cie — the only code graph that knows which tasks and tests actually implement your code.

Where this is going: vision.md — the far shore. The graph becomes the software; repositories become its cache. Software that can always explain itself. A compass, not a claim.

Code Insight Engine. No other surveyed code-graph tool can answer "which files implement this task, and are they tested?" as one query. Everything developer-facing lives in the wiki — architecture written from the codebase, and a how-to for every workflow. This README stays lean: the demo, the install, and how to help.

Every second is a real recorded session, nothing staged: this repo cloned from the public tag v0.1.4, cie index . in 1.9s (1,902 nodes · 6,581 edges · 4,169 calls), the README one-liner registering it with Claude Code (✔ Connected), then ONE question in plain words — about resolve_backend, the storage-selection rule — answered from cie's tools alone (the agent's built-ins were disabled for the take: callers → affected_by → test_map were its only path). The agent returned the 7 pinning tests with line numbers, including the test added for the explicit-auto bug fixed that same day. The GIF is edited for time only — content is never edited; the uncut sessions ship in the repo: resolve-backend-uncut.cast (the agent take) and setup-uncut.cast (index → register → connected). Full take/QC record: docs/demo/production-log.md.

Install

One-click, no clone, no Neo4j, no signup — install once from a release tag, then one command per project indexes it, registers cie with your MCP client (spawn-robust entry: absolute path, so GUI-launched clients find it), and writes the agent context files:

# once per machine (latest tag):
uv tool install "cie-mcp[mcp] @ git+https://github.com/kannamma-labs/cie.git@v0.1.5"

# per project — from inside the project:
cie index .        # ~1.9s on a 110-file repo
cie init .         # registers the client, writes AGENTS.md/CLAUDE.md

Already installed and prefer the client-side route?

claude mcp add cie -- $(command -v cie-mcp) /path/to/your/project --backend embedded --policy readonly

Plain pip works too:

pip install "cie-mcp[mcp]"   # core + MCP server (cie-mcp) — what most people want
pip install "cie-mcp[http]"  # + the HTTP tool-mount (cie/routes.py)

Package-name note (updated 2026-08-31, v0.1.1): the distribution ships as cie-mcp — the cie name on PyPI belongs to an unrelated project (cluster311/cie10, ICD-10 codes; pip install cie does NOT get you this tool — never did). Import package stays cie; console scripts stay cie and cie-mcp. GitHub installs are an equal alternative: pip install "cie-mcp[mcp] @ git+https://github.com/kannamma-labs/cie.git@v0.1.5".

Core dependencies: Pydantic v2, tree-sitter (+ Python/JS/TS/Java/Go/ Rust/C/C++/C# grammars), watchdog, Click, Rich; the Neo4j driver only when you use that backend. Python ≥ 3.10. Storage is auto-selected (serve .cie/graph.db when you indexed, else Neo4j — stated on stderr at startup, never silent).

More: serving to Cursor/Codex, multiple projects, Neo4j team mode, semantic search, HTTP, policies, troubleshooting — every workflow has a how-to in the wiki (start with install-and-serve).

Related MCP server: polycodegraph

Contributors wanted

cie is a small core with an outsized surface (135 tools, 9 languages, two storage backends, three front-ends) and a culture you can see in the commits: DoD = verified against the real environment, never written-only; misses get published, not hidden. The demo above was produced by dogfooding — and the dogfood measurement found (and fixed) a real product bug the same day. That's the working style.

Ways in, easiest first:

• Use it and report — index your repo, ask it impact questions, file what's wrong or what's missing. The troubleshoot how-to lists the known sharp edges honestly.

• A measured gap — the direct-calls TESTS heuristic shipped because a dogfood measurement showed 1 edge in a 308-test suite. Find a number like that, and the fix gets in.

• A first PR — start with the good first issue label (each names a safe entry-point module and acceptance criteria) or issues #17–#19.

• Docs — wiki how-tos count. If you wished a page existed, write it; if the wiki and code disagree, the code wins and the wiki gets a PR.

Dev setup is three commands (clone, uv venv + editable install, python -m pytest -q — 312 passing): the details, the conformance harness, and the honesty bar are in CONTRIBUTING.md and the wiki's contributing page. CONTRIBUTING.md's "Becoming a second maintainer" section is the path beyond a one-off PR.

License

cie is released under the MIT License.

By contributing, you agree your contributions are licensed under the same terms.