Athena-MCP π±
OmniRouter for AI-agent CLIs β one MCP server to detect, route, verify and economize across every AI CLI on your machine.
π§π· Leia em PortuguΓͺs Β· π¨π³ ι θ―»δΈζ
What is Athena-MCP?
Athena-MCP turns the AI CLIs installed on your machine (Codex, Claude Code, Cursor Agent, Antigravity, OpenCode, Ollama and more) into a single orchestrated council, exposed both as an MCP server (so any AI chat can delegate tasks) and a web dashboard for management.
Instead of betting on one agent, Athena:
- Detects every AI CLI on your system automatically (macOS, Windows, Linux)
- Routes each task to the best provider/model using an auto-updating ratings table
- Verifies execution reports against real project evidence β a "lie detector" that catches agents claiming work they never did
- Economizes by steering simple tasks away from expensive heavy models
- Fails over across providers with configurable combos
Features
| Feature | Description |
|---|---|
| π Cross-platform CLI scanner | Finds AI CLIs on PATH + ~/.local/bin, Homebrew, npm global, Scoop, cargo, go, flatpak, snapβ¦ 23 known CLIs listed even when not installed |
| π Dashboard | Providers, models, combos, usage stats and the "Best per Role" table at localhost:7100 |
| π Model ratings table | Scores 0β10 per role (frontend, backend, reasoning, speed), refreshed weekly from live leaderboards (SWE-bench, GPQA, Design Arena) |
| π΅οΈ Lie detector (verifier) | A cheap/free model (OpenCode free tier first) cross-checks each report against git evidence; false reports go back for fixing once, then escalate to the orchestrator |
π‘ recommend tool |
Describe a task, get the best installed provider+model with reasoning |
| π° Economy routing | Task complexity estimation; heavy models are excluded from suggestions for simple tasks |
| π Combos with failover | Chains of providers with retries, per-step models and timeout policies |
| π 10-topic report contract | Executors return lean structured reports β the orchestrator's context stays clean |
Quick start
git clone https://github.com/JoaoPauloNA/athena.git
cd athena
python -m venv .venv && source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -e .
Run the dashboard:
python -c "from athena.dashboard.app import run_dashboard; run_dashboard()" # API on :20129
cd frontend && npx vite --port 7100 # UI on :7100
Run as an MCP server (stdio), e.g. in your MCP client config:
{
"mcpServers": {
"athena": {
"command": "/path/to/.venv/bin/python",
"args": ["-m", "athena.mcp_server"],
"cwd": "/path/to/athena-mcp"
}
}
}
MCP tools
| Tool | Purpose |
|---|---|
list_providers |
Installed/available CLIs, models, roles, ratings |
ask_provider |
Send a task to one provider (verify=true enables the lie detector) |
run_combo |
Run a prompt through a failover chain |
deliberate |
Ask several agents in parallel |
recommend |
"Who should I call for this task?" β ratings Γ installed models Γ complexity |
refresh_models |
Re-scan CLI model catalogs |
list_usage |
Per-provider call counters and token estimates |
How verification works
Orchestrator β Executor CLI β 10-topic report
β
βΌ
Verifier (cheapest available model,
never the executor's own provider)
checks git status/diff + cited files
β
TRUE βββββββββ΄ββββββββ FALSE
β β
accepted sent back to executor (1x)
β
TRUE ββββββ΄ββββ FALSE again
β β
accepted ESCALATE to orchestrator
(verdict + evidence + history)
Documentation
- π Architecture Β· MCP tools reference
- π Arquitetura Β· ReferΓͺncia das ferramentas MCP
- π ζΆζ Β· MCP ε·₯ε ·εθ
Requirements
- Python β₯ 3.10
- At least one AI CLI installed (Codex, Claude Code, Cursor Agent, Antigravity, OpenCode, Ollamaβ¦)
- Node.js (only for the dashboard dev server)
License
MIT Β© 2026 JoΓ£o Paulo
