Zum Inhalt

asiai mcp

Start the MCP (Model Context Protocol) server, enabling AI agents to monitor and benchmark your inference infrastructure.

Usage

asiai mcp                          # stdio transport (Claude Code)
asiai mcp --transport sse          # SSE transport (network agents)
asiai mcp --transport sse --port 9000

Options

Option Description
--transport Transport protocol: stdio (default), sse, streamable-http
--host Bind address (default: 127.0.0.1)
--port Port for SSE/HTTP transport (default: 8900)
--register Opt-in registration with asiai agent network (anonymous)

Tools (11)

Tool Description Read-only
check_inference_health Quick health check: engines up/down, memory pressure, thermal, GPU Yes
get_inference_snapshot Full system snapshot with all metrics Yes
list_models List all loaded models across engines Yes
detect_engines Re-scan for inference engines Yes
run_benchmark Run a benchmark or cross-model comparison (rate-limited to 1/min) No
get_recommendations Hardware-aware engine/model recommendations Yes
diagnose Run diagnostic checks (like asiai doctor) Yes
get_metrics_history Query historical metrics (1-168 hours) Yes
get_benchmark_history Query past benchmark results with filters Yes
compare_engines Compare engine performance for a model with verdict; supports multi-model comparison from history Yes
refresh_engines Re-detect engines without restarting the server Yes

Resources (3)

Resource URI Description
System Status asiai://status Current system health (memory, thermal, GPU)
Models asiai://models All loaded models across engines
System Info asiai://system Hardware info (chip, RAM, cores, OS, uptime)

Claude Code integration

Add to your Claude Code MCP config (~/.claude/claude_desktop_config.json):

{
  "mcpServers": {
    "asiai": {
      "command": "asiai",
      "args": ["mcp"]
    }
  }
}

Then ask Claude: "Check my inference health" or "Compare Ollama vs LM Studio for qwen3.5".

Benchmark cards

The run_benchmark tool supports card generation via the card parameter. When card=true, a 1200x630 SVG benchmark card is generated and card_path is returned in the response.

{"tool": "run_benchmark", "arguments": {"model": "qwen3.5", "card": true}}

Cross-model comparison (mutually exclusive with model, max 8 slots):

{"tool": "run_benchmark", "arguments": {"compare": ["qwen3.5:4b", "deepseek-r1:7b"], "card": true}}

CLI equivalent for PNG + sharing:

asiai bench --quick --card --share    # Quick bench + card + share (~15s)

See the Benchmark Card page for details.

Agent registration

Join the asiai agent network to get community features (leaderboard, comparison, percentile stats):

asiai mcp --register                  # Register on first run, heartbeat on subsequent runs
asiai unregister                      # Remove local credentials

Registration is opt-in and anonymous — only hardware info (chip, RAM) and engine names are sent. No IP, hostname, or personal data is stored. Credentials are saved in ~/.local/share/asiai/agent.json (chmod 600).

On subsequent asiai mcp --register calls, a heartbeat is sent instead of re-registering. If the API is unreachable, the MCP server starts normally without registration.

Check your registration status with asiai version.

Network agents

For agents on other machines (e.g., monitoring a headless Mac Mini):

asiai mcp --transport sse --host 0.0.0.0 --port 8900

See the Agent Integration guide for detailed setup instructions.