/min, burst 15 — that the eight vault-mutating tools must pass in addition, and that PUT /transfer/upload consumes too, charged to the principal that minted the capability so the write rate cannot be escaped by minting links and redeeming them.
• A per-address budget on failed /mcp authentication — 60 failures per 5 minutes — checked before the credential lookup, so a refused probe costs no database query.

The bucket is per principal: an API key, or an OAuth grant. Refreshing an access token continues the same allowance rather than minting a fresh one, and two separate /authorize approvals for the same client hold independent allowances.

What an agent actually sees. A refusal is an ordinary tool result — never a protocol error, never a silent empty result set — and it ends with one machine-readable line:

Error: this credential exceeded its general rate limit of 120 calls per minute, so the call was refused before it ran. Nothing was read, written, or counted against the daily quota. Retry in 3 seconds, or slow the calling loop down.
MCP-REFUSAL {"code":"rate_limited","scope":"principal","limit":120,"limit_unit":"calls_per_minute","retry_after_seconds":3}

The MCP-REFUSAL sentinel is line-initial and the JSON is one line, so it survives being quoted into a transcript. A structured tool returns the identical text in its declared error field. retry_after_seconds is present only where waiting can actually help — a refusal for an unassigned vault or an unencodable argument omits it rather than invite a loop that cannot end. The same shape covers the daily quota (over_quota), the query length cap (argument_too_long), and tool-body refusals such as not_found, already_exists, and invalid_path. A partial write also carries a typed outcome: read its explanation before retrying, because some bytes may already have changed. Empty search results and successful no-op calls remain successes.

The transport refusals are outside that contract, because there is no tool call to answer: an over-budget unauthenticated request or an enforced MCP request/authentication concurrency refusal gets an HTTP 429 with Retry-After, and so does an over-rate PUT /transfer/upload — which releases its claim rather than consuming it, so the same link is still redeemable once the bucket refills.

Operational notes.

• Limiter state is in-process and is not persisted, so a restart begins with every bucket full. That is sound only because the container runs --workers 1; raising the worker count multiplies every rate above by the worker count.
• Refusals appear on /admin/performance as refusal counts, not in the latency percentiles. Repeated rate and enforced slot refusals are coalesced — one row per credential/tool/scope per MCP_REFUSAL_LOG_INTERVAL_SECONDS, each standing for 1 + suppressed refusals — so that a refusal loop cannot make writing the log the load.
• The velocity defaults are estimates against a small sample. Read /admin/performance for a week before treating any as settled, and disable one by setting it empty, null or none (zero is refused at startup).
• The daily quota is the durable ceiling and it is separate: keys created from now on get DEFAULT_DAILY_REQUEST_LIMIT (5,000), keys that already existed keep whatever they had, and OAuth grants have no daily ceiling at all — velocity bounds only.

The rationale lives in docs/architecture/rate-limits.md.

Concurrency admission

Concurrency admission ships with MCP_CONCURRENCY_MODE=shadow. It records pressure under concurrency_shadow on existing usage rows and emits bounded security events for request/authentication pressure. Calls keep their actual outcome, quota accounting and duration. Shadow mode observes current occupancy with zero wait; it does not predict how traffic would behave under enforcement.

In enforce mode, the server limits full MCP requests (including open SSE streams), authentication database sessions, tools, and usage-log writers. Tools pass velocity, vault and argument checks before acquiring slots; daily quota is checked afterward. A rejected tool receives slot_timeout without spending daily quota. Zero wait means immediate admission or refusal; a positive wait uses a bounded queue and one deadline. A retry hint is not a promise that a running call will finish by that time.

The four tool classes each default to one concurrent call: semantic_search uses embedding, find_related uses vector, the eight vault-mutating tools use write, and the remaining tools use other. Global, tenant and principal ceilings default to 4, 3 and 2. OAuth refresh keeps the same principal. Full request and per-bearer ceilings default to 32 and 4, authentication to 2, and usage writers to 1. All settings and queue limits are listed in .env.example.

Startup validates the pool budget as auth + 2 × tools + writers + 4 ≤ 15. The four connections of headroom are shared with panel, OAuth, indexing and transfer work; this arithmetic cannot guarantee availability when those other consumers exhaust it. Shadow mode does not enforce that budget. The controller is in-process and requires the existing single-worker deployment.

Review pressure observations and long-lived stream occupancy before enabling enforce. Choose off to disable concurrency admission; the existing velocity limits and daily quotas still apply. Shadow requires a zero tool wait and never adds a writer wait or drops a usage row because of its observed pressure.

Architecture

┌──────────────┐                       ┌──────────────────────┐
│ MCP clients  │   HTTP + Bearer key   │   FastAPI app        │
│  Claude Desk │ ────────────────────▶ │  ┌────────────────┐  │
│  Claude Code │                       │  │  MCP server    │  │
│  n8n agents  │                       │  │  (25 tools)    │  │
│  OpenWebUI   │                       │  └─────┬──────────┘  │
└──────────────┘                       │        ▼             │
                                       │  ┌────────────────┐  │
                                       │  │  Services:     │  │
                                       │  │  - vault       │  │
                                       │  │  - search      │  │
                                       │  │  - embeddings  │  │
                                       │  │  - links       │  │
                                       │  │  - indexer     │  │
                                       │  └─────┬──────────┘  │
                                       │        ▼             │
                                       │  ┌────────────────┐  │
                                       │  │ Postgres +     │  │
                                       │  │ pgvector       │  │
                                       │  └────────────────┘  │
                                       └──────────┬───────────┘
                                                  ▼
                                       ┌────────────────────┐
                                       │  Embedding         │
                                       │  provider          │
                                       │  (Ollama / OpenAI) │
                                       └────────────────────┘

Indexing pipeline

.md files in vault
    ↓ skip dot-dirs
parse frontmatter, extract tags (YAML + inline #hashtags)
    ↓ SHA-256 hash
skip if unchanged
    ↓
UPSERT notes_metadata (path, title, tags[], frontmatter JSONB,
                       content_hash, tsvector, modified_at)
    ↓
extract wikilinks/embeds/markdown-links → resolve targets →
note_links (source_id, target_id or NULL for dangling)
    ↓
chunk content (512 tokens, no overlap) → embed via provider →
note_embeddings (note_id, chunk_index, chunk_text, embedding[N])
    ↓
set embedded_content_hash = content_hash

The indexer runs on startup and every INDEX_INTERVAL_SECONDS (5 minutes by default). Hashes are content-only, so the change detector ignores mtime jitter. Stale embeddings are caught by the embedded_content_hash != content_hash mismatch.

Database schema

GIN indexes on content_tsvector and tags[]. B-tree indexes on the hot foreign keys. pgvector HNSW index on the embedding column (vector_cosine_ops, m=16, ef_construction=64); queries set hnsw.ef_search=80 and dedupe per note in Python after a 5x overfetch.

Project layout

src/
  main.py             FastAPI app, lifespan, MCP mount
  config.py           pydantic-settings
  database.py         async SQLAlchemy engine/session
  models/db.py        ORM models
  mcp_server/         MCP server, tools, auth middleware
  services/           vault ops, anchored filesystem, search, FTS,
                      embeddings, links, indexer, transfer
  transfer/           public /transfer/* capability-redemption routes
  auth/               login, sessions, per-request identity context
  api/                control-panel REST endpoints
  control_panel/      Jinja2 templates and static assets
  oauth/              OAuth 2.0 authorization-code flow
alembic/              database migrations
scripts/              one-off ops scripts (e.g. reset_embeddings.py)
tests/                pytest suite + smoke-test docs
openspec/             change proposals (spec-driven workflow)

Development

pip install -r requirements-dev.txt
pytest

The unit-test suite covers the embedding-provider abstraction, OpenAI batching and retry behavior, config validation, and the dimension-mismatch startup check. Network-bound tests use respx to mock httpx, so no real network access is required.

To run the server outside Docker:

DATABASE_URL=... SECRET_KEY=... VAULT_PATH=... uvicorn src.main:app --reload

Make targets

make init             First-time setup (data dirs, .env)
make build            Build Docker image (no cache)
make build-cached     Build Docker image (with cache)
make push             Push the image to the configured registry
make image            Build and push
make deploy           Build, scan, push, backup, migrate, recreate container
make up / down / restart / shell   Container lifecycle
make logs             Tail container logs
make db-init          Create database, user, and pgvector extension
make db-migrate       Run alembic migrations
make db-check         alembic check — schema vs. ORM models (must be clean)
make test-schema      Schema gate: migrations vs. models on a throwaway pgvector container
make db-backup        Dump database to backups dir
make db-restore FILE=<path>   Restore from a backup
make reindex          Explain how to trigger a reindex (panel only; there is no headless trigger)
make reset-embeddings Drop and recreate embedding column at configured dim
make rebuild-tsvectors Recompute keyword index for FTS_CONFIGS (no embeddings, no API calls)
make status           Show container and health status
make audit            Audit Python dependencies (pip-audit)
make trivy            Scan the local image for HIGH/CRITICAL CVEs (SCAN_IMAGE=obsidian-mcp:local for the bundled stacks)
make clean            Remove containers and images (data preserved)

make deploy runs the whole pipeline: build, image scan, push, database backup, alembic upgrade head, then recreate the container. Run make test-schema before any deploy that carries a migration, and make db-check after one.

Security notes

• API keys use the omcp_ prefix and are stored as SHA-256 hashes. The raw key is shown exactly once at creation.
• The control panel is intended to sit behind an external auth gateway. The included docker-compose.yml uses Traefik with an OAuth chain. Don't expose /admin directly to the internet.
• Panel sessions are server-side rows. The signed cookie carries a 256-bit random id; the database stores only its SHA-256, so a database dump contains no usable session. Logging out revokes that row, and a password change, an admin reset, a deactivation or a delete revokes every session of the account.
• The OAuth consent screen identifies the client it is asking about: the redirect host the authorization code would be sent to (taken from the URI's hostname, never its netloc, and shown in punycode rather than decoded), the server-generated client id, and the registration date. Every render says the application registered itself and is not verified by this server; a host outside OAUTH_KNOWN_REDIRECT_HOSTS is called out as unrecognised.
• The OpenAI key is rendered on the settings page as key[:8] + "..." + key[-4:] and never appears in full in HTML or JS sources.
• Path traversal is blocked at the service layer, and containment is proved by the kernel: every directory below the vault root is opened with one openat2(RESOLVE_BENEATH | RESOLVE_NO_SYMLINKS | RESOLVE_NO_MAGICLINKS) from an open root descriptor, and the rest of the operation acts on that descriptor rather than re-walking a name.
• Mutating tools act on the path as named. A final component that is a symlink is refused (naming the link's target) instead of being followed, so an in-vault alias cannot redirect a write. Reads still follow links, which is what an alias is for.
• Every path guard also refuses hidden components, so .obsidian, .git, .trash and friends are out of reach of every tool.
• Transfer links carry their token in the URL fragment, which browsers never send, and are redeemed only from an Authorization: Bearer header. Keep header logging off at your reverse proxy and APM. Unknown, expired, consumed and revoked tokens all get one identical 404 from the public routes; precise status comes from the authenticated check_upload tool.
• import_from_url fetches only genuinely public addresses, under an explicit deny list re-applied at every redirect.
• Failed /mcp authentication is budgeted per client address, counted before the credential lookup so a refused probe costs no database session and no query. The address comes from the proxy headers the app trusts, never from a header read directly, and a request with no resolvable address is charged to a shared slot rather than exempted. What it bounds is the database work an unauthenticated caller can force; it is not a defence against guessing a 256-bit key. See Rate limits.
• Parameterized queries everywhere. No string interpolation into SQL.
• Response headers include HSTS, X-Content-Type-Options: nosniff, and X-Frame-Options: DENY.

Status

Single-author, in active use as the maintainer's personal exocortex (2,500+ notes, multiple connecting agents). Public for anyone who wants to fork it. Issues and PRs welcome but expect opinionated review. This is a working system, not a generic platform.

License

MIT. See LICENSE.