convez.ai
MCP ServerDevOps & InfraestruturaIntermediário
atlassian-labs

MCP Compressor

atlassian-labs/mcp-compressor

Comprime schemas MCP para economizar tokens em agentes IA

Ver no GitHub
BomPluginScore · de 0 a 100

Servidor MCP que reduz o consumo de tokens ao comprimir descrições e schemas de ferramentas. Funciona como proxy ou biblioteca embarcada em Python, TypeScript e Rust, permitindo que agentes usem servidores MCP grandes sem desperdiçar contexto em metadados.

Instalação

$npx convez-ai atlassian-labs/mcp-compressor

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
143
Forks
14
Acompanhando
1
Saúde
75%
Último commit
há 4 dias
Versão
v0.40.1
Issues abertas
6
Contribuidores
9

Casos de uso

  • Reduzir custos de tokens em sistemas com múltiplos MCP servers
  • Otimizar contexto de agentes IA que usam servidores MCP grandes
  • Implementar padrão lazy-loading de schemas de ferramentas

Veredito

Ferramenta prática e bem-mantida (Atlassian Labs) que resolve um problema real: tokens gastos em schemas nunca usados. Implementação sólida com suporte multilingue, mas requer integração técnica deliberada em arquiteturas de agentes.

README

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

mcp-compressor

Release Build status Commit activity License

mcp-compressor helps agents use large MCP servers without spending huge amounts of context on tool descriptions and schemas.

It can run as a CLI MCP proxy, or be embedded directly from Python, TypeScript, or Rust.

Why mcp-compressor?

A powerful MCP server can expose dozens or hundreds of tools. Each tool has a name, description, input schema, and sometimes large nested JSON Schema details. Sending all of that to a model up front can waste thousands of tokens before the agent has done any useful work.

For example, large public MCP servers can spend thousands or tens of thousands of tokens on tool descriptions alone. That creates three problems:

  • models spend context on tools they may never call,
  • cost increases for every request that carries the tool list,
  • adding multiple MCP servers quickly overwhelms practical context budgets.

mcp-compressor changes the interaction pattern: the model sees a small compressed surface first, asks for the full schema only for the selected tool, and then invokes that tool.

Pattern 1: compressed MCP proxy

Use this when you want any MCP client to see a smaller tool surface.

sequenceDiagram
    participant Client as MCP client / agent
    participant Proxy as mcp-compressor
    participant Backend as Backend MCP server

    Client->>Proxy: tools/list
    Proxy->>Backend: tools/list
    Backend-->>Proxy: all backend tools and schemas
    Proxy-->>Client: compact wrapper tools
    Client->>Proxy: get_tool_schema(tool_name)
    Proxy-->>Client: full schema for selected tool
    Client->>Proxy: invoke_tool(tool_name, input)
    Proxy->>Backend: tools/call selected backend tool
    Backend-->>Proxy: tool result
    Proxy-->>Client: result

The frontend usually exposes only:

  • get_tool_schema
  • invoke_tool
  • optionally list_tools at max compression

Pattern 2: local proxy for SDKs and generated clients

Use this when your application wants to embed compression directly and call tools from code or shell commands.

flowchart LR
    App[Python / TypeScript / Rust app]
    SDK[CompressorClient]
    Proxy[Local session proxy\n/token auth]
    Generated[Generated CLI / Python / TS clients]
    Backend[MCP backend servers]

    App --> SDK
    SDK --> Proxy
    Generated --> Proxy
    Proxy --> Backend

The SDK starts a local session proxy for generated clients. Generated clients call that proxy using a session token. Your app does not need to spawn a mcp-compressor stdio subprocess.

Features

Quick example

=== "CLI"

```bash
mcp-compressor -c medium -- python server.py
```

=== "Python"

```python
from mcp_compressor import CompressorClient

with CompressorClient(
    servers={"alpha": {"command": "python", "args": ["server.py"]}},
    compression_level="medium",
) as proxy:
    print([tool.name for tool in proxy.tools])
    print(proxy.invoke("echo", {"message": "hello"}))
```

=== "TypeScript"

```ts
import { CompressorClient } from "@atlassian/mcp-compressor";

const proxy = await new CompressorClient({
  servers: { alpha: { command: "python", args: ["server.py"] } },
  compressionLevel: "medium",
}).connect();

try {
  console.log(proxy.tools.map((tool) => tool.name));
  console.log(await proxy.invoke("echo", { message: "hello" }));
} finally {
  proxy.close();
}
```

=== "Rust"

```rust
use mcp_compressor::compression::CompressionLevel;
use mcp_compressor::sdk::{CompressorClient, ServerConfig};
use serde_json::json;

let proxy = CompressorClient::builder()
    .server("alpha", ServerConfig::command("python").arg("server.py"))
    .compression_level(CompressionLevel::Medium)
    .build()
    .connect()
    .await?;

let result = proxy.invoke("echo", json!({ "message": "hello" })).await?;
```

Generated clients

CLI Mode generates shell commands:

mcp-compressor --cli-mode --server-name atlassian -- https://mcp.atlassian.com/v1/mcp
atlassian get-accessible-atlassian-resources

Code Mode generates Python or TypeScript functions:

# generated-py/atlassian.py
import atlassian

resources = atlassian.getAccessibleAtlassianResources()
import { getAccessibleAtlassianResources } from "./generated-ts/atlassian.ts";

const resources = await getAccessibleAtlassianResources();

See Code Mode and generated clients for more detail.

Where to go next

  1. Install the package you need.
  2. Run the quickstart.
  3. Read how compression works.
  4. Choose between CLI usage, SDK usage, generated clients, and Just Bash.
Ver o original no GitHub

Bom para

Relacionados