Send the work. Handle the result.

Include the completed work, its requirements, and enough evidence to check it. TUC returns JSON your workflow can act on.

Try the handoff without a key

The public preview gate checks one sanitized text, code, or JSON file with fixed rules. It calls no external model, does not save the input or result, and exits 0 on pass, 1 on fail, or 2 when more evidence is required.

curl -fsSLO https://ultimate-critic.fly.dev/tuc-preview-gate.py
python tuc-preview-gate.py generated.py --type code --task "Review this generated code safely."

Download tuc-preview-gate.py. The public limit is 10,000 characters and five runs per minute. Do not submit credentials, secrets, private source, or sensitive customer data.

Connect an AI agent

Connect a Streamable HTTP MCP client directly to https://ultimate-critic.fly.dev/mcp. Supply the tenant key as a secret X-API-Key header for paid evaluation and usage tools; omit it for the no-key preview.

The standalone stdio alternative reads the tenant key from TUC_API_KEY; the key is never a tool argument.

curl -fsSLO https://ultimate-critic.fly.dev/tuc-mcp-server.py

Download the MCP server or open the setup guide.

1. Build the request

Field	Purpose
`artifact_type`	Text, code, image, UI, game scene, JSON, or mixed.
`artifact`	Generated content, files, evidence, and builder metadata.
`intent`	Prompt, requirements, constraints, and success criteria.
`evaluation_profile`	Domain-specific judgment such as code review, factuality, visual design, or game scene.
`rubric`	Optional criteria supplied by the account owner.
`reference_materials`	Facts or source material the artifact must respect.

Quote the package before submission

Founding access includes 1,000 review credits. Standard packages use 1 credit, extended packages use 5, and full packages use 10. Call the quote endpoint with the same request body; it does not run an evaluation, store the artifact, or consume credits.

quote = critic.quote_review_credits(payload)
print(quote["tier"], quote["credits"])

Tier	Encoded package boundary	Credits
Standard	Up to 256 KiB, 8 files, no images	1
Extended	Up to 2 MiB, 16 files, 3 images	5
Full	Up to 8 MiB, 32 files, 8 images	10

2. Python SDK

from sdks.python.ultimate_critic import UltimateCriticClient

critic = UltimateCriticClient(
    base_url="https://YOUR_TUC_HOST/api",
    api_key="YOUR_TENANT_KEY",
)

result = critic.evaluate({
    "artifact_type": "code",
    "artifact": {
        "content": generated_code,
        "metadata": {"builder_model": "your-builder-model"},
    },
    "intent": {
        "prompt": "Implement a safe division helper.",
        "requirements": ["Return a typed error for division by zero."],
        "constraints": ["Do not swallow exceptions."],
    },
    "domain": "software",
    "evaluation_profile": "code_review",
})

if result["overall"]["verdict"] != "pass":
    repair_agent(result["findings"])

3. REST

curl https://YOUR_TUC_HOST/api/v1/evaluate \
  -H "Content-Type: application/json" \
  -H "X-API-Key: YOUR_TENANT_KEY" \
  -d '{
    "artifact_type": "text",
    "artifact": {
      "content": "The generated artifact",
      "metadata": {"builder_model": "your-builder-model"}
    },
    "intent": {
      "prompt": "Write the requested artifact",
      "requirements": ["State the conclusion clearly"]
    },
    "domain": "general",
    "evaluation_profile": "general_quality"
  }'

4. Run it asynchronously

job = critic.queue_evaluation(
    payload,
    idempotency_key=f"{build_id}:{artifact_sha256}",
)

while job["status"] in {"queued", "running"}:
    job = critic.get_job(job["id"])

result = job["result"]

Reuse the idempotency key only when retrying the same request. A different payload with the same key returns an error.

5. Enforce the verdict

match result["overall"]["verdict"]:
    case "pass":
        release(result["evaluation_id"])
    case "fail":
        create_atomic_repairs(result["findings"])
    case "insufficient_evidence":
        request_more_evidence(result["metadata"]["evidence_issues"])

6. Check the repair

Create one repair for each finding. After the work changes, submit it again and attach the new evaluation to the repair. Verification requires a new passing result and any evidence requested by the finding.

ticket = critic.create_repair({
    "source_evaluation_id": result["evaluation_id"],
    "finding_id": result["findings"][0]["id"],
    "invariant": "The artifact must satisfy the stated requirement.",
    "observed_error": result["findings"][0]["title"],
    "likely_root_cause": "Builder omitted a required edge case.",
    "gap_kind": "COGNITION",
    "transformation": "Implement the missing edge-case behavior.",
    "verification_method": "Re-evaluate the repaired artifact.",
    "retry_limit": 3,
})