arbor

Fit your entire codebase into an LLM's context window.

Quick Start • Why arbor • Configuration • Tools • Performance • Languages

Highlights

• 1M lines of code → 500 lines of context. arbor builds a symbol graph with tree-sitter and compresses it into token-efficient summaries an LLM can actually use.
• 9 surgical MCP tools. The LLM sees architecture first, then drills into exactly what it needs — no grep noise, no wasted tokens.
• Sub-second incremental re-index. Only changed files are re-analyzed via content hashing. Cold index of a 1M LOC project takes under 10 seconds.
• 15 languages and formats. Rust, Python, TypeScript, Go, C/C++, C#, Kotlin, plus Terraform, Ansible, SQL, Protobuf, OpenAPI, and Markdown.
• Zero configuration. One install command. No config files. Works with any project structure.

bevy (1,756 files, 21,863 functions, ~1.1M LOC)
  boot screen:      16 lines   ~400 tokens
  compact skeleton:  552 lines  ~9k tokens
  indexed in:        9.5 seconds

Quick Start

One command — installs arbor and connects it to Claude Code:

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/nikita-voronoy/arbor/main/scripts/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/nikita-voronoy/arbor/main/scripts/install.ps1 | iex

# Build from source
cargo install --git https://github.com/nikita-voronoy/arbor.git arbor-mcp

# Add to Claude Code
claude mcp add arbor -- arbor

That's it. Claude will call boot → compact → search → references as needed.

CLI mode

arbor /path/to/project --cli       # Architecture overview
arbor /path/to/project --compact   # Token-optimized skeleton

Configuration Examples

The installer configures everything automatically, but here's what it sets up and how to customize it.

The installer registers arbor as an MCP server:

claude mcp add arbor -- arbor

Verify it's registered:

claude mcp list

The installer adds a hook to ~/.claude/settings.json that nudges Claude to use arbor instead of raw grep/glob:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Grep|Glob",
        "hooks": [
          {
            "type": "command",
            "command": "echo '{\"hookSpecificOutput\":{\"hookEventName\":\"PreToolUse\",\"additionalContext\":\"STOP: Prefer arbor MCP tools (search, references, skeleton, compact, boot) over Grep/Glob for code navigation. Fall back to Grep/Glob only for string literals, comments, or regex patterns.\"}}'",
            "statusMessage": "Checking arbor preference..."
          }
        ]
      }
    ]
  }
}

This doesn't block grep — it adds context that helps Claude choose the right tool.

The installer appends a block to ~/.claude/CLAUDE.md that teaches Claude when to use each arbor tool:

## Code navigation: use arbor MCP first

- **Instead of grep for a symbol** → use `mcp__arbor__search`
- **Instead of grep for "who calls X"** → use `mcp__arbor__references`
- **Instead of reading many files** → use `mcp__arbor__boot`, then `mcp__arbor__skeleton` or `mcp__arbor__compact`
- **Instead of tracing dependencies** → use `mcp__arbor__dependencies` or `mcp__arbor__impact`
- **After making changes** → call `mcp__arbor__reindex`

Start every session with `mcp__arbor__boot`.

You can edit ~/.claude/CLAUDE.md to fine-tune this behavior. For project-specific instructions, add a CLAUDE.md in the project root.

arbor auto-detects the project root from the working directory. For monorepos with multiple languages, it indexes all detected facets automatically:

# Index from repo root — detects Rust + TypeScript + Terraform + Markdown
arbor /path/to/monorepo --compact

For separate repos that share types, use the tunnels tool to discover cross-project connections.

arbor works through Claude Code's IDE extensions. After installing arbor:

1. Install the Claude Code extension for your IDE
2. arbor is automatically available — Claude will use boot and compact to understand your project

No additional IDE configuration needed.

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/nikita-voronoy/arbor/main/scripts/uninstall.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/nikita-voronoy/arbor/main/scripts/uninstall.ps1 | iex

This removes the binary, MCP registration, hooks, and CLAUDE.md instructions.

How It Works

flowchart LR
    Source[".rs .py .ts .go\n.c .cpp .cs\n.tf .yml .sql"] -- "tree-sitter\nparse" --> Graph["Symbol graph\nfunctions • structs\ncalls • imports"]
    Graph -- "query" --> Tools["MCP tools\nboot • compact\nsearch • refs • impact"]
    Graph -- "persist" --> DB[".arbor/index.bin\n(incremental)"]

1. Index — tree-sitter parses source files into ASTs. arbor extracts functions, structs, traits, enums, calls, imports, and type references.
2. Persist — the graph is saved to .arbor/. On re-index, only changed files are re-analyzed (xxh3 content hashing).
3. Serve — 9 MCP tools let the LLM explore the graph at any granularity.
4. Resolve — cross-file call edges are resolved in a second pass after all files are indexed.

MCP Tools

Performance

Tested on real-world projects (Apple Silicon, parallel parsing with rayon):

Incremental re-index (only changed files) is typically <100ms.

arbor's MCP tools return structured, compressed output — dramatically fewer tokens than raw grep + file reads for the same information.

Supported Languages

graph TB
    subgraph arbor-mcp["arbor-mcp"]
        MCP["MCP server<br>(rmcp over stdio)"]
        CLI["CLI entry point"]
        H["9 tool handlers"]
    end

    subgraph arbor-analyzers["arbor-analyzers"]
        TS["tree-sitter<br>10 languages"]
        IAC["Ansible / Terraform"]
        SCH["SQL / Protobuf / OpenAPI"]
        DOC["Markdown"]
    end

    subgraph arbor-core["arbor-core"]
        G["Graph<br>(Node, EdgeKind)"]
        Q["Query engine<br>search / refs / impact"]
        SK["Skeleton<br>boot / compact"]
    end

    subgraph arbor-persist["arbor-persist"]
        ST["Store (bincode)"]
        FH["FileHashes (xxh3)"]
    end

    DET["arbor-detect<br>Facet detection"]

    MCP --> H
    H --> Q
    H --> SK
    arbor-analyzers --> G
    DET --> arbor-analyzers
    G --> ST
    FH --> arbor-analyzers

License

MIT — see LICENSE.

Built with tree-sitter and MCP