Pick the right action in milliseconds. Know when to abstain.

// any planner, any language:
const decision = await fetch("http://localhost:8100/v1/decide", {
  method: "POST",
  headers: { "content-type": "application/json" },
  body: JSON.stringify({
    state: { app: "Finder", context: { goal: "delete the old folder" } },
    candidate_actions: [
      { id: "a1", kind: "click", description: "Delete button" },
      { id: "a2", kind: "click", description: "Cancel" },
    ],
  }),
}).then((r) => r.json());

// -> { "action_id": "a1", "confidence": 0.75,
//      "abstained": false, "latency_ms": 0.4 }

# MCP: system1_decide / system1_verify in any MCP planner
{
  "mcpServers": {
    "system1": {
      "command": "/opt/system1/.venv/bin/system1",
      "args": ["mcp"]
    }
  }
}

# or speak the SystemOne API everyone already ships:
POST /v1/systemone/choice   # noul / choice / score

git clone https://github.com/speer-ai/system1
cd system1
pip install -e ".[dev]"

# zero-GPU baseline policy, HTTP + MCP + bench:
system1 serve --port 8100  # SystemOne-compatible API
system1 mcp                # MCP stdio server
system1 bench points.jsonl # score decision points

# score a benchmark file: one JSONL record = one decision point
{"state": {"app": "Finder", "context": {"goal": "delete folder"}},
 "candidate_actions": [
   {"id": "a1", "kind": "click", "description": "Delete button"},
   {"id": "a2", "kind": "click", "description": "Cancel"}],
 "gold_action_id": "a1"}

# -> top-1 accuracy, risk-coverage curves, ECE,
#    p50/p95 latency per hardware tier