convez.ai
MCP ServerIntegração & APIsAvançado
Matrix-aas

Mercedes MCP — Consultor de Codificação ECU

Matrix-aas/mercedes-mcp

Decodifique e consulte configurações Mercedes ECU em 3 idiomas

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP que converte arquivos Vediamo CBF da Mercedes-Benz em conhecimento consultável por LLM, oferecendo busca multilíngue, decodificação de dumps hexadecimais e recomendações honestas (somente leitura). Ideal para diagnóstico, análise de configurações e compreensão de parâmetros de ECU sem capacidade de escrita no veículo.

Instalação

$npx convez-ai Matrix-aas/mercedes-mcp

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
1
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
v1.0.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Diagnóstico e troubleshooting de módulos Mercedes
  • Decodificação de dumps de codificação em linguagem natural
  • Busca multilíngue de fragmentos de código ECU (EN, DE, RU)
  • Referência técnica e aconselhamento sobre parâmetros de variantes

Veredito

Especialista de nicho bem executado para diagnóstico Mercedes. A busca híbrida (léxica + embeddings) e o enriquecimento multilíngue resolvem um problema real em comunidades de tuning. Honesto ao alertar sobre ambiguidades estruturais e limitar-se apenas à leitura — não é ferramenta de flashing, é referência consultiva.

README

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

mercedes-mcp

License: MIT Runtime: Bun MCP

English · Русский

An MCP server that turns Mercedes-Benz Vediamo CBF files into coding knowledge an LLM can actually reason over. Ask in plain language — in English, German or Russian — and it finds the right ECU coding fragment, explains what each value does, flags where a setting is ambiguous, and decodes a coding dump back into human-readable settings.

It ships with the parsed data, the embeddings and the enrichment already built, so it runs straight after bun install — no indexing step, no API key.

Scope (v1): knowledge and advice only. It reads and decodes in the diagnose direction. It does not write to the car and does not encode settings back to bytes. Think of it as an expert reference, not a flashing tool.

What it gives an LLM

Seven tools over stdio and Streamable HTTP:

ToolWhat it does
list_ecusEvery ECU/CBF in the library with variant and domain counts
list_variantsThe variant (coding family) names of one ECU
searchHybrid plain-language search across all CBFs — one ranked row per fragment
get_domainA whole coding block with all its fragments and value options
get_fragmentThe full card for one fragment: values, bit location, meaning, per-value glosses, enthusiast effect, risk, related fragments
suggest_for_intent"I want X" → the best fragment, a likely enable value, alternatives, and honest caveats
decode_codingA hex coding dump → the current human-readable value of every fragment

Under the hood:

  • Hybrid, cross-lingual search. MiniSearch lexical matching fused (RRF) with multilingual-e5 embeddings, so a Russian query reliably reaches a German label.
  • Enrichment overlay. Every fragment carries an LLM-generated literal meaning, DE/EN/RU synonyms and per-value glosses, plus a hand-curated layer of ground truth that always wins (the kind of tribal knowledge that lives on forums, not in the CBF).
  • Honest signals. Results self-describe their confidence as search relevance, not certainty of meaning, expose the raw retrieval signals, and flag structural ambiguity (a bit shared across qualifiers, or a meaning that varies by variant) so the model knows when to ask for the exact variant.

Quick start

Requires Bun ≥ 1.3.

git clone https://github.com/Matrix-aas/mercedes-mcp.git
cd mercedes-mcp
bun install
bun run mercedes-mcp serve --stdio

That's it — the CBF data, the e5 document vectors and the enrichment are all committed, so there is no build step. On the first search the small e5 model (~120 MB) downloads itself once into .models/; everything else is already in the repo.

HTTP transport instead of stdio:

bun run mercedes-mcp serve --http --port 8765

Wire it into an MCP client

Point any MCP client at the stdio server. For Claude Desktop / Cursor (mcpServers in the client config):

{
  "mcpServers": {
    "mercedes-mcp": {
      "command": "bun",
      "args": ["run", "mercedes-mcp", "serve", "--stdio"],
      "cwd": "/absolute/path/to/mercedes-mcp"
    }
  }
}

Claude Code, from the repo directory:

claude mcp add mercedes-mcp -- bun run mercedes-mcp serve --stdio

A worked example

"I want my exhaust to pop and crackle on a cold start."

search returns Emotionsstart and Sound_Design (MED177) at the top — for the English and the Russian phrasing. get_fragment on Emotionsstart returns the two values, the bit location, the meaning, and the enthusiast effect (exhaust pop on startup) with a risk note. suggest_for_intent proposes the enable value but attaches the curated caveat that overrun crackle actually comes from Sound_Design, and that Erweiterte_Schubabschaltung works against it. Paste a coding dump into decode_coding and it tells you which of those are currently on.

Architecture

A small monorepo; dependencies point one way: cbf-core ← knowledge ← server.

  • packages/cbf-core — parses the CBF binary into a normalized model (Ecu → Variant → CodingDomain → CodingFragment → Subfragment) and decodes coding bytes. Pure, no I/O. A focused TypeScript port of the C# Caesar library from CaesarSuite.
  • packages/knowledge — flattens the parsed model into searchable records, runs the hybrid search, and owns the versioned on-disk embedding cache and the enrichment overlay.
  • packages/server — the seven MCP tools over stdio + Streamable HTTP, and the mercedes-mcp CLI.

Flow: parseCbf → buildRecords → Library.search/getFragment → MCP tool JSON → LLM.

Development

bun test            # whole workspace — fully offline, deterministic, no model or network
bun run mercedes-mcp serve --stdio

Regenerating the committed artifacts (rarely needed — only if you add CBFs or change the model/prompt):

bun run mercedes-mcp build-index            # re-embed all fragments with the real model → .cache/
bun run mercedes-mcp build-index --enrich   # regenerate the LLM enrichment overlay (uses the codex CLI by default)

Environment variables are all optional and documented in .env.example; serving and tests need none of them.

Data, ownership & disclaimer

The CBF files and decoded labels describe Mercedes-Benz / Daimler control-unit coding and remain the property of their respective owners. This project is an independent, non-commercial tool with no affiliation to or endorsement by Mercedes-Benz Group AG. It is provided for educational and personal use.

Variant coding changes how your car behaves. Even though this tool only reads and advises, acting on that advice with a real coding tool is entirely at your own risk — verify against your own vehicle and variant before changing anything. No warranty; see LICENSE.

Credits & license

Released under the MIT License.

Ver o original no GitHub

Bom para

Relacionados