contextatlas

ContextAtlas is an MCP server that gives Claude Code a pre-computed "atlas" of a codebase, so architectural questions can be answered from one fused context bundle instead of many grep and file-read calls. What it does - At index time it extracts structured claims from ADRs, READMEs and design docs, docstrings of exported symbols, and filtered git commit messages, keyed to specific code symbols. - At query time it fuses four signals per symbol: LSP structural data (definition, references, types, diagnostics), architectural intent claims, git history (recent commits, hot/cold, co-change), and test associations. - Three MCP tools are exposed: get_symbol_context (full bundle for a known symbol, multi-symbol mode up to 10), find_by_intent (BM25 search over indexed claim text in local SQLite FTS5), and impact_of_change (blast radius using co-change and test impact). - Query resolution is local: SQLite lookup plus local language-server calls, with no network traffic or model calls during normal use. Language support TypeScript, Python, Go, Ruby and C# are supported through language servers (typescript-language-server, Pyright, gopls, ruby-lsp, csharp-ls), which are peer dependencies the user installs. Node.js 20+ is required. Two setup paths - Skills path: install the package, run init, then use /generate-adrs, /index-atlas and /prime-atlas inside Claude Code; no API key needed. - CLI path: install, set ANTHROPIC_API_KEY, run init, generate-adrs, index and doctor. Repos that already have ADRs under docs/adr/ can skip generate-adrs, since existing ADRs are extracted automatically. Committed atlas pattern The generated atlas.json is intended to be committed to the repository alongside code and ADRs. Teammates cloning the repo import it directly into a local cache without re-running extraction, and contributors can regenerate it when changes affect architectural claims. Teams that cannot commit it can set atlas.committed: false and each developer runs their own extraction. Privacy notes At index time, ADR/doc files matched by configuration, docstring text of exported symbols, and filtered commit messages are sent to the Anthropic API; source code, git history beyond those messages, LSP data and runtime query contents are not. A docs.include glob that matches source files would send those files whole. Extraction streams can be limited (for example to ADRs only). All extracted claims and query resolution stay in a local SQLite file. Reported benchmarks The README reports a benchmark suite over honojs/hono, encode/httpx and spf13/cobra with pre-registered prompts and blind grading, citing 45-72% token reduction on architectural-intent prompts and a hono reference run with 39% fewer tool calls in aggregate. It also reports paired quality deltas (factual correctness, hallucination, actionability, completeness) with confidence intervals, and states that tie and trick prompt buckets showed ContextAtlas as net-negative, which the authors present as expected. These figures come from the project's own benchmark repository and should be treated as author-reported, not independently verified. Status and licensing README states v0.9.0 shipped 2026-05-16 and that the package is not yet published to npm, so install instructions describe the intended shape. Licensed MIT.