Search-Mesh

Search-Mesh is a Rust-native codebase intelligence service for autonomous coding agents.

The goal is to reduce agent latency and token waste by moving repository search, syntax-aware filtering, context extraction, and precise patching into a local MCP-compatible process.

Status

This repository is in early setup. The current focus is a small, correct MVP:

• A Rust workspace with separate core and MCP crates.
• A JSON-RPC over stdio server implementing the MCP initialize lifecycle handshake, suitable for Claude Code, OpenCode, and other MCP-compatible agents.
• A first search tool, scan, backed by multi-keyword scanning.
• A syntax probe tool, ast_probe, backed by tree-sitter for Rust, Python, JavaScript, and TypeScript.
• A context extraction tool, squeeze, that returns AST-bounded source blocks.
• A patch tool, patch, that applies precise line/column edits and reports syntax validity.

SIMD acceleration, tree-sitter verification, semantic squeezing, and atomic patching are planned phases, not current guarantees.

Intended Tools

• scan: scan target directories for multiple keywords in one pass.
• ast_probe: validate raw hits against syntax tree node types.
• squeeze: return the smallest useful AST-bounded code block around a hit.
• patch: apply byte-offset edits and verify syntax after mutation.

See docs/usage.md for local examples and docs/mcp-protocol.md for the draft protocol.

Install

Claude Code Plugin (Recommended)

This repository is a self-hosted Claude Code plugin marketplace. It bundles the search-mesh MCP server registration and the agent skill together, so /plugin install wires up both in one step:

/plugin marketplace add jamestkelly/search-mesh
/plugin install search-mesh@search-mesh

On macOS and Linux, the plugin downloads and verifies the correct prebuilt search-mesh-mcp binary automatically the first time it's used — no separate install step. Windows isn't supported by the automatic installer yet; use cargo install search-mesh-mcp or a manual binary install there.

Manual Install

For OpenCode, or Claude Code configured manually rather than through the plugin, search-mesh-mcp needs to be on your PATH first.

search-mesh-mcp is published to crates.io and as prebuilt GitHub Release binaries.

Via cargo:

cargo install search-mesh-mcp

Via prebuilt binary: download search-mesh-mcp-<target>.tar.gz for your platform (macOS arm64, macOS x64, or Linux x64) from the Releases page, extract it, and put search-mesh-mcp on your PATH.

Configure Claude Code Manually

Add to .mcp.json at your project root:

{
  "mcpServers": {
    "search-mesh": {
      "command": "search-mesh-mcp",
      "args": []
    }
  }
}

Configure OpenCode

Add to opencode.jsonc:

{
  "mcp": {
    "search-mesh": {
      "type": "local",
      "command": ["search-mesh-mcp"],
      "enabled": true
    }
  }
}

Agent Skill

The Claude Code plugin install above already includes this skill. To install it manually for other agents/setups, skills/search-mesh/SKILL.md teaches an agent when to prefer scan, ast_probe, squeeze, and patch over shell tools like grep, cat, and sed. Copy it into your own skills directory:

mkdir -p ~/.claude/skills/search-mesh
cp skills/search-mesh/SKILL.md ~/.claude/skills/search-mesh/SKILL.md

Repository Layout

.claude-plugin/
  plugin.json       Claude Code plugin manifest.
  marketplace.json  Self-hosted marketplace catalog listing this plugin.
.mcp.json           Bundled MCP server registration (used by the plugin).
crates/
  search-mesh-core/   Core search, parsing, squeezing, and patching logic.
  search-mesh-mcp/    JSON-RPC/MCP stdio server.
docs/
  architecture.md     System shape and phased design.
  mcp-protocol.md     Tool schemas and response shapes.
  roadmap.md          Near-term implementation plan.
  usage.md            Local MCP usage examples.
examples/
  scan-request.jsonl  Example newline-delimited JSON-RPC requests.
  initialize-request.jsonl
  ast-probe-request.jsonl
  squeeze-request.jsonl
  patch-request.jsonl
  patch-target.txt
scripts/
  search-mesh-mcp-launcher.sh  Installs search-mesh-mcp on first use (macOS/Linux), then execs it.
skills/
  search-mesh/SKILL.md  Agent skill teaching when to use these tools.

Development

Install Rust with rustup, then run:

just check

Equivalent commands:

cargo fmt --all -- --check
cargo clippy --workspace --all-targets -- -D warnings
cargo test --workspace

CI And Releases

Pull requests and pushes to main run the Review workflow:

• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace
• MCP JSONL example smoke tests

Releases are prepared with release-plz from Conventional Commits:

• release-plz opens a release PR that bumps Cargo.toml/Cargo.lock versions and updates the changelog.
• Merging that PR to main creates a git tag and GitHub Release per changed package, and publishes both crates to crates.io.
• When a search-mesh-mcp-v* release is published, the Release Binaries workflow builds and attaches search-mesh-mcp binaries (plus a .sha256 checksum file for each) for macOS (arm64, x64) and Linux (x64) as release assets.

Design Principles

• Prefer the smallest correct implementation before optimization.
• Keep protocol handling separate from core repository intelligence.
• Measure performance before adding specialized acceleration paths.
• Avoid .unwrap() and .expect() outside tests.
• Return agent-friendly payloads instead of whole files when possible.

License

MIT. See LICENSE.