What is the best MCP server for code search?

No single MCP server is best for every code search. The choice depends on the kind of search you need (by meaning, by symbol, by structure or by keyword), where the embeddings and the index run, how the index stays fresh, the license and your client. For local semantic search without an API key, cocoindex-code is one option.

Checked Markdown version

6 criteria separate these servers

  • Search type. There are 4 kinds: by meaning (code from a description), by symbol (definitions and references), by structure (syntax patterns) and by keyword (exact text). The table below shows which server does which.
  • Chunking. A semantic index splits files into chunks before it embeds them. Chunks that follow function and class boundaries return a whole function when it fits in one chunk. Fixed-size windows can cut a function in half.
  • Freshness. Code changes while the agent works. Check what triggers a re-index and how much work each re-index repeats.
  • Local or hosted embeddings. A local model keeps the code on your machine and needs no API key. A hosted model takes less local compute and sends code to a provider.
  • License. Check whether your company allows copyleft (GPL) tools and whether the server needs a paid plan. Of the 6 servers below, 4 are MIT or Apache 2.0.
  • Client support. A stdio server runs as a local process under the client. A remote server needs a client that speaks MCP over HTTP.

How 6 servers compare

Checked on November 5, 2026 against each project’s README or documentation, linked in the first column. The table lists documented features and holds no benchmark results.

ServerSearch typeIndex and embeddingsFreshnessLicense
cocoindex-codeBy meaning, over Tree-sitter chunksLocal SentenceTransformers model with no API key, or a cloud provider through LiteLLM. The index is embedded in the project folder, with no database to set up.Re-indexes changed files. The MCP search tool refreshes the index before each query by default.Apache 2.0
Claude ContextHybrid: BM25 plus dense vectors, over AST chunksEmbeddings from OpenAI, VoyageAI, Ollama or Gemini. Vectors are stored in Milvus or Zilliz Cloud.Re-indexes changed files, found with Merkle trees.MIT
SerenaBy symbol: find symbol, find references, renameLanguage servers (LSP) for more than 40 languages, or a paid JetBrains plugin. It uses no embeddings.Not applicable: it keeps no embedding index.GPL-3.0-or-later for the application
ast-grep MCPBy structure: AST patterns and YAML rulesIt runs the ast-grep CLI on the files and uses no embeddings.Not applicable: it keeps no embedding index.MIT
GitHub MCP ServerBy keyword, through GitHub code search (the search_code tool)GitHub’s hosted index. The server is remote (hosted by GitHub) or local.Covers code that is on GitHub, not local edits.MIT
Sourcegraph MCP ServerKeyword search, natural-language search, go to definition, find referencesYour Sourcegraph instance.Set by the Sourcegraph instance.Commercial, on Enterprise plans

The ast-grep MCP README calls that server experimental.

Where cocoindex-code fits

cocoindex-code fits when you want search by meaning over a repository on your own machine:

  • Local by default. The [full] install embeds with a local model (Snowflake/snowflake-arctic-embed-xs), so the code is not sent to an embedding provider. The tool sends anonymous usage telemetry that excludes code, paths and queries. COCOINDEX_DISABLE_USAGE_TRACKING=1 turns it off.
  • Setup. It needs no vector database. A background daemon starts on first use.
  • Syntax-aware chunks. The tool tries to split files at function and class boundaries. Its README lists 33 supported languages.
  • Incremental updates. A changed file is re-indexed, and chunks whose text did not change keep their embeddings.
  • Clients. The README documents Claude Code, Codex, OpenCode, Kilo Code, Grok and Oh My Pi. Other MCP clients that launch a stdio server can start ccc mcp from inside the project. An agent skill is the recommended route for Claude Code.

Where it does not fit

  • Symbol navigation. Its MCP server has one tool, search. For callers, references and renames, use Serena, Sourcegraph, or CocoIndex Code Plus (ccx), which is a separate self-hosted product, licensed per deployment.
  • Exact keyword or regex search. The client’s built-in search, or a hybrid index such as Claude Context, matches exact text.
  • Code hosted elsewhere. To search an organization’s repositories, use the GitHub MCP Server, Sourcegraph or CocoIndex Code Plus.
  • Remote MCP. ccc mcp runs in stdio mode. A client that only connects to remote servers needs a different server.
  • Small installs. The [full] extra brings about 1 GB of local model dependencies. The slim install avoids them but needs a LiteLLM embedding provider: a cloud API with its key, or a local server such as Ollama.

Install cocoindex-code

bash
pipx install 'cocoindex-code[full]'
# or with uv:
uv tool install --upgrade 'cocoindex-code[full]'

cd your-repo
ccc index
bash
# Claude Code: skill (recommended in the README)
npx skills add cocoindex-io/cocoindex-code

# Claude Code: MCP
claude mcp add cocoindex-code -- ccc mcp

# OpenAI Codex CLI
codex mcp add cocoindex-code -- ccc mcp

Cursor and other MCP clients that launch a stdio server take the generic entry. The client must start ccc mcp with the project as its working directory:

json
{
  "mcpServers": {
    "cocoindex-code": {
      "command": "ccc",
      "args": ["mcp"]
    }
  }
}

To compare servers on your own code, write 5 questions you asked your agent this week. Run each one against every server you consider, and check whether the right function is in the first 5 results.

Sources and further reading