StateArk

Never explain your project to an AI twice.

Your chat is a workspace, not an archive. StateArk turns a working session into a versioned, portable project state that any new chat can pick up. Open terminal and write:

npx stateark

Then quit Claude Desktop completely (Cmd+Q) and reopen it. That is the whole install.

(Prefer to install from source? npm install && npm run build && npm run setup — see INSTALL.md. Undo any time with npx stateark remove.)

Here what we had in mind developing stateark - we address a problem millions of users have...

You:  ...three hours of work...
You:  Savepoint
      -> my-project v0.3 written to ~/StateArk

      (next day, new chat, empty context)

You:  Resume my-project
      -> decisions, constraints, rejected approaches, open questions,
         and the actual files - all current, none of the detours

Local by default. Savepoints are ordinary folders of Markdown, JSON and your real files. Nothing is uploaded. If StateArk disappears tomorrow, you still have everything.

~/StateArk/projects/my-project/v0.3/
  state.md        <- the canonical state, readable in any editor
  state.json
  manifest.json   <- every artifact with its SHA-256
  artifacts/
    app.py
    schema.sql

Commands

What makes it more than a summary

• Carry-forward. Files you do not re-submit are copied into the new version with their original hash, so a forgetful model cannot silently lose your schema.
• Integrity checks. StateArk cannot see your chat - so it checks what it can: truncation markers (// ... rest unchanged), files that collapsed in size, files the state describes but never handed over, savepoints identical to their predecessor. It warns, it never blocks: the savepoint is always written.
• Journal. Between savepoints the model quietly records turning points, so a Savepoint at the end of a long session does not depend on a degraded context. Only for projects you already saved once - scratch conversations are never touched.

Requirements

Node 20+. Claude Desktop, or Claude Code (claude mcp add --transport http ...). It also runs under other local MCP hosts — users have it working with Codex and Hermes against the same store.

Hosted clients that dial your server from the vendor's cloud (Claude web, ChatGPT web) cannot reach localhost. See INSTALL.md.

Several agents at once

Supported, and worth saying plainly because it is the setup that broke first. Writes to a project are serialised across processes, so two agents cannot claim the same version number. If a savepoint reports that another process is busy, repeat it in a moment — nothing is lost. A process that dies mid-write releases its lock automatically.

If your store lives in Dropbox, OneDrive or iCloud Drive: several agents on one machine are fine. Two machines writing the same store simultaneously is a sync conflict, which the sync client resolves and StateArk cannot.

Licence

Elastic License 2.0. Use it for anything including commercially, read and modify the source, share the package. You may not offer it to third parties as a hosted or managed service. Your savepoints are yours and are not covered by this licence.

Architecture

ChatGPT / Claude / Gemini-capable MCP client
                 |
                 v
          StateArk Local Agent      <- 127.0.0.1 only, by default
                 |
                 v
        ~/StateArk/projects         <- SOURCE OF TRUTH
                 |
                 | optional
                 v
             Supabase               <- mirror / backup / transport

A Savepoint is an ordinary directory, not a proprietary database:

StateArk/projects/my-project/
  journal.ndjson      # entries recorded since the last savepoint
  project.json        # index (self-healing: rebuilt from disk if corrupt)
  v0.3/
    state.md          # human- and LLM-readable canonical state + integrity warnings
    state.json        # the same state, structured
    meta.json         # version, lineage, platform, sync status, warnings
    manifest.json     # artifacts with SHA-256, stored/pending, carried_forward_from
    journal.ndjson    # the entries this savepoint consolidated
    artifacts/
      app.py
      schema.sql
      prototype.zip

If StateArk disappears, those files remain usable.

Install

Requires Node 20+.

npm install
npm run typecheck
npm test          # 97 checks against a throwaway store — run this first
npm start

No .env is needed for local-only mode. Copy .env.example to .env only if you want to change the port, the store location, or enable the Supabase mirror.

On first run StateArk generates a random access key and stores it in ~/StateArk/.access-key (mode 0600). The startup banner prints your endpoints:

MCP:         http://localhost:8787/mcp/<key>
Upload page: http://localhost:8787/upload/<key>
Health:      http://localhost:8787/health

Security model

The HTTP entrypoint is a local server on your own machine, so StateArk assumes any web page you visit is hostile:

Before exposing the agent over HTTPS: set a long STATEARK_ACCESS_KEY, set STATEARK_ALLOWED_HOSTS to your tunnel hostname, and put a real reverse proxy in front.

Platform reality

Local-first and hosted LLMs are different networking environments. Claude Desktop/Code and other local MCP clients can talk to the local agent directly. A hosted ChatGPT/Gemini client generally cannot reach localhost on your computer. For those you would need a secure HTTPS route to your running local agent — remote access is on the roadmap and does not exist yet.

If you are setting this up, in this order:

1. run StateArk locally;
2. test Savepoint and Resume from a local MCP client;
3. enable Supabase sync only if you actually want a mirror;
4. expose the agent over HTTPS only after the three steps above work.

Local-only mode — the default

Leave SUPABASE_URL, SUPABASE_SECRET_KEY and STATEARK_OWNER_ID unset. Cloud sync stays off and files never leave the machine. Default store: ~/StateArk (override with STATEARK_LOCAL_ROOT).

This is what you get out of the box. The Supabase section below is opt-in.

Local + Supabase sync

Run the migrations in the Supabase SQL editor, in order:

1. supabase/migrations/001_stateark.sql
2. supabase/migrations/002_artifacts.sql
3. supabase/migrations/003_hardening.sql

Then set SUPABASE_URL, SUPABASE_SECRET_KEY (server-side Secret key, never a publishable key), STATEARK_OWNER_ID, and STATEARK_STORAGE_BUCKET.

Cloud is a mirror, not the master:

• the local savepoint is committed first and a cloud failure cannot invalidate it;
• cloud deletion does not delete local data;
• sync status lives in meta.json (disabled / pending / synced / failed);
• retry a failed push with the sync_savepoint tool.

Text/code artifacts up to 2 MB are mirrored inline in Postgres; everything else goes to the private Storage bucket. The mirror is not end-to-end encrypted — the Supabase project can read what it stores. Sync is off by default for exactly that reason; turn it on only for a project you would be comfortable putting in any hosted database.

Savepoint behaviour

When you say Savepoint, the host LLM should:

• reconstruct the latest valid state rather than summarise chronology;
• retain governing decisions, requirements and constraints;
• keep rejected approaches only when the reason matters;
• send exact text/code as transfer=text;
• send binary bytes as transfer=base64 only when truly available;
• otherwise mark the artifact transfer=pending instead of fabricating it;
• omit files that have not changed — they are carried forward automatically;
• never abbreviate a file with ... rest unchanged.

Local creation is atomic: StateArk writes a temporary bundle and renames it only when complete. Artifact names are sanitised; if a name had to be changed, manifest.json records the original under original_name.

Integrity checks

Every savepoint is compared against its predecessor. Findings are returned in the tool result and rendered at the top of state.md. The savepoint is always written — a check never costs you work, it only tells the model to come clean.

Read the full README on GitHub →