model-radar

MCP server that pings free coding LLM models across HTTPS providers and subscription CLIs (Claude Code, Grok, Antigravity/agy, Codex), ranks them by latency, and helps AI agents pick the fastest available model — or pin several subscriptions for a parallel review.

Inspired by free-coding-models.

Install

pip install model-radar-mcp

Related MCP server: beacon-mcp

Quick Start

1. Configure an API key

# Option A: Save to ~/.model-radar/config.json
model-radar configure nvidia nvapi-xxx

# Option B: Environment variable
export NVIDIA_API_KEY=nvapi-xxx

Or copy the template: cp config.example.json ~/.model-radar/config.json and edit it.

2. Add to your MCP client

Claude Code (~/.claude/settings.json):

{
  "mcpServers": {
    "model-radar": {
      "command": "model-radar",
      "args": ["serve"]
    }
  }
}

Cursor (~/.cursor/mcp.json):

Stdio (Cursor starts the server):

{
  "mcpServers": {
    "model-radar": {
      "command": "/path/to/your/.venv/bin/model-radar",
      "args": ["serve"]
    }
  }
}

Streamable HTTP (persistent server — recommended):

{
  "mcpServers": {
    "model-radar": {
      "url": "http://127.0.0.1:8743/mcp",
      "transportType": "streamable-http"
    }
  }
}

Start the server first:

model-radar serve --transport sse --port 8743

OpenClaw (~/.openclaw/config/mcporter.json):

{
  "mcpServers": {
    "model-radar": {
      "type": "http",
      "url": "http://127.0.0.1:8743/mcp"
    }
  }
}

Web dashboard: Add --web for a localhost UI at http://127.0.0.1:8743/ for status, config, discovery, and running prompts. The server binds to 127.0.0.1 only; keys never leave your machine.

model-radar serve --transport sse --port 8743 --web

Auto-restart wrapper:

while true; do model-radar serve --transport sse --port 8743; sleep 1; done

Then call restart_server() from any MCP client to reload with updated code.

3. CLI usage

# Scan models
model-radar scan --min-tier S --limit 10

# List providers
model-radar providers

# Save a key
model-radar configure nvidia nvapi-xxx

Catalogs are live

Model ids are not a hardcoded list. On startup, once an hour, and after a completion 404, model-radar fetches each provider’s /v1/models (Ollama /api/tags, grok models / agy models) and replaces that provider’s catalog — new ids in, retired ids gone. GET /v1/models is free; completions are what you pay for.

Seed tuples in the package are a fallback plus SWE-bench overlays for known ids. See Catalog playbook.

model-radar db refresh              # force live replace
python scripts/catalog-report.py    # seed vs live vs missing keys (no secrets)

Providers

HTTPS providers take an API key (configure_key or env). Call list_providers() for the current count and key status.

CLI subscriptions

If you already pay for a monthly plan, model-radar can ride that subscription — no API key. The official CLI is auto-detected from $PATH at startup.

The old gemini CLI was deprecated (June 2026) in favor of Antigravity CLI (agy). Install: curl -fsSL https://antigravity.google/cli/install.sh | bash. agy models may also list Claude and GPT-OSS on the same login. Codex-in-agy is a conversation mode; for model-radar use the standalone codex CLI.

These never join get_fastest() / default ask() — that would spend quota by accident. Pin them:

ask(prompt="Review this paragraph…", providers=["claude", "grok", "gemini"])
ask(prompt="…", model_ids=["sonnet", "grok-4.6"])

MCP Tools

Discovery

• list_providers() — See all providers, API-key status, and installed subscription CLIs

• list_models(tier?, provider?, min_tier?, free_only?) — Browse the catalog (refreshes a provider if its list is older than an hour)

• scan(verify?) — Ping models in parallel, ranked by latency. verify=True checks for non-empty output.

• get_fastest(min_tier?, count?, free_only?, verified?) — Best N models right now

• get_workers(count?, min_tier?, verified?) — N verified-alive models from N distinct providers

• provider_status() — Per-provider health check

Execution

• run(prompt, model_id?, free_only?) — Execute on fastest model with auto-fallback

• ask(prompt, count=3, model_ids?, providers?) — Same prompt on N models (Ollama sequential, remotes parallel)

• recommend(job) — Short diverse lineup for translate / rewrite / review / code / dict

• quality_probe(job) — Time + pass/fail on a fixed prompt (dict = Paper B five headwords)

• still_free(speed?) — Which Lane A hosts still answer; up to 3 chat models each in parallel (speed=fast prefers small/flash ids)

• batch_run(prompts, results_file?) — Batch execution with incremental JSONL, resume support, adaptive concurrency

Evaluation (LLM-as-Judge)

• judge(prompt, rubric, count=3, exclude_providers?) — Rate a single item with N diverse judges (pass the producer to exclude)

• compare(item_a, item_b, blind=True) — Blind A/B comparison, randomized order per judge

• batch_judge(items, rubric, results_file?) — Evaluate at scale with incremental results

• backtranslate_eval(..., exclude_providers?) — Back-translation quality metric; do not use the producer

Quality & Setup

• benchmark(model_id?) — Quality-test with 5 coding challenges

• refresh_models() — Fetch live lists and replace each provider’s catalog (purge retired ids)

• setup_guide(provider?) — Setup instructions for unconfigured providers

• configure_key(provider, api_key) — Save an API key

• restart_server() — Restart for code updates (SSE mode)

• server_stats() — Uptime and start time

Tier Scale (SWE-bench Verified)

Documentation

• Architecture — Module map, live catalogs, transport, rate limiting

• MCP Transport — Transport options, stateless HTTP, client configuration

• Catalog playbook — Live vs seed, TTL, purge, 404 refetch

• Local MCP ops — systemd restart, keys, compare runs

• Translation Pipeline Playbook — Batch translation patterns

• LLM-as-Judge Playbook — Evaluation patterns and judge selection

License

MIT