All your real context,
in one local lake.
contextlake mirrors your repositories, indexes them into a knowledge graph, and serves it to your AI tools over MCP, so agents answer from real source instead of guessing.
View on PyPIYour AI is only as good as what it can see.
Point it at one file and it's sharp. Ask it about the system across dozens of repos, and it starts guessing.
contextlake gives your tools the real source to read: mirrored to your machine, indexed into a queryable graph, and served to your editor over MCP. Everything runs locally and offline, so no code leaves your machine.
Three layers, adopted one at a time
The mirror is useful on its own. Each layer above it is optional, turn on only what you need.
contextlakeMirror
Clone every repo in a GitLab group, GitHub org, Bitbucket workspace, or Codeberg owner into a faithful mirror, each parked on its most active branch and kept fresh with one command, never touching the branch you're on.
Knowledge optional
Parse the mirror into a code and dependency graph, then add semantic search, a council-verified wiki, and Atlassian / Figma / GitLab connectors.
Serve
Expose it all over MCP and an offline graph visualizer, so agents answer “where is X defined?” instead of grepping.
A whole codebase as one navigable graph
contextlake kb graph
renders a bounded, offline, interactive map, fleet overview, a symbol's neighbourhood, or a single repo. Type glyphs,
language lettermarks, confidence-coded edges, level-of-detail labels, and a navigator minimap keep even dense graphs legible.
The real visualizer, running right here on contextlake's own code. Drag, zoom, and search. Vendored and offline, no network. Export to Mermaid, DOT, or JSON.
Built for real working machines
- OllamaEmbedder
- BuiltinEmbedder
- OpenAIEmbedder
A real blast_radius call, answered from the graph inside your editor. No model, no network.
Offline-first
The core tool is stdlib-only; nothing leaves your machine. The knowledge layer is opt-in and runs locally, so your source and your context stay on your disk.
- 0.62kb/embeddings/store.py class VectorStore
- 0.61kb/embeddings/base.py embed
- 0.57kb/embeddings/base.py build_embedder
- 0.57kb/embeddings/builtin.py embed
- 0.56kb/embeddings/base.py _build_builtin_embedder
Real cosine-ranked hits over contextlake's own code. The built-in CPU model needs no key and no network.
Semantic search →A real sync run over a local test fleet. Uncommitted work is detected and left untouched, and the same pass scales to hundreds of repos, concurrent and resumable.
Syncing →Curated wiki
LLM-synthesized, council-verified pages grounded strictly in graph facts, with provenance footers.
Editor steering
Generate AGENTS.md, .mcp.json, rules, and a skills library, wired to your graph.
When you reach for contextlake
The same graph answers a handful of everyday questions that are painful across many repos.
Land in an unfamiliar fleet
New to a codebase spread across dozens of repos? Ask where a symbol is defined, what calls it, and how repos connect, without cloning your way through it by hand.
$ contextlake kb query "RateLimiter"Ground your AI editor
Point Claude Code, Cursor, Windsurf, or Kiro at the graph over MCP so they answer from your real source with cited file:line, instead of guessing from training data.
$ contextlake kb serveCheck blast radius before a change
See what depends on a class or module across every repo before you touch it, so a small edit does not quietly break a service three repos away.
$ contextlake kb impact EmbedderFind who knows it
Surface the likely owners and reviewers for a repo or path from commit history, so you know who to ask before a change lands.
$ contextlake kb owners billing-apigrep, or a hosted vector database?
vs. grep across repos
grep finds strings, one repo at a time. It cannot tell a definition from a call, cannot cross repo boundaries, and buries the answer in matches. contextlake parses code into a typed graph, so you ask for callers, dependents, or owners and get cited file:line answers across the whole fleet.
vs. a hosted vector DB
A hosted vector database means standing up infra and usually shipping your code to a third party. contextlake is offline and graph-first: exact edges instead of fuzzy nearness, token-cheap for agents, and semantic vectors are an opt-in local layer, not a dependency. Your source and your context stay on your disk.
vs. other local code-graph tools
Tools like Graphify and GitNexus index a codebase, on demand, one at a time. contextlake mirrors and continuously keeps fresh a whole fleet of repos, on their real branches, with a graph, search, and a wiki that stay in sync across all of them, automatically.
vs. AI wiki generators (DeepWiki)
DeepWiki turns a single public repo into an AI-written wiki, on demand, in the cloud, one at a time, and your own code needs a paid account to go through it privately. contextlake mirrors your whole fleet locally and keeps a graph, search, and wiki offline and current across every repo you have, nothing leaves your disk to get there.
Otters keep a favorite stone. Pebble keeps yours.
Every repo you care about, mirrored to your own lake and kept on its most active branch, so the context is always within reach for you and your agents. No stone goes missing, and the work you have not committed yet is never dropped.
How branch-safety works →Up and running in a minute
No GitLab or config needed to try it on a repo you already have. Point it at any local git repo and open the graph.
Open source, and small on purpose
contextlake is MIT-licensed and deliberately focused. Small, sharp tools live or die by their edges, so bug reports, fixes, and well-scoped features are all welcome.
- Keep it lean. One job, done well; features that stray from it are a hard sell.
- No network in tests, and every change ships with the test that proves it.
- Conventional Commits, atomic and well-scoped, with a clear why.