codetopo
Registry identity: mcp-name: io.github.nrupala/codetopo
See your codebase the way an agent sees it. codetopo turns a source repository into a queryable code graph — symbols, calls, references, and structure — with JSON answers an AI coding agent can act on directly.
codetopo holds the OpenSSF Best Practices Passing badge — 100% of 67 criteria.
You are here because an agent just touched your codebase and you want to answer structural questions fast: what calls this function? what breaks if I change this? how are these two modules connected? codetopo indexes the tree once and answers those questions as data — counts first, full lists only when asked — over a CLI, an HTTP API, or the MCP tools your agent already speaks.
It is deliberately domain-blind: no scoring, no rankings, no opinions about what the code means. It answers structural questions about source trees, and only that. Domain meaning belongs to the agents and teams composing on top.
Live demo: codetopo.aimlds.org — codetopo's real map of this repository, drawn by codetopo itself and refreshed on every push to main.
60-second quickstart
Requires a stable Rust toolchain (install via rustup).
# Install the four binaries: codetopo (CLI), codetopo-server (HTTP API),
# codetopo-mcp (stdio MCP server), codetopo-acp (stdio ACP agent).
# The packages must be named — a bare install URL fails because the
# repository contains multiple binary packages.
# Index a repository (Rust and TypeScript/JavaScript are parsed via tree-sitter)
# Find a symbol's id by name (every surface has this discovery step)
# Ask what calls what (the package prefix is your directory's name,
# exactly as indexed — hyphen included)
That's the whole loop: index once, query as data. Aggregate-first
output — totals before lists — keeps token budgets and pipes sane by
default; --expand brings the full list when you want it.
Example queries
Against the fixture repository used in the project's own end-to-end
verification (docs/phase-2-verification.md) — output shown is the real
output of the CLI:
A few more things you can ask (same engine, same data):
The path answer renders as typed steps, e.g. a --calls--> x --imports--> b
(illustrative; exact ids depend on your index).
Giving it to your agent
HTTP API (codetopo-server)
The same queries over JSON, behind Bearer API-key auth, with per-request metering — one transport for any client, any language:
&&
# {"index_id": "<uuid>", "files_indexed": 3, "files_failed": 0, "nodes": 15, "edges": 23, "diagnostics": 2}
Eight endpoints under /v1, one index operation plus seven read-only
queries: POST /v1/index, GET /v1/graphs/{id}/{descendants, ancestors, blast-radius, path, stats, verify, snapshot, neighborhood, symbol, search}, GET /v1/schema. Full contract in
schemas/openapi.yaml (OpenAPI 3.1.0), env knobs
(CODETOPO_BIND, CODETOPO_DATA_DIR, CODETOPO_KEYS_FILE,
CODETOPO_METERING_LOG) documented in schemas/README.md.
MCP (codetopo-mcp)
A stdio JSON-RPC server exposing the same handler layer as twelve tools —
the eight query/integrity tools (codetopo_index, codetopo_descendants,
codetopo_ancestors, codetopo_blast_radius, codetopo_path,
codetopo_stats, codetopo_verify, codetopo_snapshot) plus discovery
(codetopo_search_symbols, codetopo_get_symbol,
codetopo_neighborhood, codetopo_describe_schema) — so an in-process
agent learns one surface and gets both. MCP answers are identical to the HTTP answers by
construction: both call the same codetopo-core queries.
ACP (codetopo-acp)
A stdio agent speaking the Agent Client Protocol — for editors and clients that converse with agents instead of calling tools. Open a session on a working directory and ask code-structure questions in plain language:
→ {"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}
← {"jsonrpc":"2.0","id":1,"result":{"agent":"codetopo","protocolVersion":"1.0",
"agentCapabilities":{"read_only":true,"code_structure":true},"agentInfo":{"name":"codetopo"}}}
→ {"jsonrpc":"2.0","id":2,"method":"session/new","params":{"cwd":"/path/to/repo"}}
→ {"jsonrpc":"2.0","id":3,"method":"session/prompt",
"params":{"sessionId":"<id>","prompt":"blast radius of auth::login"}}
← {"jsonrpc":"2.0","id":3,"result":{"answer":"Blast radius of 'auth::login': 14 nodes — …"}}
Deterministic intent routing (descendants / ancestors / blast-radius / path /
stats / snapshot / verify) straight to the same codetopo-core queries —
no LLM, no network, no file writes. Edit requests get an honest read-only
refusal. stdout carries only protocol frames; diagnostics go to stderr.
Architecture
┌─────────────┐ tree-sitter ┌──────────────┐
│ source tree │ ───────────────► │ codetopo- │ extract symbols,
└─────────────┘ Rust, TS/JS │ extract │ calls, references
└──────┬───────┘
▼
┌──────────────┐
│ codetopo-core│ graph model:
│ │ nodes, edges,
│ │ traversal queries
└──────┬───────┘
▼
┌──────────────┐
│ codetopo- │ SQLite store +
│ store │ hash-chained
│ │ audit log
└──────┬───────┘
│
┌──────────────────────────────────┼─────────────────────────────┐
▼ ▼ ▼ ▼
┌──────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ codetopo │ │ codetopo- │ Bearer auth │ codetopo- │ stdio │ codetopo- │ stdio
│ (CLI) │ │ server │ JSONL │ mcp │ JSON-RPC │ acp │ ACP
└──────────────┘ └──────────────┘ metering └──────────────┘ └──────────────┘
| Crate | Role |
|---|---|
codetopo-core |
Graph model, node/edge kinds, all traversal queries, snapshots |
codetopo-extract |
tree-sitter extractors (Rust, TypeScript/JavaScript) |
codetopo-store |
SQLite persistence + tamper-evident, hash-chained audit log |
codetopo-cli |
codetopo headless CLI — the reference query surface |
codetopo-server |
codetopo-server axum HTTP API (thin adapter over core) |
codetopo-mcp |
codetopo-mcp stdio MCP server (thin adapter over core) |
codetopo-acp |
codetopo-acp stdio ACP agent — conversational sessions over the Agent Client Protocol (thin adapter over core) |
The CLI, HTTP, MCP, and ACP surfaces are all thin adapters over the same
handlers — they cannot disagree. schemas/ keeps the machine-readable
contracts (OpenAPI 3.1, MCP tool catalogue) checked against the running
servers. The store's audit log follows the trust-primitive pattern used
across this product line: a hash-chained, append-only record, verified by
codetopo verify / GET …/verify.
Docs
docs/render.md— deterministic render: the frozen layout spec (JSON) + SVG rules behindcodetopo renderdocs/protocol-proof.md— MCP/ACP conformance evidence: strict stdio clients + official-schema validationdocs/scale-gate.md— the 10,000-file scale gate and its measured recordDockerfile+docker-compose.yml— self-hosted container: the HTTP API over your own indexesdocs/phase-2-design.md— API design: endpoints, auth, metering, aggregate-first disciplinedocs/phase-2-verification.md— end-to-end verification report (fixture numbers above come from here)docs/phase-3-acp-design.md— ACP agent design: wire protocol, intent routing, read-only guaranteedocs/phase-3-acp-verification.md— ACP verification reportdocs/quickstart-verification.md— clean-machine verification of the quickstart and example queries aboveschemas/README.md— the two agent-facing contracts and where the spec departs from the design docCODE_SCHEMA.md— the graph schema (§6 sets the aggregate-first rule)
Privacy Policy
codetopo runs entirely on your machine. The MCP server (codetopo-mcp), the CLI, and the index databases they create never send your code, your queries, or any usage data anywhere: there are no accounts, no API keys, no telemetry, and no network calls at runtime. Indexing reads only the repositories you point it at and writes only the database paths you give it. Questions about privacy: info@nrupalakolkar.com.
License
- Engine (
codetopo-*crates): AGPL-3.0-or-later (seeLICENSES/) - Agent-facing contracts (
schemas/: OpenAPI 3.1 spec, MCP tool catalogue): Apache-2.0 (seeschemas/LICENSE-APACHE)
Owned by Nrupal Akolkar · Built with Muse by Meta — join Muse, then redeem code 101EEK in Settings within 48 hours of joining and we both earn tokens.