Gen-Zero API & MCP Updates

01 · Live Endpoint Status

This panel opens a real MCP session from your browser: GET /health, an SSE handshake on /sse, then initialize and tools/list. Nothing here is hard-coded. If a step fails, the panel says which step and why.

02 · Measured Figures, With Sources

We list a latency only when a run or a file backs it. Each row says where the number comes from and how far it can be trusted. Different rows come from different runs, so do not add them up.

Read before quoting. "0 generated tokens" is not "0 tokens processed": the semantic ask path still reads a prompt (20 prompt tokens in the sample below). No end-to-end speed or cost comparison against an LLM exists yet. "Zero hallucination" has not been measured, so we do not claim it.

03 · MCP Tool Reference

Schemas below are copied from the server's own tools/list answer. Responses are real calls to the public endpoint on 2026-10-01, trimmed only where a field holds 1024 floats. Call a tool with tools/call and {"name": ..., "arguments": {...}}.

{
  "action": "ask",
  "state": "disk is 95% full; nightly backup job pending",
  "candidates": [
    "delete /var/log/*",
    "rotate logs and compress",
    "rm -rf /"
  ]
}

{
  "isError": true,
  "content": [{ "type": "text", "text":
    "ConfirmationRequired: the semantic risk classifier rates this request as possibly dangerous (p=0.705 >= escalate 0.449)" }],
  "_meta": {
    "engine": "native_qwen",
    "best_action": "rotate logs and compress",
    "confidence": 0.98397,
    "tier": "Escalate",
    "gate_status": "requires_confirmation",
    "risk": {
      "p_dangerous": 0.70468,
      "thresholds": { "escalate": 0.4494, "hard_stop": 0.762 },
      "forward_ms": 1427.83
    },
    "scorer": { "prompt_tokens": 20, "forward_ms": 2278.27 }
  }
}

{
  "edges": [1, 2, 3],
  "genders": ["Male", "Male", "Male", "Male"],
  "axioms": [
    { "r1": 1, "r2": 2, "gender": "Male", "result": [7] },
    { "r1": 7, "r2": 3, "gender": "Male", "result": [9] }
  ]
}

{
  "isError": false,
  "content": [{ "type": "text",
    "text": "Concluded: relation 9 in 2 steps (strategy chart)" }],
  "_meta": {
    "engine": "relation_semiring_fold",
    "causal_fold": {
      "strategy": "chart", "edges": 3, "axioms": 2,
      "conflict_keys": 0, "steps": 2,
      "predicted": 9, "proof_path": [1, 2, 7, 3, 9]
    }
  }
}

{
  "edges": [1, 2],
  "genders": ["Male", "Female", "Male"]
}

{
  "isError": true,
  "content": [{ "type": "text", "text":
    "CausalFoldRefused: no bracketing of the chain closes under the learned table" }],
  "_meta": {
    "reject": {
      "code": "CausalFoldRefused",
      "stage": "causal_fold",
      "http_status": 422
    },
    "causal_fold": { "strategy": "chart", "edges": 2,
                     "axioms": 0, "step_failed": 2 }
  }
}

{
  "op": "audit_action",
  "state": [0.0, 0.0, 0.0],
  "action": 1,
  "horizon": 5
}

{
  "isError": false,
  "content": [{ "type": "text",
    "text": "audit_action: Approved (risk_score 0.0511)" }],
  "_meta": { "engine": "production_pipeline", "pipeline": {
    "op": "audit_action",
    "audit": {
      "verdict": "Approved",
      "gate_tier": "Tier0Proceed",
      "risk_score": 0.05105,
      "warn_risk": 0.3,
      "survival_horizon": 5,
      "first_hazard_step": null,
      "safety_calibrated": false,
      "reasons": [
        "no hazard in 5 step(s) of horizon 5; risk_score 0.0511 < warn_risk 0.3",
        "safety estimate is uncalibrated (latent_norm_boundary_margin): a model-internal margin, not an observed failure rate"
      ]
    } } }
}

04 · Fail-Closed Policy Gate

Every decision passes a four-tier gate before it is released. Unknown, non-finite or out-of-range input does not fall through to a default: it is refused with a typed code.

Tier definitions: /llms.txt. Thresholds: Paper 5, confirmed in the live _meta.risk.thresholds field above. Known gap, stated in Paper 5: a bare chmod -R 777 / scores below the escalate threshold, so this classifier is one layer of defence, not the only one.

05 · Client Install & Config

Pick your client. Each block has a copy button. Restart the client after you edit a config file.

claude mcp add --transport sse gen-zero https://api.gen-zero.ai/sse --header "Authorization: Bearer gz_public_free"

{
  "mcpServers": {
    "gen-zero": {
      "url": "https://api.gen-zero.ai/sse",
      "headers": {
        "Authorization": "Bearer gz_public_free"
      }
    }
  }
}

{
  "mcpServers": {
    "gen-zero": {
      "serverUrl": "https://api.gen-zero.ai/sse",
      "headers": {
        "Authorization": "Bearer gz_public_free"
      },
      "disabled": false
    }
  }
}

{
  "mcpServers": {
    "gen-zero": {
      "url": "https://api.gen-zero.ai/sse",
      "headers": {
        "Authorization": "Bearer gz_public_free"
      }
    }
  }
}

curl -N -H "Authorization: Bearer gz_public_free" https://api.gen-zero.ai/sse
# event: endpoint
# data: /message?session_id=<id>&token=gz_public_free

MSG="https://api.gen-zero.ai/message?session_id=<id>&token=gz_public_free"
curl -s -X POST "$MSG" -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"curl","version":"0"}}}'
curl -s -X POST "$MSG" -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","method":"notifications/initialized"}'
curl -s -X POST "$MSG" -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/list"}'

curl -i -N -m 3 "https://api.gen-zero.ai/sse?token=gz_public_free"
# HTTP/2 200 · content-type: text/event-stream · no token → HTTP 401

Codex CLI, self-hosted Rust (gen-zero serve --mode sse) and the Python package: /agent-install.md and /llms-full.txt §5.

06 · Release Log

The MCP server reports gen-zero 0.1.0, and that is the only engine release so far. The entries below are dated changes to the API surface and this site, each with its commit. A new engine version will appear here when the server reports it.

1. gen-zero 0.1.0Current · liveserver-reported version
   • Feature · Three MCP tools: zero (12 verbs), causal_fold, pipeline. Protocol 2024-11-05 over SSE.
   • Feature · pipeline decide modes: auto (K-MoE router), mcts, mpc_cem, astar, manifold_gflownet, cfr_nash, reflex. World models: residual, symplectic, contact.
   • Performance · Tier 0 Proceed gate path: P99 ≤ 80 µs (self-reported, /llms.txt). Scoring is prefill only: 0 generated tokens per decision.
   • Safety · Four-tier fail-closed gate (Proceed / Confirm / Escalate / HardStop). Non-finite, out-of-range or unknown input is refused with a typed code, never clamped.
   • Safety · Token auth on the public endpoint: a missing token gets HTTP 401. The sandbox cannot touch the caller's files or processes.
2. Site · Wasm game engine2026-09-30167485d · 00e6eaf
   • Feature · Sokoban, Snake and 2048 demos on a Rust → Wasm engine (games-wasm 0.1.0, under 150 KB). They run in the browser and make no calls to the MCP server.
3. Site · Quant risk-gate demo2026-09-30b1cb7e8 · b89d735
   • Feature · Jump-diffusion market simulation with a five-rule pre-trade gate, on the home page benchmarks section.
4. API · Endpoint moved to api.gen-zero.ai2026-09-2993ec01b
   • Breaking · Public endpoint changed from https://gen-zero.ai/sse to https://api.gen-zero.ai/sse. Update saved client configs.
   • Feature · The old host still forwards /sse, /message, /health, /ready and /metrics to the new one, so same-origin browser probes (like §01) work.
5. Research · Breakthroughs portfolio2026-09-29645d521
   • Performance · Folded adapter head reports 7.8–12.8 µs median per decision; CP-SAT safe dispatch reports 0 violations in 5,000 adversarial tests. Self-reported, not yet replicated.