Local-first codebase intelligence MCP server that indexes repositories with tree-sitter, stores searchable chunks in DuckDB, and gives Claude semantic search, regex search, daemon status, and deep code research tools.
ChunkHound reads source files, Markdown, text, PDFs, and supported config files under the target directory and stores indexed chunks in a local database., Realtime indexing and daemon mode can continue watching project files after the initial MCP connection., Code research and web search tools require embedding, reranking, and LLM configuration and may invoke local CLIs or external model APIs depending on settings., Exclude generated files, vendored dependencies, secrets, large artifacts, and unrelated repositories before indexing broad workspace roots., Review MCP client configuration carefully when using an absolute project path in a global Claude Desktop config.
Privacy notes
Indexed chunks, file paths, symbols, comments, Markdown, PDFs, configuration values, database files, daemon state, and search results can reveal proprietary source code and internal architecture., Embedding, reranking, LLM, and web search providers may receive code-derived queries or snippets if configured., Local ChunkHound database files, logs, daemon state, and MCP transcripts may retain code-derived context after the session ends., Avoid sharing ChunkHound databases, config files with API keys, verbose logs, research outputs, and screenshots from private repositories.
Author
ChunkHound
Submitted by
oktofeesh1
Claim status
unclaimed
Last verified
2026-06-06
Decision playbook
Review trust signals before you adopt
Signals are present but mixed. Use the checklist below to confirm the source and operational safety for your environment.
Compare context
Selected
0
Current score
63
Baseline
—
Delta
No baseline selected
No major trust-signal divergence detected in the current selection.
Source and provenance checks
Needs review
Confirm ownership and provenance before trusting install instructions.
Source link availableRequired
Open the canonical repository and verify ownership.
Done
Source provenance statusRequired
Marked as source-backed.
Done
Metadata reviewed
No reviewed flag detected in metadata.
Pending
Safety and privacy checks
Complete
Validate risk disclosures before installation or API wiring.
Safety notes presentRequired
Review the listed safety guidance before running commands.
Done
Privacy notes presentRequired
Review data handling notes before connecting accounts or secrets.
Done
Trust level risk gateRequired
Trust level does not block evaluation.
Done
Package and install checks
Needs review
Check package metadata and artifact integrity signals.
Install payload available
Install or copy payload is available for review.
Done
Package verification flag
No package verification flag provided.
Pending
Checksum metadata
No checksum provided for downloaded artifact.
Pending
Compare-driven decision checks
Needs review
Use compare context to validate trade-offs before adoption.
Compare tray has multiple entries
Add at least one more entry to compare trust differences.
5 safety and 4 privacy notes across 4 risk areas. Review closely: credentials & tokens, third-party handling.
4 areas
SafetyLocal filesChunkHound reads source files, Markdown, text, PDFs, and supported config files under the target directory and stores indexed chunks in a local database.
SafetyLocal filesRealtime indexing and daemon mode can continue watching project files after the initial MCP connection.
SafetyGeneralCode research and web search tools require embedding, reranking, and LLM configuration and may invoke local CLIs or external model APIs depending on settings.
SafetyCredentials & tokensExclude generated files, vendored dependencies, secrets, large artifacts, and unrelated repositories before indexing broad workspace roots.
SafetyLocal filesReview MCP client configuration carefully when using an absolute project path in a global Claude Desktop config.
PrivacyLocal filesIndexed chunks, file paths, symbols, comments, Markdown, PDFs, configuration values, database files, daemon state, and search results can reveal proprietary source code and internal architecture.
PrivacyThird-party handlingEmbedding, reranking, LLM, and web search providers may receive code-derived queries or snippets if configured.
PrivacyCredentials & tokensLocal ChunkHound database files, logs, daemon state, and MCP transcripts may retain code-derived context after the session ends.
PrivacyCredentials & tokensAvoid sharing ChunkHound databases, config files with API keys, verbose logs, research outputs, and screenshots from private repositories.
Disclosure: MIT-licensed open-source codebase intelligence MCP server. Review provider configuration before enabling semantic search, LLM research, reranking, or web search features in sensitive repositories.
Safety notes
ChunkHound reads source files, Markdown, text, PDFs, and supported config files under the target directory and stores indexed chunks in a local database.
Realtime indexing and daemon mode can continue watching project files after the initial MCP connection.
Code research and web search tools require embedding, reranking, and LLM configuration and may invoke local CLIs or external model APIs depending on settings.
Exclude generated files, vendored dependencies, secrets, large artifacts, and unrelated repositories before indexing broad workspace roots.
Review MCP client configuration carefully when using an absolute project path in a global Claude Desktop config.
Privacy notes
Indexed chunks, file paths, symbols, comments, Markdown, PDFs, configuration values, database files, daemon state, and search results can reveal proprietary source code and internal architecture.
Embedding, reranking, LLM, and web search providers may receive code-derived queries or snippets if configured.
Local ChunkHound database files, logs, daemon state, and MCP transcripts may retain code-derived context after the session ends.
Avoid sharing ChunkHound databases, config files with API keys, verbose logs, research outputs, and screenshots from private repositories.
Prerequisites
Python 3.10 or newer and the `uv` package manager.
A local repository or workspace you are authorized to index.
ChunkHound JSON config reviewed for database path, excludes, embeddings, and LLM provider settings.
Optional embedding provider credentials for semantic search, or regex-only usage when no embedding key is configured.
Optional LLM provider or local CLI reviewed before enabling `code_research`.
Schema details
Install type
cli
Troubleshooting
No
Source repository stats
Scope
Source repo
Collection metadata
Estimated setup
15 minutes
Difficulty
advanced
Tool listing metadata
Disclosure
MIT-licensed open-source codebase intelligence MCP server. Review provider configuration before enabling semantic search, LLM research, reranking, or web search features in sensitive repositories.
ChunkHound MCP Server connects Claude to a local-first codebase intelligence
index. It parses repositories with tree-sitter, stores searchable chunks in
DuckDB, and exposes MCP tools for code search, daemon status, deep code
research, and optional web-backed research.
Use it when Claude needs targeted context from a large repository without
reading broad files repeatedly. Regex search can work without embedding keys,
while semantic search and research workflows depend on the embedding, reranking,
and LLM providers configured in .chunkhound.json or environment variables.
These sources were reviewed on 2026-06-06. Prefer the live repository,
CLI reference, PyPI metadata, README, license, package metadata, configuration
docs, MCP tool registry, stdio server, and MCP CLI command source for current
setup and behavior.
Features
Run a stdio MCP server with chunkhound mcp.
Index local repositories into a DuckDB-backed code intelligence database.
Parse many programming, configuration, Markdown, text, and PDF formats.
Search code with semantic search when embeddings are configured.
Search with regex when no embedding provider is available.
Use realtime indexing and daemon mode for active repositories.
Inspect daemon status through MCP.
Run code_research when embeddings, reranking, and an LLM provider are
configured.
Use an explicit project path only when the global MCP client configuration must
launch ChunkHound outside the repository directory.
Use Cases
Ask Claude to find authentication, routing, migration, or error-handling code
across a large repository.
Run regex searches against indexed files when semantic embeddings are not
configured.
Keep a local code index updated while switching branches.
Scope search to an approved repository instead of exposing a whole workspace.
Use code_research for higher-level architecture questions when the LLM and
reranking providers are approved.
Safety and Privacy
ChunkHound is local-first, but it still creates a derived index of source code.
Limit the target path, configure excludes, and verify what will be indexed
before connecting a model to private repositories. Keep .chunkhound databases,
logs, API keys, MCP transcripts, and research outputs out of public commits.
Semantic search, reranking, LLM research, and web search can involve external
providers depending on configuration. Review every provider endpoint, model, API
key, timeout, and custom base URL before enabling those features for sensitive
codebases.
Duplicate Check
No chunkhound/chunkhound, ChunkHound MCP Server, chunkhound mcp, or matching
source URL entry was found in content/mcp or README.md. Existing code index
and codebase memory entries cover different projects; this entry documents
ChunkHound's local-first tree-sitter, DuckDB, realtime indexing, and
code_research MCP workflow.
Show that ChunkHound MCP Server is listed on HeyClaude. Paste this Markdown into your README — it renders the badge and links back to this page.
[](https://heyclau.de/entry/mcp/chunkhound-mcp-server)
How it compares
ChunkHound MCP Server side by side with 3 alternatives on trust, install, platform support, and disclosed safety notes — all from reviewed registry metadata.
1 trust signal differ across this comparison (Submitter).
Local-first codebase intelligence MCP server that indexes repositories with tree-sitter, stores searchable chunks in DuckDB, and gives Claude semantic search, regex search, daemon status, and deep code research tools.
Local code indexing MCP server that lets AI assistants search repositories, build symbol indexes, inspect file summaries, monitor changes, and navigate multi-language codebases.
Code intelligence MCP server with a Zig core for local project indexing, structural outlines, symbol lookup, search, dependency graphs, snapshots, remote public-repo queries, and fallback edits.
Index any documentation — websites, GitHub repos, npm/PyPI packages, local files — and give Claude always-current, version-specific answers with the open-source Grounded Docs MCP server. An offline-first alternative to Context7, Nia, and Ref.Tools.
✓ChunkHound reads source files, Markdown, text, PDFs, and supported config files under the target directory and stores indexed chunks in a local database.
Realtime indexing and daemon mode can continue watching project files after the initial MCP connection.
Code research and web search tools require embedding, reranking, and LLM configuration and may invoke local CLIs or external model APIs depending on settings.
Exclude generated files, vendored dependencies, secrets, large artifacts, and unrelated repositories before indexing broad workspace roots.
Review MCP client configuration carefully when using an absolute project path in a global Claude Desktop config.
✓Code Index MCP reads repository files, builds shallow and deep indexes, exposes file summaries, and can return source snippets through search and file resources.
Point `--project-path` at one approved repository; avoid broad parent directories, home directories, generated dependency trees, or secret-heavy paths.
Deep indexing and watcher refreshes can consume CPU, memory, disk IO, and temp storage on large codebases.
The server can create and clear index/cache settings and temporary directories used for stored index data.
Search results, file summaries, and symbol bodies can surface credentials, private code, customer data, internal URLs, or proprietary implementation details if they exist in the repository.
✓CodeDB indexes local projects and exposes file tree, outline, search, symbol, caller, dependency, read, snapshot, project, and context tools to the MCP client.
CodeDB's `codedb_edit` tool exists as a fallback editing tool and can create, replace, insert, delete, or modify files when used by a client without native edit tooling.
The npm package runs a postinstall step that downloads a native binary from GitHub Releases; review package and release provenance in environments that restrict native binaries.
Remote repo queries use the public `api.wiki.codes` service and should be treated as network access outside the local repository.
The upstream README marks the project as alpha software, with parser coverage and snapshot formats still stabilizing.
✓The server stores its index locally, but MCP tool arguments and returned search/fetch results can be visible to your MCP client and model provider transcript.
The `scrape_documentation` and `fetch_url` tools can fetch URLs, GitHub repos, packages, or local paths — only index sources you are authorized to access and constrain local paths and network targets.
Privacy notes
✓Indexed chunks, file paths, symbols, comments, Markdown, PDFs, configuration values, database files, daemon state, and search results can reveal proprietary source code and internal architecture.
Embedding, reranking, LLM, and web search providers may receive code-derived queries or snippets if configured.
Local ChunkHound database files, logs, daemon state, and MCP transcripts may retain code-derived context after the session ends.
Avoid sharing ChunkHound databases, config files with API keys, verbose logs, research outputs, and screenshots from private repositories.
✓Tool calls may expose file paths, filenames, source code snippets, imports, classes, methods, functions, symbols, search matches, complexity summaries, and project settings.
Index data and temp directories can persist derived repository structure or symbol metadata outside the source tree.
File watcher status, exclude patterns, and settings can reveal local project layout and development conventions.
If an MCP client forwards tool results to a model provider, repository contents and search snippets may leave the local machine.
Exclude secrets, generated artifacts, vendored dependencies, build output, and private data before indexing.
✓Local indexes, snapshots, file trees, symbol names, dependency graphs, snippets, read results, and search results can reveal proprietary code structure and implementation details.
Upstream documents sensitive-file blocking for patterns such as environment files, credentials, and keys, but users should still review ignore rules and avoid indexing secret-heavy directories.
CodeDB writes telemetry to `~/.codedb/telemetry.ndjson` unless `CODEDB_NO_TELEMETRY=1` is set, then syncs aggregate tool counts, latency, startup, file count, line count, language, version, and platform data on MCP session close.
Upstream telemetry docs state that source code, file contents, file paths, and search queries are not collected.
Remote public-repo queries and local MCP responses may still be logged by MCP clients, model providers, and terminal history.
✓Indexed documentation, local file contents, URLs, search queries, extracted text, and search results may be exposed to Claude or another MCP client when tools are invoked.
Scraped content (HTML, Markdown, PDF, Office documents, source code) is stored in a local SQLite database on your machine; avoid indexing secrets, private repositories, internal URLs, or sensitive folders unless you intend them to be available through MCP results.
Prerequisites
Python 3.10 or newer and the `uv` package manager.
A local repository or workspace you are authorized to index.
ChunkHound JSON config reviewed for database path, excludes, embeddings, and LLM provider settings.
Optional embedding provider credentials for semantic search, or regex-only usage when no embedding key is configured.
Python 3.10 or newer.
uv or another Python package runner that can install `code-index-mcp`.
An approved repository path that the MCP client is allowed to inspect.
Optional native search tools such as ripgrep, ugrep, ag, or grep for faster search backends.
macOS or Linux on x64 or arm64 for the published native binary launcher.
Node.js 18 or newer when using the `codedeebee` npm launcher.
A local project directory that the MCP client exposes through roots or launches from.
Review of which local repositories, file types, generated artifacts, and secrets patterns may be indexed.
Node.js 22+ with `npx` available.
Start the server with `npx @arabold/docs-mcp-server@latest` before connecting — it listens at http://localhost:6280.
An MCP client such as Claude Code or Claude Desktop.