Code Atlas

Code Atlas builds a graph database of your entire codebase — code structure, documentation, and dependencies — and exposes it as 23 MCP tools that AI coding agents can query to understand, navigate, and reason about your code. It combines three search modalities in one system: graph traversal (who calls this function?), semantic search (find code by meaning), and BM25 keyword search (exact name/message matches). All powered by Memgraph as the backend, with a SQLite fallback when Docker is not available.

The core problem it solves: every time an AI agent touches a codebase, it burns tokens figuring out where things are. On large projects, agents can spend 30–50% of their context window on orientation before writing a single line of code. Code Atlas externalizes the mental model a developer builds of a codebase — structure, meaning, names — as a queryable graph, so agents can orient in one tool call instead of grepping through five files.

Apache 2.0 licensed, self-hosted, MCP-compatible with Claude Code, Cursor, Windsurf, and any MCP client. Available on PyPI as code-atlas-mcp.

• Your AI coding agent is burning context on codebase orientation before it gets to the actual task.
• You have a monorepo with cross-project dependencies that agents struggle to navigate.
• You want agents to understand documentation alongside code (ADRs, READMEs, design docs linked to the code they describe).
• You need "blast radius" analysis — what breaks if I change this function — as an agent query.
• You want all of this self-hosted with no data leaving your machine and no additional API costs.

• Three search modes, one system. Graph traversal, semantic search, and BM25 keyword search are fused via reciprocal rank fusion (RRF) in the hybrid_search tool. The agent does not need to know which search mode to use — hybrid search auto-adjusts weights by query shape.
• AST-level incremental indexing. Only re-indexes entities that actually changed, not entire files. This makes re-indexing fast even on large codebases.
• Documentation as first-class. Markdown docs, ADRs, and READMEs are indexed with links to the code they describe. An agent can ask "what does the authentication documentation say about token refresh?" and get a grounded answer.
• Monorepo-native. Auto-detects sub-projects, tracks cross-project dependencies, supports scoped queries.
• Blast radius analysis. The blast_radius tool computes transitive closure of callers and callees — "what breaks if I change this" — with every hit reporting the path it was found via.
• Token-efficient by design. Tool definitions are self-documented so agents can one-shot any tool without calling a guide first. Search results are budget-aware — the tool prioritizes what matters most rather than dumping everything.
• Self-hosted, no API costs. Local embeddings via TEI (Text Embeddings Inference), no extra API keys. All intelligence runs through your existing agent subscription.
• Human-readable too. atlas ui serves the same graph as a local web interface with an architecture-health view (DSM, propagation cost, dependency cycles) for spotting codebase decay.

• Very small community. 6 GitHub stars as of September 2026. This is a single-developer project, not an ecosystem. Expect to read source code for edge cases and file issues yourself.
• Requires Python 3.14+. This is an unusually high floor. Most production environments are on 3.10–3.12. If your agent runtime does not have 3.14 available, you cannot use it without a separate environment.
• Memgraph dependency for full features. The SQLite fallback works but loses community detection (find_communities) and some analyses differ. You need Docker running with Memgraph for the full feature set.
• 23 MCP tools is a lot of context. The full tool definitions total 5,791 tokens. For agents with tight context budgets, this is a significant overhead cost just for registration. The "Search" subset (name + description only) is 2,360 tokens, which is more manageable but still substantial.
• No hosted option. Self-hosted only. If you do not want to run Docker and Memgraph, this is not for you.
• New project, untested at scale. The README mentions a ~1,400 entity test repo. Performance on codebases with 100K+ entities is unverified.
• No multi-language benchmark. Works on Python codebases. Support for other languages depends on AST parsing coverage, which is not documented.

Code Atlas is free and open source under Apache 2.0. You pay for:

• Compute infrastructure if you self-host: Docker, Memgraph, and TEI all run locally. A modest workstation is sufficient for the SQLite fallback; Memgraph wants Docker with adequate RAM.
• No additional API costs. Embeddings are local via TEI. No OpenAI or Anthropic API keys needed for the intelligence layer.
• Your agent's existing API costs are the only ongoing cost. Code Atlas reduces those by cutting orientation tokens, but it does not eliminate them.

Install: pip install code-atlas-mcp (requires Python 3.14+).