The builder finishes
The agent packages the work with its task, requirements, constraints, and evidence.
Model Context Protocol
Connect by URL or run one local stdio file. The agent can check generated work, read the verdict, and decide whether to repair, collect evidence, or continue.
The agent packages the work with its task, requirements, constraints, and evidence.
tuc.evaluate sends the package through the configured tenant account.
The returned verdict is pass, fail, or insufficient_evidence.
Install io.github.marksignals/the-ultimate-critic from the official MCP Registry, or point a Streamable HTTP client directly at the live endpoint. Omit the header to use tuc.preview. For paid tools, store the tenant key through the secret or header setting provided by your client.
{
"mcpServers": {
"tuc": {
"type": "streamable-http",
"url": "https://ultimate-critic.fly.dev/mcp",
"headers": {
"X-API-Key": "YOUR_TENANT_KEY"
}
}
}
}
The remote endpoint is stateless. The key is accepted only as the secret X-API-Key header; it is never a tool argument or returned in a result.
The standalone bridge requires Python 3.10 or later and no package install. Paid tools read TUC_API_KEY from the process environment.
curl -fsSLO https://ultimate-critic.fly.dev/tuc-mcp-server.py
{
"mcpServers": {
"tuc": {
"command": "python",
"args": ["C:/absolute/path/tuc-mcp-server.py"]
}
}
}
Use the preview only with sanitized, non-sensitive work. Paid submissions follow the service and privacy terms for your account.
tuc.previewChecks one sanitized text, code, or JSON artifact with the deterministic no-retention preview. Does not consume the paid entitlement.
tuc.quoteReturns the package tier and credit cost before you submit it. The artifact is not evaluated or stored.
tuc.evaluateChecks a complete supported artifact request and reports the charged tier in the response.
tuc.usageReads review-credit usage and remaining access without submitting an artifact.
Credit tiers: standard is up to 256 KiB, 8 files, and no images (1 credit); extended is up to 2 MiB, 16 files, and 3 images (5 credits); full is up to 8 MiB, 32 files, and 8 images (10 credits). Limits apply to the encoded request package.
A useful pipeline handles each result explicitly. A missing-evidence result is not a pass, and a failed result should carry the supported findings into the repair step.
Continue only if the result also satisfies your own release policy.
Send the findings and verification steps to the next repair stage.
Collect the requested proof and submit the package again.
$249 once includes one tenant key, 90 days, 1,000 review credits, and up to two hours of first-integration support. Use tuc.quote to see whether a package costs 1, 5, or 10 credits before submitting it.