LocaLLama MCP 服务器（测试分支）

LocalLama MCP Server

Local-first, provider-neutral Model Context Protocol server for coding-agent workflows. Routes tasks across local models (Ollama, LM Studio, llama.cpp), free OpenRouter models, and paid frontier models using cost, latency, context capacity, and benchmark history.

Node.js: >=22

⚠️ Early / experimental — not yet a stable release. This project is under active, rapid development and has not been fully verified end-to-end. MCP tool signatures, configuration, and behavior may change between releases without notice.

Version numbers follow SemVer mechanically (they're derived from Conventional Commit messages, not hand-picked), so a 1.x number signals only "a public surface exists" — it is not a promise of stability or completeness. If you depend on this server, pin to an exact version.

• Tagged releases on main are the relatively safer builds.
• The testing channel publishes bleeding-edge pre-releases (x.y.z-testing.n) for trying unproven changes early.

Overview

LocalLama MCP reduces token costs without sacrificing quality. Tasks are queued asynchronously — route_task returns a task_id immediately; callers poll get_task_status for results. The decision engine chooses local → free → paid based on measured provider capabilities and configurable thresholds.

Supported MCP clients: Codex, Claude Code, Claw Code, Cursor, GitHub Copilot Agent mode, and any generic MCP stdio client.

Requirements

• Node.js 22+
• npm
• At least one of: Ollama, LM Studio, llama.cpp server, or an OpenRouter API key

Installation

git clone https://github.com/Heratiki/locallama-mcp.git
cd locallama-mcp
npm install
npm run build

Configuration

Copy .env.example to .env and edit with your values. The server resolves .env from its own root directory (or LOCALLAMA_ROOT_DIR when set), not from the MCP host's CWD.

# Local LLM Endpoints
LM_STUDIO_ENDPOINT=http://localhost:1234/v1
OLLAMA_ENDPOINT=http://localhost:11434/api
# LLAMA_CPP_ENDPOINT=http://localhost:8080   # leave unset to disable

# Routing thresholds
DEFAULT_LOCAL_MODEL=qwen2.5-coder-3b-instruct
TOKEN_THRESHOLD=1500
COST_THRESHOLD=0.02
QUALITY_THRESHOLD=0.7

# Provider concurrency
PROVIDER_HEALTH_PROBE_INTERVAL_MS=60000
PROVIDER_MAX_CONCURRENT_LOCAL=1
PROVIDER_MAX_CONCURRENT_REMOTE=5
PROVIDER_TIMEOUT_MS=120000
OLLAMA_TIMEOUT=120

# Code search (native BM25, no Python required)
CODE_SEARCH_ENABLED=true
CODE_SEARCH_EXCLUDE_PATTERNS=["node_modules/**","dist/**",".git/**"]
CODE_SEARCH_INDEX_ON_START=true
CODE_SEARCH_REINDEX_INTERVAL=3600

# Benchmarks
BENCHMARK_RUNS_PER_TASK=3
BENCHMARK_PARALLEL=false
BENCHMARK_MAX_PARALLEL_TASKS=2
BENCHMARK_TASK_TIMEOUT=60000
BENCHMARK_SAVE_RESULTS=true
BENCHMARK_RESULTS_PATH=./benchmark-results
RELIABLE_BENCHMARK_COUNT=3
MIN_VALIDATOR_SCORE=0.6
VALIDATION_RETRY_BUDGET=1

# Lock file
LOCK_FILE_CHECK_ACTIVE_PROCESS=true
REMOVE_STALE_LOCK_FILES=true

# OpenRouter (optional)
OPENROUTER_API_KEY=your_openrouter_api_key_here
OPENROUTER_FREE_ONLY=false

# Logging
LOG_LEVEL=debug

# Operational testing
# EXPECT_LOCAL_PROVIDER_DOWN=true

Key environment variables

MCP Client Configuration

Build the server, then point your MCP client at node dist/index.js:

{
  "mcpServers": {
    "locallama": {
      "command": "node",
      "args": ["/path/to/locallama-mcp/dist/index.js"],
      "env": {
        "LM_STUDIO_ENDPOINT": "http://localhost:1234/v1",
        "OLLAMA_ENDPOINT": "http://localhost:11434/api",
        "DEFAULT_LOCAL_MODEL": "qwen2.5-coder-3b-instruct",
        "TOKEN_THRESHOLD": "1500",
        "COST_THRESHOLD": "0.02",
        "QUALITY_THRESHOLD": "0.07",
        "OPENROUTER_API_KEY": "your_openrouter_api_key_here"
      }
    }
  }
}

Claude Code users can place this in .mcp.json (project-scoped) or ~/.claude/settings.json (global).

Tools

Core tools (always available)

OpenRouter tools (require OPENROUTER_API_KEY)

Async task flow

route_task → { task_id }
                ↓ poll
get_task_status → { status: "pending" | "in_progress" | "completed" | "failed", result? }

When local providers are contended by benchmark workloads, route_task surfaces contention metadata:

{
  "task_id": "...",
  "status": "queued",
  "queue_position": 2,
  "benchmark_contention": {
    "local_slot_contended": true,
    "active_benchmark_runs": 1,
    "queued_benchmark_runs": 2,
    "message": "Local execution slot currently contended by benchmark workloads."
  }
}

Resources

Static resources

Resource templates

Usage

Starting the server

npm start

A lock file prevents multiple instances. Stale locks from crashed processes are detected and cleaned up automatically.

Running benchmarks

npm run benchmark
npm run benchmark:comprehensive

Results are stored in benchmark-results/ as JSON and Markdown summaries.

Dashboard

When the server is running, a web dashboard is available at http://localhost:3001 (server-local).

Features:

• Real-time job queue with status, provider/model, and queue position
• Task monitoring with per-job details and ETA
• Manual route_task submission form
• Task and job cancellation
• Benchmark history

REST API endpoints:

Example submission:

curl -X POST http://localhost:3001/api/tasks \
  -H "Content-Type: application/json" \
  -d '{"task": "Refactor parser for readability", "context_length": 4096, "complexity": 0.6, "priority": "quality"}'

Live monitoring metadata

When the JobTracker WebSocket server is running, task-executing tools include:

{
  "task_id": "task-123",
  "monitoring": {
    "websocketUrl": "ws://127.0.0.1:8081",
    "activeJobsUri": "locallama://jobs/active",
    "jobProgressUriTemplate": "locallama://jobs/progress/{jobId}",
    "note": "Connect to websocketUrl for live updates, or use MCP resources."
  }
}

websocketUrl is scope: server-local — in SSH/container/Codespaces/WSL setups, forward the port before connecting.

_server_reminder ambient metadata

Tools attach a _server_reminder field at most once every 30 minutes to surface monitoring info:

{
  "_server_reminder": {
    "schemaVersion": 1,
    "kind": "monitoring-reminder",
    "status": "reachable",
    "scope": "server-local",
    "message": "Optional monitoring available from MCP server host.",
    "monitoringUrl": "http://127.0.0.1:3001",
    "lastCheckedAt": 1747699200000
  }
}

Remote access

If your MCP client is not on the same machine as the server:

# SSH
ssh -L 8081:127.0.0.1:8081 -L 3001:127.0.0.1:3001 user@host

• Dev Containers / Codespaces: forward ports 8081 (WebSocket) and 3001 (dashboard) via the VS Code Ports view.
• WSL client + WSL server: use the WebSocket URL directly. Windows client + WSL server: forward port 8081 via VS Code or a local tunnel.

Provider integrations

Ollama

Set OLLAMA_ENDPOINT in .env. The server probes for available models on startup.

LM Studio

Set LM_STUDIO_ENDPOINT in .env. Exposes an OpenAI-compatible API.

llama.cpp (llama-server)

# Single model
llama-server -m /path/to/model.gguf --port 8080

# Router mode (multiple models)
llama-server --model /path/model1.gguf --model /path/model2.gguf --port 8080

Set LLAMA_CPP_ENDPOINT=http://localhost:8080 in .env. If the endpoint is unset or unreachable, the provider initialises silently — other providers are unaffected. The server does not manage the llama-server process lifecycle.

OpenRouter

Set OPENROUTER_API_KEY. The server fetches ~240 available models on startup (30+ free). Use clear_openrouter_tracking to force a refresh. Set OPENROUTER_FREE_ONLY=true to restrict to free-tier models.

Code search

Code search uses a native TypeScript BM25 engine — no Python or external dependencies required.

# Via MCP tool
retriv_init { "directories": ["/path/to/repo"], "force_reindex": true }
retriv_search { "query": "pagination logic" }

Development

npm run build        # compile TypeScript + copy assets
npm start            # run compiled server
npm run dev          # TypeScript watch mode
npm test             # build + run Jest (23 suites, 186 tests)
npm run lint         # ESLint (note: eslint-plugin-import not installed — lint currently fails)
npm run lint:fix     # ESLint with auto-fix

All test files mock server state to prevent multiple real instances during test runs.

Architecture

src/
  index.ts                        entry point, lock file, MCP lifecycle
  modules/
    api-integration/              tool definitions, resources, routing adapters
    decision-engine/              task analysis, model selection, coordination
    cost-monitor/                 token accounting, cost estimation
    benchmark/                    execution, scoring, summaries, DB storage
    lm-studio/                    LM Studio provider
    ollama/                       Ollama provider
    llama-cpp/                    llama-server provider
    openrouter/                   OpenRouter provider
    core/provider/                shared provider registry and execution queue
    updater/                      self-update logic (check_for_updates, update_server)
    job-store/                    persistent Task/Job store
    websocket-server/             live monitoring side channel

Decision engine uses two model data stores:

• ModelRegistry + CapabilityDetector: benchmark-derived capability scores (authoritative for full routing)
• modelsDbService: heuristic performance data seeded from ModelRegistry at startup; used by preemptiveRouting()

Project docs

Troubleshooting

Server won't start — lock file detected

1. Check if another instance is running (ps aux | grep locallama).
2. Stale locks from crashes are cleaned up automatically (REMOVE_STALE_LOCK_FILES=true).
3. If needed, manually remove locallama.lock from the project root.

OpenRouter models not appearing

Use clear_openrouter_tracking through the MCP interface to force a fresh fetch.

npm run lint fails

eslint-plugin-import is referenced in the config but not installed. Known issue. Build and tests are unaffected.

Security notes

• API keys belong in .env, which is excluded from version control.
• All log output goes to stderr; stdout is reserved for MCP JSON-RPC. Never write non-JSON to stdout.
• Treat MCP tools as model-controlled surfaces. Avoid mutations without user approval.

License

ISC