Git-native structural memory for AI coding agents. Index once — stays fresh on every pull, merge, and branch switch. MCP server for Claude Code, Cursor & any agent.
View on GitHub ↗A git-native, self-updating structural memory for your codebase. Index once, stay fresh forever, and let any LLM agent retrieve exactly the code context it needs over MCP — function-level chunks under a token budget instead of grep-and-read-the-whole-file.
Benchmarks with receipts: see EVALLOG.md — including the target we missed.
- Package & imports: codesherpa (PyPI release imminent — until then, install from GitHub) - Command line & MCP server: sherpa · index lives in .sherpa/
That's it. init installs post-commit/post-checkout/post-merge/ post-rewrite hooks that run sherpa sync --quiet, so the index follows your HEAD automatically. First run downloads the embedding model (~0.5 GB, one-time) to ~/.cache/sherpa/ and prints progress; after that everything is fully local and offline. Try it from the shell with sherpa search "where is the retry logic for http requests".
Agents burn most of their context window navigating: grepping, opening whole files, re-reading. sherpa gives them seven MCP tools that answer structurally:
tool what it answers ------ searchcode(query, budgettokens) hybrid lexical+semantic+symbol search, packed to a token budget getdefinition(symbol) jump to a definition with signature + breadcrumb getcallers(symbol) / getreferences(symbol) ranked callers / references (same-package first) getrecentchanges(since) commits + symbol-level diffs since a ref or ISO date expand(expandid) full body of any compact result, on demand indexstatus() freshness, counts, warming state
From the project README.
Add the radar badge to your README — it shows your project was picked up by MCP Radar and links to this page:
[](https://mcp.liqiwa.com/s/farhankhwaja--codesherpa.html)
Local-first MCP context engine for AI coding agents — cuts token cost ~90% via a per-project code graph + code folding. Works with Cursor, Claude Code, Codex & any MCP client. Web + ML/PyTorch codebas
bybit-exchange/kaasTurn scattered notes, docs and transcripts into a queryable Markdown wiki — an LLM knowledge-base compiler with MCP access, no embeddings, self-hosted.
TedHaley/pensieveLocal search across your files, git repos and AI agent sessions on your Mac. An MCP server gives your agents the best context, a 3D map shows your work by topic, and git blame shows who knows what.
nassim-arifette/jevgrepJev-powered semantic code search for coding agents — find behavior across repositories via CLI or MCP, with exact source excerpts and line numbers.
naw103/foremergeCatch intent conflicts before code conflicts. The open-source coordination protocol for coding agents, built above Git.
bar181/bar-observatoryDeterministic, local-only audit reports for Claude Code AI agent sessions. Rust + SQLite. Zero network calls. MCP server for agents.
The top new MCP servers of the week, every Monday. No spam, unsubscribe anytime.