Patronus AI MCP server: announced, now unavailable

The announcement post is still live. The repository it links went offline between June and August 2026, and Patronus's Python SDK and API are the working route.

Checked 3 October 2026 · from Patronus's MCP announcement post, the GitHub API, the Wayback Machine, PyPI and the official MCP registry

One of 91 guides to monitoring and developer tools. Each covers what an AI coding agent can get out of the tool through MCP, the protocol that lets the agent query the tool directly.

How Patronus AI works with MCP

Patronus AI evaluates LLM and agent output. Its hosted evaluators score hallucination, toxicity and custom criteria, and Percival, its agent debugger, reads traces and names failure modes.

Its homepage now says "We are a frontier lab training the first Digital World Models." In June 2026 it announced a $50M Series B led by Greenfield Partners, with Datadog participating.

The announced server let Claude Desktop pick evaluators, run a batch evaluation and log the traces to Patronus. The surviving code registers six tools: evaluate, run_experiment, batch_evaluate, list_evaluator_info, create_criteria and custom_evaluate.

All six start or configure an evaluation, and none returns a stored trace, span or past result. The post's first example, Cursor finding "the latest AI failures from Patronus", has no matching tool in that code.

Every vendor here gets the same four checks: official, maintained, distributed and usable. Patronus AI passes 0 of them.

Patronus is reachable from code today. An agent with a shell and a PATRONUS_API_KEY can call the SDK directly, which also covers tracing, something the six MCP tools never did.

What can an AI agent do with Patronus AI?

Do you need the MCP server at all?

Use the patronus SDK. Install it with pip, set PATRONUS_API_KEY, and call a RemoteEvaluator for one check or run_experiment for a dataset.

For tracing, the SDK exports OpenTelemetry to https://otel.patronus.ai:4317 by default. Its documentation shows integrations for OpenAI, Anthropic, smolagents, PydanticAI, OpenAI Agents, LangChain and CrewAI.

A community server exists. settlegrid/settlegrid-patronus-ai, created on 16 May 2026 and last pushed two days later, wraps six Patronus calls with per-call billing through SettleGrid.

That wrapper is a third-party project with 0 GitHub stars, not a Patronus release.

Call Patronus from code through its SDK. Treat any MCP wrapper, including the March 2025 copy, as code you maintain yourself.

Patronus AI integration setup

Patronus documents no MCP client setup. The working setup is the Python SDK, configured by environment variables or a patronus.yaml file.

pip install patronus
export PATRONUS_API_KEY=<your-key>
export PATRONUS_PROJECT_NAME=my-agent

from patronus import init
from patronus.evals import RemoteEvaluator

init()
judge = RemoteEvaluator("judge", "factual-accuracy")
result = judge.evaluate(task_input="What is the capital of France?",
                        task_output="Paris", gold_answer="Paris")
print(result.pass_, result.score, result.explanation)

Every other client

Each block below is the configuration for one client, with the file path and the key that client expects.

import patronus

patronus.init()

@patronus.traced()
def answer(query):
    with patronus.start_span("retrieval"):
        ...

Read-only access, permissions and security

• The MCP copy holds your whole API key

  Its documented invocation passes the key on the command line, so the process gets every permission that key carries.

• docs.patronus.ai now requires login

  It answers with HTTP 401 and "contact support@patronus.ai for access". The SDK reference at patronus-ai.github.io/patronus-py stays public.

• The SettleGrid wrapper charges per call

  Its README lists prices from 1 to 5 cents per method, billed through SettleGrid.

What the vendor documents

The announcement post lists three uses. Cursor finds recent AI failures and suggests fixes; Claude Desktop acts as a chat interface over Patronus; and Patronus evaluators score Claude chat outputs.

Should you connect an agent to Patronus AI?

No, not over MCP. Connect through the `patronus` SDK or the Patronus API instead. Best for teams that already use Patronus evaluators and can call them from code or CI.

Strongest fit

No, not over MCP. Connect through the patronus SDK or the Patronus API instead.

Main advantage

Patronus's hosted evaluators, including its Lynx hallucination judge, are a few lines of Python away.

Main weakness

The announced server's repository returns 404, and the server was never packaged on PyPI.

Operational risk

Running the March 2025 copy means maintaining it yourself against the SDK's current releases.

Frequently asked questions