codeintel

One MCP tool that lets a coding agent search, trace, and understand a codebase — structurally, not by grepping. codeintel unifies three engines — a call/import graph, an LSP for exact symbols, and semantic embedding search — behind a single code.query call that routes to the right engine, caches the answer, and never throws. The agent always gets back a clean, well-formed result to reason over.

Status: beta (0.x), and young. The code.query surface is one call and has been stable since 0.8, the suite is thorough, and every release is gated by a canary that runs a real query against a built wheel. But this is a new project with a single maintainer, and each time it has been pointed at an unfamiliar codebase it has found real bugs. Use it locally, on a developer machine, for a single user — that is the case it is built and tested for. Before relying on it for anything beyond that, read Project status.

codeintel visualizing its own codebase. One command — codeintel graph <repo> --html — turns any indexed repo into a self-contained, interactive call graph you can open offline or share as a file. Layouts, complexity-sized nodes, click-to-inspect metrics, and JSON/Markdown/SVG/PNG export. See docs/graph-viewer.md.

Prefer plain text? codeintel map writes a readable architecture overview to CODE_INTEL.md — node/edge counts, ranked symbols by caller count, and entry points — for skimming or for MCP hosts that can't render a graph:

What CODE_INTEL.md is for. It's a static, committable snapshot of a codebase's shape — meant to be read (by a person or an agent) first, instead of reconstructing structure by grepping. It covers the cases the live code.query tool doesn't:

• Agents & hosts that don't speak MCP. Not every agent supports MCP, and the server isn't always running. codeintel map writes a plain file any agent can read; codeintel map --inject also drops a pointer into CLAUDE.md / AGENTS.md, so an agent picks up the codebase's structure automatically at the start of a session.
• A committed, diffable overview. It lives in the repo — reviewable in a PR, browsable on GitHub, available offline. Re-run codeintel map after codeintel index to refresh it.
• The load-bearing code at a glance. Ranking symbols by caller count surfaces what most of the codebase depends on (the risky-to-change core) plus the entry points — the first things a newcomer, or an agent, should understand before touching anything.

See docs/map-file.md for the format and the --inject flow.

Why an agent needs it

Without structural tools, an agent dropped into unfamiliar code falls back on grep and reads whole files to reconstruct relationships by hand — burning tokens, missing call sites, and guessing at blast radius before it edits anything. codeintel answers those questions directly instead:

• "What calls this? What breaks if I change it?" → the real call graph, which catches cross-file and module-level callers a text search silently misses.
• "Where is this symbol defined, and everywhere it's used?" → the language server, with exact locations.
• "Where's the code that does X?" (when you don't know the name) → semantic search over the repo.
• Always a clean answer. Every call returns the same JSON envelope. A missing or broken backend degrades to a safe null with a reason — so the agent falls back to grep instead of crashing on an exception it can't reason its way out of.

Net effect: fewer, sharper tool calls, less re-reading, and an agent that can see structure — callers, impact, call chains — that plain search can't.

What your agent can ask

It's one call: code.query(op, target, engine="auto"). In auto mode (the default) codeintel picks the engine per operation:

Pin one engine with --engine graph│lsp│semantic, or fan out with --engine both / all to merge results.

deadcode is a candidate list, not a delete list

deadcode is the one op whose output invites a destructive action, so it gets an explicit caveat. Every hit is re-read and verified against the source before it is reported, which removes the common false positives — but no reachability analysis sees every caller. Dynamic dispatch, registries and decorators, getattr lookups, entry points declared in packaging metadata, plugin discovery, reflection, and calls from languages the graph does not parse are all invisible to it. Through 0.14.x it was systematically wrong on callback-heavy code and confident about it; that class of defect is fixed, but the underlying limit is structural and permanent.

So: review each hit before deleting anything, and never wire deadcode into an agent that deletes without a human in the loop. Used as a ranked list of places worth looking, it is genuinely useful. Used as a work order, it will eventually remove live code.

Example — "who uses safe_null_result?"

// request
{ "op": "callers", "target": "safe_null_result", "engine": "auto" }

// response — always this exact envelope; `result` is ready-to-read markdown
{
  "ok": true, "op": "callers", "target": "safe_null_result",
  "engine": "graph", "cached": false,
  "result": "## Callers of safe_null_result (7)\n- …gateway [USAGE] (src/codeintel/gateway.py)\n- …providers.graph [USAGE] (src/codeintel/providers/graph.py)\n- …server [USAGE] (src/codeintel/server.py)\n- … (4 more)"
}

The agent hands result straight to the model. If the graph backend isn't installed, the identical call returns "result": null, "reason": "engine-unavailable" — no exception, and the agent just falls back to its own search.

What makes it good

• Local-first and private. One process on your machine — no cloud service, no API keys, no telemetry, no per-query network. Safe to point at a private repo, even with --engine all. (The one-time exception: fastembed downloads its embedding model once, then runs fully offline.)
• It never throws. Every call returns the same JSON envelope; a missing or broken backend degrades to null with a reason. No exceptions, no 500s, no malformed output for the agent to trip over — so you never wrap code.query in a try.
• One tool, not three. Register a single MCP server and it auto-routes each question to graph, LSP, or semantic — instead of wiring up three backends with three response shapes and three failure modes.
• Degrades instead of breaking. No graph backend installed? That engine returns null and the agent falls back to grep. The semantic engine needs nothing external, so codeintel is useful the moment it's installed and only gets sharper as you add backends.
• Fast on repeat, and the cache never lies. A content-hash cache returns instantly for unchanged code and self-invalidates when a background reindex advances the index, so you never read a cached answer for code that moved on. The cache is bounded (LRU), so a long-running server holds steady memory. (The cache is always consistent with the index; how current the index itself is depends on the engine — see Keeping answers fresh.)
• Concurrency-safe. The HTTP transport handles requests on threads, so one slow query (an LSP session warming, a first-time index) can't block every other agent.
• Honest about its own health. codeintel doctor answers three separate questions per engine — installed? runnable? is this repo indexed? — with the single command to fix each gap, so "installed" is never mistaken for "working". And a readiness claim is one a query can actually honor: install a missing backend mid-session and the running server picks it up on the next call, rather than reporting the engine healthy while quietly routing around it until you restart the host.

Quickstart

pip install codecortex

This installs the codeintel CLI; the semantic engine works out of the box. (On PyPI the distribution is codecortex because codeintel was taken; the CLI and import stay codeintel.)

One command prepares the rest and indexes your repo:

codeintel setup --all /path/to/your/project

This installs uv (for the LSP engine), warms serena, downloads the embedding model, indexes the repo, and prints a health report ending in a Next: list — exactly what's ready and the one remaining step. It's idempotent, so re-running is safe. The graph engine (codebase-memory-mcp) is an optional external binary that adds who-calls / impact / hotspots / changed; codeintel is fully usable without it.

Or from source:

git clone https://github.com/hamilton-sky/codeintel.git
cd codeintel
pip install -e .

Register with your AI agent(s), then query:

codeintel install            # registers with the agents you actually have installed
codeintel query --op search --target "authentication middleware"

Enable native Codex integration

codeintel is an MCP server, so Codex can call its tools directly rather than invoking the CLI. After installing the package, explicitly register it with Codex:

codeintel install --agent codex

This safely adds a [mcp_servers.codeintel] entry to ~/.codex/config.toml (or $CODEX_HOME/config.toml when that is set) without changing your other Codex settings. Registration is deliberately opt-in: installing a Python package should not silently modify an agent's configuration. Start a new Codex task (or restart Codex) after registration; the refreshed task will have native code.query, code.status, code.doctor, and code.map MCP tools available.

For a fully prepared local setup, run:

codeintel setup --all /path/to/your/project && codeintel install --agent codex

Enable native Claude Code integration

After installing the package, explicitly register it with Claude Code:

codeintel install --agent claude

This adds the codeintel MCP server to ~/.claude.json (or $CLAUDE_CONFIG_DIR/.claude.json) while preserving your existing configuration — that is the file Claude Code reads for user-scope MCP servers, and you can confirm the entry with claude mcp list. Start a new Claude Code session after registration so it can load the native code.query, code.status, code.doctor, and code.map MCP tools.

Upgrading from ≤ 0.11.1? Earlier versions wrote this block to ~/.claude/settings.json, which Claude Code ignores for MCP registration — so codeintel never actually loaded. Re-run codeintel install --agent claude; it registers in the right place and points out the stale entry so you can delete it.

For a fully prepared local setup, run:

codeintel setup --all /path/to/your/project && codeintel install --agent claude

Registration is verified, not assumed

It only touches agents you have. codeintel install defaults to --agent auto: it registers the hosts whose config root already exists on this machine and names the ones it skipped. Installing a Python package should not create ~/.gemini/ and ~/.config/zed/ for someone who has neither. Force a specific host with --agent claude|codex|gemini|zed, or every supported host with --agent all.

It registers an absolute path. The bare name codeintel is resolved by the host, not by the shell you ran install in — and a GUI-launched desktop agent does not source your shell profile, so a command your terminal finds is routinely invisible to the app. That is the one failure a handshake run in your terminal cannot catch, because it inherits the PATH that works. Pass --relative-command for the bare name. If a later upgrade moves the binary, re-running codeintel install repairs the stale path in place, leaving the rest of your config untouched.

Then it launches the exact command it registered and drives a real MCP handshake — initialize → tools/list — and reports what came back:

v claude: registered at /Users/you/.claude.json

v verified: codeintel 0.11.2 — 4 tools (code.query, code.status, code.doctor, code.map)

If the command is not on PATH, or the server fails to start, install says so and exits non-zero instead of reporting a success your agent cannot use. Pass --no-verify to skip the handshake.

The same principle gates releases. Because every result is a safe envelope with ok: true and the CLI never throws, an exit-code smoke test passes against a build that boots cleanly and answers nothing — so scripts/release_canary.py runs before every publish against the built wheel in a clean environment: it registers Codex and Claude Code into a throwaway HOME, launches the command those config files name, and asserts on the answer text of a real code.query over a fixture repo. A release that writes a config no host reads, or that returns ok: true with nothing in it, fails there instead of on your machine.

Full reference — what each host reads, the absolute-path rationale, and troubleshooting: docs/install.md.

How it works

A Gateway receives every query and dispatches it to one of three providers — graph (structural relationships), LSP (precise symbol resolution), or semantic (embedding-based search) — based on the operation type. Each provider is fully isolated: if it is unavailable or raises an exception, the gateway catches it and returns a safe-null envelope. The caller always gets a well-formed response with no exception to catch.

flowchart LR
    A["AI agent · MCP"] --> GW
    H["Harness · HTTP"] --> GW
    C["Developer · CLI"] --> GW
    GW["Gateway<br/>route · cache · safe-null"] -->|"auto: search"| SP[SemanticProvider]
    GW -->|"auto: impact / callers / …"| GP[GraphProvider]
    GW -->|"auto: symbol"| LP[LspProvider]
    GP --> GB[("codebase-memory-mcp")]
    LP --> LB[("language server")]
    SP --> SB[("fastembed + sqlite-vec")]

Full walkthrough: docs/architecture.md · docs/query-flow.md.

Safe-null contract

Every Gateway.query() call returns a dict with exactly these keys:

{"ok": true, "op": "search", "target": "auth", "result": null, "engine": "semantic", "cached": false}

ok is always true. result is null when no provider has an answer — never an exception, never a 500. Callers must check result is not None before using the value.

The optional keys are the ones worth reading when an answer surprises you:

codeintel query --json prints this envelope from the CLI.

Engines

Run codeintel doctor at any time to see which engines are actually ready for a repo and how to fix the ones that aren't.

Keeping answers fresh

The three engines have genuinely different freshness models, and it's worth knowing which you're reading:

So a callers/impact/hotspots answer is only as current as your last index. If a result describes code you just changed — or a symbol you just added comes back reason: "not-in-graph" — that's the signal to re-run:

codeintel index /path/to/repo

The reply names the fix when it can: a missing symbol now returns a hint with the exact command rather than a bare reason.

One more honest caveat. For targets that are symbol names or free text (most of them — callers, impact, hotspots, search), there is no file whose content hash could change, so a cached answer is invalidated only when a background reindex completes, and those are debounced (~30s). An edit followed immediately by the same query can therefore return the pre-edit answer. Targets that are real file paths are content-hashed and refresh as soon as the bytes change.

Pass --engine auto (the default) and codeintel chooses the best engine per operation. Pass --engine both or --engine all to fan out to multiple engines and merge results.

Documentation

Full system docs live in docs/ — start with the index:

• Architecture — layers, the CodeProvider protocol, the safe-null contract, caching, freshness (ASCII + Mermaid).
• Install & registration — what each agent host actually reads, why the registered command is an absolute path, and the three levels of proof that registration worked.
• Query flow — request lifecycle, engine selection, fan-out & merge, and why it never throws.
• Map file — the static CODE_INTEL.md orientation layer for hosts with no MCP support.
• Benchmarks — real numbers at scale: 25 k chunks indexed in ~8 min, ~235 ms warm queries, 60 MB index.
• Engine references: graph · lsp · semantic.

CLI reference

Human-facing commands (doctor, status, query, setup, reset) honor --no-color / NO_COLOR and --ascii, and auto-degrade to plain text when piped.

Exit codes, so a make target or CI step can gate on $?:

Config

Create .codeintel.toml at your project root to override defaults:

backend          = "auto"                   # auto | graph | lsp | semantic
semantic         = "on"                     # on | off
reindex          = "on-demand"              # on-demand | never
cosine_floor     = 0.25                     # minimum similarity score for semantic hits (0–1)
max_chunks       = 500                      # max chunks to embed per file
max_total_chunks = 100000                   # safety ceiling on chunks embedded in one index pass
model            = "BAAI/bge-small-en-v1.5" # fastembed embedding model

Config is validated on load — an out-of-range number, a misspelled enum, or a wrong type falls back to that key's default (with a logged warning) instead of breaking every query.

Environment variables:

Privacy & dependencies

codeintel is local-first — one local process, no cloud service, no API keys, no telemetry, and no per-query network. Its own code makes zero outbound HTTP calls, and the HTTP transport binds to 127.0.0.1 only by default — binding a non-loopback host requires --allow-remote, and --token (or CODEINTEL_HTTP_TOKEN) then gates every request behind a bearer token. The server bounds concurrent connections, but for exposure to a hostile network you should still front it with a reverse proxy (TLS, rate-limiting) — the built-in http.server is not hardened for the open internet.

Bundled (installed with the package, run locally): mcp (the tool interface) · sqlite-vec (the semantic index, a local DB file) · fastembed (the local embedding model).

Optional external backends — auto-detected on PATH; if one is absent, that engine returns a safe-null and the agent simply degrades to grep:

Not sure what's installed? codeintel doctor reports exactly which backends are present, whether this repo is indexed, and the command to fix each gap.

The only network touch is first-run setup: fastembed downloads the BAAI/bge-small-en-v1.5 weights once (cached under ~/.cache, fully offline thereafter); the optional backends also install on first use if you opt in. After that, no code or data leaves your machine — which is what makes --engine all safe to run on a private repo.

For agents

Register codeintel as an MCP server (codeintel install) and the agent gets four tools:

Over MCP the agent calls code.query directly. Over HTTP, start the server and POST to /code/query:

codeintel serve-http &   # listens on 127.0.0.1:8766 by default

For a shared or remote deployment, start it with --allow-remote --token "$CODEINTEL_HTTP_TOKEN" and send Authorization: Bearer <token> on each request — a missing or wrong token gets a clean 401. Requests are handled concurrently, so one slow query never blocks another.

import urllib.request, json

def code_query(op: str, target: str, engine: str = "auto") -> dict:
    body = json.dumps({"op": op, "target": target, "engine": engine}).encode()
    req = urllib.request.Request(
        "http://127.0.0.1:8766/code/query",
        data=body,
        headers={"Content-Type": "application/json"},
    )
    with urllib.request.urlopen(req) as resp:
        return json.loads(resp.read())

result = code_query("search", "authentication middleware")
if result["result"] is not None:
    print(result["result"])   # ranked semantic matches

The response is always JSON-safe. Check result["result"] is not None before use. Never catch an exception from the gateway — it never raises.

Operations & deployment

Running codeintel as a shared service? It ships with what ops teams expect:

Plus bearer-token auth — or RBAC (per-token roles + op scopes via auth.toml; a disallowed op returns 403, and the role is server-authoritative so a client can't escalate) — structured JSON logs (CODEINTEL_LOG_FORMAT=json) with optional per-request access logs, graceful SIGTERM shutdown, a bounded connection pool, and a non-root Dockerfile with a healthcheck.

Full guide → docs/deploy.md: systemd, Docker / Compose, Kubernetes (liveness + readiness probes, token from a Secret), reverse-proxy TLS, RBAC + SSO-via-auth-proxy, a Prometheus scrape config, and a security checklist.

docker build -t codeintel . && docker run -p 127.0.0.1:8766:8766 \
  -e CODEINTEL_HTTP_TOKEN="$(openssl rand -hex 32)" codeintel

Project status

An honest picture, so you can decide what to trust this with.

What's solid. The suite is large and real — more test code than source, a coverage floor enforced in CI, and fault-injection tests behind the never-raise contract. CI runs lint, mypy, and the full suite on Python 3.11/3.12/3.13, then builds the wheel, installs it into a clean environment, and runs a release canary that registers the build with Codex and Claude Code in a throwaway HOME, boots the server those configs name, and asserts on the answer text of a real code.query. A release that installs into a file no host reads, or that boots and answers nothing, fails before it ships. The code.query envelope has been stable since 0.8.

What's young. The project is pre-1.0 and moves fast. The honest signal is in the CHANGELOG: 0.15.0 and 0.15.1 were written almost entirely from pointing the tool at four repositories it had never seen, and several of those defects had survived multiple adversarial review rounds. The rate at which new codebases surface new bugs has not yet flattened. Expect to be the first person to hit something, and please report it — see below.

Use it for. Local, single-user code intelligence on a developer machine. That is the designed case, it is the tested case, and the blast radius of a wrong answer is a wasted tool call: the safe-null contract means a failing engine degrades to null with a reason, so your agent falls back to grep rather than crashing.

Be careful with.

On the test numbers. The suite is large and the coverage floor is enforced, but read the figure with its caveat: the graph and LSP backends are external binaries that are not installed in CI, so those two engines are exercised against hand-authored mocks rather than the real wire contract, and the release canary — which does assert on real answer text against a built wheel — currently covers the semantic engine only. Line coverage measures how much of the intended behavior runs, not how much of reality it has met.

The honest one-paragraph version. codeintel has been run on very few repositories its author did not write, and that is where its bugs have come from — every fix in 0.15.x came from pointing it at an unfamiliar codebase. Its characteristic failure mode is answering confidently from the wrong index rather than failing loudly, which the never-raise contract makes harder to notice: a wrong answer and a right one are the same shape. Run codeintel doctor before trusting a repo-wide answer, treat deadcode as candidates for review, and if something looks off please report it — an issue from someone who is not the author is the single most useful thing this project can receive right now.

Engine coverage depends on external binaries. Semantic search works out of the box. The graph engine needs codebase-memory-mcp and the LSP engine needs uvx on PATH — without them those engines safe-null and you get a fraction of the capability table above. codeintel setup --all installs what it can and codeintel doctor tells you exactly what is missing and how to fix it. Run doctor first if the tool seems quieter than the docs suggest.

Maintenance. One maintainer, MIT licensed, issues and PRs welcome. There is no support guarantee — factor that into anything load-bearing.

Reporting a problem

codeintel doctor --json prints a complete, machine-readable picture of what's installed, what's runnable, and whether this repo is indexed — per engine, with the remediation for each gap. Paste it into an issue and the report is actionable immediately instead of needing a round trip:

codeintel doctor --json

It reports only local engine and index state. Over the HTTP transport the registrations field — which names agent config files on the machine running the server — is deliberately omitted.

If a result looks wrong rather than a command failing, send the envelope rather than the rendered text:

codeintel query --op callers --target yourSymbol --json

engine says which engine answered, cached whether it came from the cache, reindexing whether the index was mid-rebuild, and reason/hint why an empty answer was empty. Those five fields are usually the whole diagnosis.

Development

git clone https://github.com/hamilton-sky/codeintel.git
cd codeintel
pip install -e .[dev]

pytest tests/ -q            # ~620 tests, ~30s; fails under 83% coverage
ruff check src tests        # lint
mypy                        # types (src/ only)

Your local run is not CI's run. A dev machine usually has codebase-memory-mcp and uvx installed; CI has neither, so the live graph/LSP tests skip there and the never-raise envelopes take different reason/hint paths. A bug reachable only on the no-backend path passes at your desk and fails in CI. To see CI's shape before you push:

env PATH="$(dirname "$(which python)"):/usr/bin:/bin" pytest -q

Release gate. The unit suite runs against the source tree, so it cannot see a packaging break, a missing entry point, a host config written where nobody reads it, or a server that boots and answers nothing. Run the canary against the built wheel in a clean environment — the same check CI runs before publishing:

python -m build && python -m venv /tmp/canary && /tmp/canary/bin/python -m pip install dist/*.whl && /tmp/canary/bin/python scripts/release_canary.py

It exits non-zero on the first failed check and cleans up the temporary HOME it installs into.