The best MCP servers for searching a local codebase in 2026

TL;DR — There is no single best local code-search MCP server; there are four jobs. Serena owns symbol navigation and refactoring. Semble and cocoindex-code own embedding-based search with no API key (XERJ's default embedder is lexical feature hashing). ast-grep MCP owns structural patterns. XERJ owns one index over code and everything beside it. No head-to-head was run.

Pick one in ten seconds

• Pick Serena if the agent must find references, rename symbols or edit at the symbol level. Language servers do that; nothing else here does.
• Pick Semble if you want embedding-based code search on a CPU with no API key and an installer that wires itself into Claude Code, Codex or OpenCode.
• Pick cocoindex-code if you want the same with an AST-based index and a plugin for Claude Code or Grok, and can afford its roughly 1 GB of local embedding dependencies.
• Pick Claude Context if you already run Milvus or Zilliz Cloud and are fine sending code to an embedding API.
• Pick ast-grep MCP if the question is a syntax pattern ("every call to unwrap() inside a loop"), not a topic.
• Pick codedb if you want a single binary with callers and dependency graphs and accept alpha status. XERJ has no call graph.
• Pick XERJ if the agent asks about code and the docs, logs, JSONL, CSV or SQLite files next to it, through one MCP server that also stores memory.

Summary table

Serena: the one with an IDE behind it

Serena wraps language servers and exposes symbol-level retrieval, editing and refactoring as MCP tools. Its README claims support for over 40 languages through that backend, and a JetBrains plugin backend for more.

That is a capability XERJ does not have. XERJ records where a symbol is defined, has no call graph, and does not edit. If the agent's next step is a rename across files, Serena is the right tool.

Semble and cocoindex-code: embedding-based search, no key

Semble chunks files with tree-sitter and scores chunks with a static embedding model that runs on a CPU. It needs no API key, GPU or external service; the model downloads from Hugging Face on first use. Its README claims about 99 % fewer tokens than grep plus read, measured by the vendor on about 1,250 queries over 63 repositories. We did not reproduce that.

cocoindex-code builds an AST-based index. Its full install pulls in sentence-transformers so local embeddings work with no key; the slim install requires a cloud embedding provider.

Both beat XERJ's default path on synonym-style questions, because XERJ's default embedder is lexical feature hashing. XERJ needs --embed-mode neural for that.

Claude Context and codedb: hosted embeddings by default

Claude Context's documented setup asks for a vector database (Milvus or Zilliz Cloud) and an OpenAI API key for embeddings. It is a good fit if that infrastructure already exists; it is not an offline tool.

codedb is a single binary that registers itself in many MCP clients. Its README marks it alpha and says the default hybrid retrieval uses hosted embeddings, with semantic=local for local-only retrieval.

ast-grep MCP: patterns, not topics

The ast-grep MCP server is experimental and exposes four tools, including one that dumps a syntax tree so the agent can write a working pattern. It builds no index, so it never goes stale, and it needs ast-grep and uv installed.

Use it beside an index, not instead of one. Choose ast-grep when the question is a pattern rather than a topic. The XERJ vs ast-grep page splits those jobs in detail.

XERJ: one index over the whole folder

XERJ is a search engine first. xerj autoindex reads a repository plus the docs, logs, JSONL, CSV, PDFs and SQLite files around it, writes a defs field and a symbols array for 36 tree-sitter languages, and xerj mcp serves 17 tools over that node, memory included.

What it gives up: no references, no edits, a lexical default embedder, single-node only, and heavy memory while indexing. Our own JSONL benchmark measured the node peaking at about 9.1 GiB resident while indexing a 9.7 MB file. For code-only questions, compare it with ripgrep first: on 30 disclosed tasks ripgrep won 5 outright (XERJ vs ripgrep). Choose ripgrep when the agent searches a tree once.

How this page was checked

Every licence, install line and capability above was read from the project's own README or docs on 2026-10-19. Stars and latest releases came from the GitHub REST API the same day. Nothing was benchmarked head to head, and every vendor performance number is labelled as the vendor's.

FAQ

Which MCP server for local code search needs no API key?

Serena, Semble, cocoindex-code with its full install, ast-grep MCP and XERJ all run without an API key. Claude Context's documented setup needs a vector database and an OpenAI key for embeddings, and codedb's default hybrid retrieval uses hosted embeddings unless you choose semantic=local.

Which one is best for finding references and renaming symbols?

Serena. It drives real language servers, so it can find references and perform symbol-level edits. XERJ extracts definitions only and rewrites nothing, and the embedding-based servers return passages, not reference graphs.

Which one searches by meaning offline?

Semble and cocoindex-code ship local embedding models. XERJ's default embedder is lexical feature hashing; embedding-based retrieval needs the opt-in --embed-mode neural, which downloads a small model once.

When does XERJ beat the code-only servers?

When the agent's questions span code and the files around it: docs, JSONL, CSV, logs, PDFs and SQLite. XERJ indexes all of them into one Elasticsearch-compatible node behind one MCP server, and the same node stores agent memory.

Were these servers benchmarked against each other?

No. This page compares documented capabilities and GitHub data checked on 2026-10-19. No head-to-head retrieval benchmark was run, and vendor token-saving claims are quoted as vendor claims.

How were stars and releases collected?

From the GitHub REST API on 2026-10-19: stargazers_count from the repository endpoint and the latest release from the releases/latest endpoint. A server with no GitHub release shows none.

Evidence

• Serena provides semantic code retrieval, editing and refactoring tools through language servers, supports over 40 programming languages with its language-server backend, and installs with uv tool install -p 3.13 serena-agent. — https://github.com/oraios/serena
• Claude Context needs a vector database (Milvus or Zilliz Cloud) and, in its documented configuration, an OpenAI API key for the embedding model; the MCP server runs with npx @zilliz/claude-context-mcp@latest. — https://github.com/zilliztech/claude-context
• Semble runs on CPU with no API keys, GPU or external services, downloads its embedding model from Hugging Face on first use, chunks files with tree-sitter, and exposes a search MCP tool. — https://github.com/MinishLab/semble
• cocoindex-code's full install pulls in sentence-transformers for local embeddings with no API key; the slim install requires a cloud embedding provider and an API key. — https://github.com/cocoindex-io/cocoindex-code
• codedb is alpha; structural search runs locally, and its default hybrid retrieval uses hosted embeddings unless semantic=local is chosen. — https://github.com/justrach/codedb
• ast-grep MCP is an experimental server exposing four tools and requires ast-grep and uv to be installed. — https://github.com/ast-grep/ast-grep-mcp
• ripgrep searches files directly with no index; on one 331-file tree it won 5 of 30 disclosed tasks outright against the XERJ index. — https://github.com/BurntSushi/ripgrep
• XERJ's tree-sitter registry covers 36 languages and extracts definitions; xerj mcp served 17 tools in a live tools/list. — engine/crates/xerj-autoindex/src/extract/code.rs
• In the JSONL benchmark of 2026-10-19 the XERJ node peaked at 9,578,548 kB (about 9.1 GiB) resident while indexing one 9.7 MB file of 50,000 records. — benchmarks/best-of-jsonl-search-2026/README.md
• Semble's README claims about 99 % fewer tokens than grep plus read, measured by the vendor on about 1,250 queries over 63 repositories. — https://github.com/MinishLab/semble

Related

• I need my agent to find a passage and cite the file, not dump the whole tree. How?
• Local code search MCP for Claude Code that uses fewer tokens than grep
• The best grep and ripgrep alternatives for AI coding agents in 2026
• Is ripgrep enough for a coding agent, or do I need a code index?
• ast-grep vs a code index for an agent?
• How do I stop my agent from reading the whole repo into context?
• How do I audit a PHP codebase with an index instead of grepping 1,400 files? · same intent: code search for agents
• My codebase indexer says indexed but I don't see my code. How do I check it actually finished? · same intent: code search for agents
• How do I set up reference coding so the agent retrieves a mechanism before writing code? · same intent: code search for agents