convez.ai
MCP ServerAutomaçãoAvançado
basidiocarp

Rhizome

basidiocarp/rhizome

Inteligência de código estruturada para agentes IA via tree-sitter e LSP

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP de inteligência de código agnóstico de editor que fornece navegação em nível de símbolo, análise estrutural e refatoração para agentes de IA. Usa tree-sitter para análise offline rápida e LSP para semântica entre arquivos, eliminando leituras brutas de arquivos e buscas frágeis.

Instalação

$npx convez-ai basidiocarp/rhizome

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
2
Acompanhando
0
Saúde
50%
Último commit
há 4 meses
Versão
v0.14.1
Issues abertas
10
Contribuidores
2

Casos de uso

  • Navegação e extração de símbolos em projetos de código
  • Refatoração e edição estrutural automatizada
  • Análise de dependências e referências cruzadas entre arquivos

Veredito

Rhizome é uma solução robusta e bem arquitetada para agentes de IA que precisam de inteligência estrutural de código em vez de leitura bruta. Parte de um ecossistema maduro com separação clara de responsabilidades; recomendado para quem já trabalha com agentes de codificação complexos ou quer evitar LSP brittle.

README

A tradução para o português deste README ainda está na fila.

Rhizome

Editor-agnostic code intelligence MCP server. Gives AI agents symbol-level navigation, structure, and refactor surfaces instead of forcing them to read raw files.

Named after fungal rhizomorphs, root-like structures that spread through an environment and expose hidden pathways.

Part of the Basidiocarp ecosystem.


The Problem

Most MCP-compatible clients have no built-in code intelligence. They can read files, but they cannot navigate symbols, find references, inspect structure, or perform project-wide refactors without expensive raw reads and brittle search.

The Solution

Rhizome provides a standalone code-intelligence layer for any MCP client. Tree-sitter handles fast offline structure work, LSP takes over when you need cross-file semantics, and the MCP surface keeps the result compact enough to use in real sessions.


The Ecosystem

ToolPurpose
rhizomeCode intelligence via tree-sitter and LSP
capWeb dashboard for the ecosystem
cortinaLifecycle signal capture and session attribution
hyphaePersistent agent memory
lamellaSkills, hooks, and plugins for coding agents
myceliumToken-optimized command output
sporeShared transport and editor primitives
stipeEcosystem installer and manager
volvaExecution-host runtime layer

Boundary: rhizome owns code intelligence and structural editing tools. It does not own memory, shell filtering, lifecycle capture, or installation. Backend selection stays centralized in rhizome-core::BackendSelector; rhizome-mcp routes tools through that shared policy instead of inventing ad hoc LSP branches, and rhizome-lsp stays focused on live protocol clients and translation.


Quick Start

# Install via ecosystem installer
curl -fsSL https://raw.githubusercontent.com/basidiocarp/stipe/main/install.sh | sh
stipe setup

Or build from source:

# Build
cargo build --release

# Recommended: full ecosystem setup
stipe init

# Alternative: rhizome-only config guidance
rhizome init
# Inspect status and servers
rhizome status
rhizome lsp status
rhizome lsp install python

# Use directly
rhizome symbols src/main.rs
rhizome structure src/lib.rs

How It Works

MCP client             Rhizome                      Backend
──────────             ───────                      ───────
tool call        ─►    backend selector      ─►    tree-sitter
deeper query     ─►    capability check      ─►    LSP if needed
edit or export   ─►    tool handler          ─►    project-aware result
  1. Receive MCP requests: navigation, diagnostics, editing, and export calls arrive through the MCP server.
  2. Select a backend: choose tree-sitter, generic fallback, or LSP based on the tool and language.
  3. Run project-aware analysis: resolve structure, symbols, references, or diagnostics.
  4. Return structured results: emit compact machine-readable data instead of raw file dumps.
  5. Export knowledge: send code graph data to Hyphae when requested.

Supported Languages

TierLanguagesHow it works
Full query patterns10 languagesLanguage-specific tree-sitter queries
Generic fallback4 languages in the default buildGeneric AST walker over common node types
Optional grammar pack3 niche languageslang-all enables C#, Swift, and Haskell tree-sitter support
LSP only14 or more languagesLanguage server required

All 31 supported languages have LSP server configs, and 20 or more have auto-install recipes. The default binary keeps the heaviest niche grammars out of the shipped tree-sitter set; C#, Swift, and Haskell stay available through LSP unless you opt into rhizome-treesitter/lang-all.


What Rhizome Owns

  • Symbol navigation and structure queries
  • Tree-sitter and LSP backend orchestration
  • Safe structural editing tools
  • Code graph export to Hyphae

What Rhizome Does Not Own

  • Long-term memory and retrieval: handled by hyphae
  • Token filtering: handled by mycelium
  • Lifecycle signal capture: handled by cortina
  • Installation and host registration: handled by stipe

Key Features

  • Dual backend model: uses tree-sitter by default and upgrades to LSP when the task requires it.
  • MCP-first surface: works with Claude Code, Codex, Cursor, Continue, and other MCP clients.
  • Structural editing: exposes targeted edit tools instead of line-oriented shell patching.
  • Project-aware export: can ship code graph data into Hyphae memoirs.
  • Managed LSP support: includes status, install, and per-language config surfaces.

Architecture

rhizome/
├── rhizome-core        backend selection, config, installer
├── rhizome-treesitter  offline parsing and symbol queries
├── rhizome-lsp         LSP client backend
├── rhizome-mcp         MCP server and tool handlers
└── rhizome-cli         CLI entry point
rhizome serve                     start MCP server
rhizome symbols <file>            list symbols
rhizome structure <file>          show symbol tree
rhizome status                    show backend status
rhizome lsp install <language>    install an LSP server
rhizome export                    export code graph to Hyphae
rhizome export-understanding      export bounded repo understanding
rhizome summarize                 summarize project structure and entry points
rhizome doctor                    diagnose common installation issues

Configuration

# <platform config dir>/rhizome/config.toml

[languages.python]
server_binary = "ruff"
server_args = ["server"]

[lsp]
disable_download = false

Logging

Rhizome reads RHIZOME_LOG first, then falls back to RUST_LOG. If neither is set, it defaults to warn.

# General debugging
RHIZOME_LOG=debug rhizome serve

# Narrow to a noisy module
RHIZOME_LOG=rhizome_mcp::tools=debug rhizome serve

rhizome serve keeps stdout reserved for newline-delimited MCP JSON-RPC traffic. Logs go to stderr so they do not corrupt the transport.

Repo-Understanding Export Status

rhizome export-understanding --json and the MCP export_repo_understanding tool emit two separate status surfaces:

  • update_class: display-only summary label for humans
  • export_status: machine-facing orchestration contract

Use export_status for automation. It exposes:

  • outcome: complete_success, partial_success, cached_reuse, no_supported_files, or full_failure
  • refresh_kind: full_refresh, partial_refresh, cached_reuse, or no_refresh
  • any_exports_succeeded and any_exports_failed
  • safe_to_consume

Orchestration rule: treat partial_success as a degraded input even when an artifact exists. Only auto-consume repo understanding when safe_to_consume is true.

cached_reuse means Rhizome found supported files and reused the cached export without reprocessing them. no_supported_files is a bounded no-op result for paths that contain nothing exportable; do not treat it as equivalent to cached reuse.


Documentation

Development

# Fast test loop
cargo nextest run --workspace

# Targeted benchmark for tree-sitter symbol extraction
cargo bench -p rhizome-treesitter --bench parse_symbols

# Opt into every tree-sitter grammar when you need the niche language pack
cargo build --workspace -p rhizome-cli --features rhizome-treesitter/lang-all

# Full build and lint
cargo build --release
cargo clippy
cargo fmt

# Use command timing when you need to inspect a real command path
time cargo run -p rhizome-cli --bin rhizome -- symbols /absolute/path/to/file.rs

License

See LICENSE for details.

Ver o original no GitHub

Bom para

Relacionados