jarvis

Local-first code intelligence for coding agents. Precomputed SCIP navigation (go-to-definition, find-references, call hierarchy, document symbols), Zoekt lexical search, cross-repo blast radius, and semantic search — exposed as MCP tools to Claude Code, Cursor, or any MCP client.

Runs as a single stdio process reading local SQLite files. No server, no auth, no network, nothing leaves your machine.

Quick Start

# 1. External indexer binaries (scip, zoekt, per-language indexers)
curl -fsSL https://raw.githubusercontent.com/jarvis-intelligence/jarvis-index/main/setup.sh | sh

# 2. jarvis itself
uv tool install jarvis-mcp

# 3. Index a repo (slug defaults to the directory name)
jarvis index /path/to/your/repo

4. Register the MCP server. Using Claude Code, install the plugin and it registers itself:

/plugin marketplace add jarvis-intelligence/jarvis-index
/plugin install jarvis@jarvis

Any other MCP client (or Claude Code without the plugin) registers manually:

claude mcp add jarvis --scope user -- jarvis-server

That's it — ask your agent "find all references to AuthService" and it will call findReferences instead of grepping.

{
  "mcpServers": {
    "jarvis": {
      "command": "jarvis-server"
    }
  }
}

git clone https://github.com/jarvis-intelligence/jarvis && cd jarvis
uv sync
claude mcp add jarvis --scope user -- uv --directory "$(pwd)" run jarvis-server

MCP tools

Every nav tool takes repo (the slug from jarvis index) plus a tool-specific symbol or path. All tools report failure the same way — a {"error": "..."} payload rather than a transport-level error, so a query bug never kills the stdio server.

Requirements and limits

Read this before installing — jarvis is deliberately narrow.

• macOS and Linux only. Windows is not supported.

• One language per repo. Language is detected by extension plurality across git-tracked files; a polyglot monorepo gets indexed as whichever language has the most files. Multi-language merge is out of scope. Override with --language.

• SCIP navigation (goToDefinition, findReferences, etc.) covers four language families: TypeScript/TSX, Python, Java/Kotlin, Swift. jarvis index --search-only additionally covers Go, Ruby, Rust, C, C++, C#, PHP, Scala, shell, and SQL for searchCode/semanticSearch only — no navigation.

• Navigation and search only — jarvis never edits code. If you want an agent that can perform semantic renames and refactors, you want Serena; the two are complementary.

• Indexing is a separate, explicit step. Nothing is live-analyzed. Run jarvis index (or jarvis watch) to publish an index before querying.

• Requires external binaries that setup.sh installs:

  Purpose	Binary	Source
  SCIP → SQLite conversion	scip	prebuilt, pinned v0.9.0 (minimum — older versions silently drop occurrence ranges)
  Lexical search	zoekt-index · zoekt-webserver	cross-compiled by our CI — upstream publishes no binaries
  TypeScript indexing	scip-typescript	npm install -g
  Python indexing	scip-python	npm install -g
  Swift indexing	scip-swift	prebuilt, macOS arm64 only
  Java/Kotlin indexing	scip-java	detect-only — Docker image, asks before pulling

  Options: --only <name> to install one dependency, --force to reinstall, --help for usage. Re-running is safe: anything already present is skipped.

Optional extras:

uv tool install "jarvis-mcp[watch]"      # + watchdog, for `jarvis watch`
uv tool install "jarvis-mcp[semantic]"   # + lancedb/sentence-transformers/tree-sitter, for semanticSearch

Why it's built this way

Storage is the seam. The runtime half only ever reads down into it; the indexing half only ever writes up into it; the two share no other contract:

Three things worth reading the diagram for:

• The runtime path never writes. Queries open a published index-<sha>.db read-only (mode=ro&immutable=1). Index files are never mutated in place.
• Publishing is atomic. A reindex writes a new versioned .db, populates the package graph, and runs zoekt-index — only once all of that succeeds does os.replace (POSIX rename(2)) flip the small current pointer. A query already reading the old file keeps working; there is no downtime window, and a failure anywhere leaves the previously published index live.
• The package graph is rebuilt, not accumulated. Each reindex clears that repo's own outgoing edges before recomputing them, so a removed dependency's edge is retracted — blastRadius always reflects each repo's last index run.

Editable source: docs/assets/jarvis-layers.dot (Graphviz). Layer-by-layer detail, the full index pipeline, and the semantic path are in docs/system-architecture.md.

Core query/search logic is ported from an internal reference implementation; the enterprise shell (FastAPI, Postgres, hosted-git auth, Cloud Build) is dropped in favor of a single stdio process reading local SQLite files.

Indexing a repo

jarvis index /path/to/your/repo            # slug defaults to the directory name
jarvis index /path/to/your/repo --slug foo # or pick one explicitly
jarvis index /path/to/your/repo --scheme MyScheme # Swift repo with an ambiguous Xcode scheme
jarvis index /path/to/your/repo --language python # force the language instead of detecting it from git-tracked files
jarvis index /path/to/your/repo --semantic-include vendor/generated # force-include a path the generated-file filter would otherwise skip
jarvis index /path/to/your/repo --search-only # skip SCIP indexing; publish only Zoekt + semantic search
jarvis index /path/to/your/repo --fallback-search-only # opt in: if the indexer fails mid-build, degrade to search-only (status `degraded`) instead of failing; the next reindex retries the full build
jarvis list
jarvis status foo
jarvis reindex foo
jarvis forget foo

status (as shown by both list and status) is usually indexed or failed, but can also be partial: the index published real symbols but no navigable positions (an indexer/converter bug) — check the stderr warning from jarvis index for details.

--semantic-include is repeatable — pass it once per path prefix to force-include several. Like --scheme and --language, once set there is no flag to clear it; change it by re-running jarvis index with the new value(s).

Language detection counts source files by extension across git-tracked files and picks the winner — one language per index:

Ties break by fixed priority (.ts → .tsx → .py → .java → .kt → .swift). .git, node_modules, .venv, __pycache__, dist, and build are skipped. Reading git rather than walking the filesystem is deliberate: a walk also counts gitignored scratch directories, which can outnumber a repo's own code and pick a language it doesn't use.

The pipeline then runs: chosen indexer → scip expt-convert → populate the package dependency graph (packages/edges tables in registry.db) → zoekt-index into ~/.jarvis/.zoekt → copy to ~/.jarvis/scip/_/<slug>/_/index-<sha>.db → atomic current pointer flip → registry update.

The scip/_/<slug>/_/ path shape reuses the vendored IndexConnectionCache's (project, repo, branch) 3-tuple layout with the outer two pinned to _ (see src/jarvis/config.py). It is not a user-facing contract — only <slug> matters when calling tools.

Swift indexing works end-to-end. It requires scip >= v0.9.0: older converters cannot read scip.proto's typed_range oneof, which is the only range encoding scip-swift emits, and silently produce an index with no navigable positions. jarvis index refuses an older scip rather than publishing one.

Indexing a Swift repo with code-signed app-extension targets additionally requires scip-swift >= v0.1.2: earlier versions pass no code-signing overrides to xcodebuild, which then fails provisioning for every signed target before compiling anything. Because setup.sh skips any dependency that is merely present, an existing install is not upgraded by re-running it — use sh ./setup.sh --only scip-swift --force.

Watching a repo (auto-reindex)

jarvis watch /path/to/your/repo             # debounce defaults to 5s
jarvis watch /path/to/your/repo --debounce 3
jarvis watch /path/to/your/repo --scheme MyScheme
jarvis watch /path/to/your/repo --language python

Runs in the foreground (not a daemon) using watchdog — install it with the watch extra. A burst of file changes (e.g. an editor's atomic save touching several files) coalesces into exactly one reindex. The reindex fires once --debounce seconds (default 5) have passed since the last file change — this prevents thrashing on rapid edits. .git, node_modules, .venv, __pycache__, dist, and build are ignored.

Tool details

• getIndexStatus takes an optional repo_path (the repo's local git working directory) to compare the published commit against git rev-parse HEAD. Omitted, freshness is reported without a staleness check — never stale: true without evidence.
• searchCode takes query plus an optional repo filter. On first call it lazy-spawns an embedded zoekt-webserver (pidfile'd so a second jarvis process reuses it instead of spawning a duplicate; killed on clean exit via atexit).
• blastRadius takes repo plus symbol_or_package (e.g. "npm:@scope/ name", the same "{manager}:{name}" string jarvis index derives from each repo's SCIP symbols). Returns every other indexed repo whose package depends on it, up to 2 hops, each tagged with its hop distance. The package graph has no per-node timestamp, so freshness is always "unknown" here — an honest limitation of the schema, not a bug. Cross-repo edges resolve by exact package name against whatever has already been indexed: index the dependency first, or re-run jarvis index/reindex after indexing it, for an edge to appear. Each reindex retracts that repo's own stale edges before recomputing them, so a removed dependency's edge disappears too — the graph always reflects each repo's last index run, not an accumulation of every run it's ever had.
• semanticSearch takes repo plus a natural-language query. Requires the optional semantic extra. Results fuse a LanceDB vector search over tree-sitter-chunked code with searchCode's Zoekt hits via reciprocal rank fusion. Raises a clear error if the repo has never been indexed with the extra installed (jarvis reindex <slug> after installing it builds the missing table); indexing itself is non-fatal — a failure there never blocks the rest of jarvis index. Semantic indexing also respects .gitignore (on top of the hardcoded ignore-directory list) and skips any file over 1 MB, in addition to the existing generated-file banner/long-line detection — --semantic-include overrides all three.

Known upstream limitations

These are real behaviors of the underlying SCIP tooling (scip expt-convert as of v0.9.0, scip-java, scip-kotlinc), not jarvis bugs:

• typeHierarchy returns an explicit {"error": ...}, not empty arrays, on indexes built with an unpatched upstream scip — that converter declares global_symbols.relationships in its schema but never writes it. An empty result would wrongly assert "no supertypes"; the error says "cannot tell" instead. setup.sh installs a fork build carrying the fix, so a fresh jarvis reindex <slug> makes the tool work. Reported upstream: scip-code/scip#464, fix scip-code/scip#465 (open, CI green).
• displayName / kind are backfilled from the symbol string. The converter never populates global_symbols.display_name/.kind, so query.py's _display_and_kind parses both from the SCIP symbol string whenever the database columns are empty (which they still normally are) — documentSymbols returns real values in practice; only a genuinely unparseable symbol falls through to null.
• searchCode's repo filter matches Zoekt's own repository name, which jarvis index now names after the slug via zoekt-index -meta — so this no longer diverges for repos indexed with current code. Shards published by an older jarvis still carry their old directory-derived name until you jarvis reindex <slug>.
• scip-java can't index Android/Gradle repos at all — its Gradle plugin keys off Gradle's standard source sets, which AGP replaces with its variant model, so the build emits zero SCIP shards (scip-java#177).
• Kotlin indexing requires an exact Kotlin version match — scip-kotlinc is compiled against one pinned Kotlin release (SCIP_JAVA_KOTLIN in setup.sh, currently 2.2.0); its compiler-plugin API is internal and unstable even across patch releases, so any other version fails. Both cases are detected automatically from the indexer's own failure output and degrade to --search-only rather than failing outright.
• Maven-built Java repos need bash >= 4.4 on macOS — scip-java's generated javac wrapper (#!/usr/bin/env bash, set -eu) expands "${LAUNCHER_ARGS[@]}" unguarded, which errors on bash < 4.4; macOS ships only 3.2, so the build dies at default-compile with LAUNCHER_ARGS[@]: unbound variable. setup.sh works around it by linking ~/.jarvis/shims/bash to a newer bash and putting that one directory first on PATH for the indexer. If no bash >= 4.4 is installed, indexing fails with the remedy rather than degrading to --search-only — unlike the two cases above, this one is fixable (brew install bash), and a persisted --search-only cannot be un-set.

Configuration

Data directory (default ~/.jarvis):

JARVIS_DATA_DIR=/custom/path jarvis index /path/to/repo

Environment variables:

• JARVIS_DATA_DIR — override default ~/.jarvis for all indexes and registry
• JARVIS_FALLBACK_SEARCH_ONLY — default the opt-in degrade-to-search-only fallback on for repos indexed without an explicit --fallback-search-only / --no-fallback-search-only flag. Accepts 1/true/yes/on (case-insensitive); any other value is treated as off with a one-line warning.
• JARVIS_EMBEDDING_QUERY_PREFIX / JARVIS_EMBEDDING_DOC_PREFIX — override the query/document instruction prefix applied before embedding. Auto-detected for bge-m3, e5, and nomic-embed; set these if using a different model that needs one — semanticSearch warns when an unlisted model has no prefix configured.

Agent skills

Three agent skills ship in the Claude Code plugin, under plugin/skills/:

• jarvis-setup — install, register, index, verify.
• jarvis-use — prefer jarvis for structural queries (find references, go-to-definition, hierarchy).
• jarvis-issues — file jarvis bugs/features via gh.

Install them, and register the MCP server, with:

/plugin marketplace add jarvis-intelligence/jarvis-index
/plugin install jarvis@jarvis

See Quick Start above for the manual registration alternative.

Standards

Blob decoding follows the SCIP protocol: scip_pb2.py is generated from scip.proto at scip-code/scip tag v0.9.0 (regenerated up from v0.7.0, which lacked the typed_range oneof scip-swift requires), and occurrence/relationship blobs are decoded as real scip.Document / scip.SymbolInformation messages.

The SQLite layer (documents, chunks, global_symbols, mentions, defn_enclosing_ranges) is not part of that published spec — it is the output shape of the experimental scip expt-convert sub-command, verified by hand against a real index. Treat it as a moving target across scip releases.

Tests

uv run pytest

Integration tests that shell out to the real scip-python / scip / zoekt-index binaries are marked integration:

uv run pytest -m "not integration"   # unit only
uv run pytest -m integration         # real-binary pipeline

Documentation

• docs/project-overview-pdr.md — scope, value prop, out-of-scope items
• docs/system-architecture.md — architectural guarantees, storage layout, query paths
• docs/codebase-summary.md — module map, test coverage
• docs/code-standards.md — code patterns and conventions
• docs/project-roadmap.md — all phases complete, future ideas

All 4 planned phases are shipped — see plans/0724-2316-jarvis-mcp-implementation/plan.md.

License

MIT