Patronus AI MCP server: withdrawn, and what to use
The announcement is still live and the repository it links is gone. What remains is a personal copy with six tools, and the integration that works is OpenTelemetry.
Last verified 26 August 2026 · from Patronus documentation, public repositories and package registries
This page is one of 90 in a directory of monitoring and developer tools. Each page checks what an AI coding agent can get out of the tool through its MCP server, the connector that lets the agent query the tool directly.
Summary
Patronus AI announced an official MCP server, and there is now nothing to connect. The post introducing the Patronus AI MCP server is still live, and the repository it links returns HTTP 404.
Only a personal copy survives, anwarchk/patronus-mcp-server, with no commits since 26 March 2025. It was never published to PyPI and is absent from the official MCP registry. Its six tools only run or configure evaluations, and none reads a trace, a span or a past run.
Patronus itself is alive: it evaluates LLM and agent output for hallucination and quality, and its Percival debugger reads agent traces and names failure modes.
The way in is OpenTelemetry, not MCP: send traces over OTLP, the OpenTelemetry wire protocol, and read the results inside Patronus.
What is the Patronus AI MCP server, and is it usable?
MCP is an open protocol that lets an AI coding agent such as Claude Code or Cursor call a tool's functions directly, without anyone writing code against the tool's API.
Has an MCP server and has a usable one are different questions.
Every one of the 90 vendors in this directory gets the same four checks. Patronus AI passes 0 of them.
| Check | Result | Evidence |
|---|---|---|
| Official | Fail | It was official and is now unreachable. Patronus announced it on its own blog as the first MCP server for agent testing and optimization, and that post still links patronus-ai/patronus-mcp-server, which returns HTTP 404. The organization's 6 public repositories hold no MCP repository. |
| Maintained | Fail | Last commit 26 March 2025, 25 commits all in one week, all by one author, zero external contributions. |
| Distributed | Fail | PyPI returns 404 for patronus-mcp, patronus-mcp-server and patronus-mcp-py; zero results in the official MCP registry; no hosted endpoint. |
| Usable | Fail | The README's 430 lines contain no client configuration, and the install step is an unedited template placeholder. |
Patronus belongs to the agent observability and evaluation stage. It has a side role in code review, because Percival suggests prompt fixes. It has no infrastructure telemetry, no CI product and no cost product.
| Part of the work | What Patronus AI has here | Can an agent reach it |
|---|---|---|
| Code, tests & review | Percival prompt-fix suggestions | Partial, inside Patronus rather than through the connector |
| CI & release | None | No pipeline product, and no tool that reaches one |
| Production observability | None | No infrastructure telemetry of any kind |
| Agent observability & evals | Evaluators, Percival trace debugger | Partial. Over OpenTelemetry. The MCP server cannot read traces |
| AI cost management | None | No spend or billing surface |
The one stage Patronus belongs to is the one its connector cannot serve. Evaluation and trace debugging are real products, and an agent reaches neither through MCP. The six tools start and configure evaluations. None of them reads a trace back. That gap, rather than the abandoned repository, is the reason the working route below is OpenTelemetry.
What can an AI agent do with Patronus AI?
The server source defines six tools: evaluate, run_experiment, batch_evaluate, list_evaluator_info, create_criteria, custom_evaluate. Each one runs or configures an evaluation.
Asked "what happened in production last night?", an agent connected to this server has no tool that can answer.
There is no tool for reading traces, spans or past runs. An agent connected to this server and asked an observability question will either say it cannot answer or, worse, invent an answer.
Guard: Do not connect this server. Even for scoring, you would run a repository by hand. It has not moved since March 2025 and ships no client setup. Send traces over OpenTelemetry instead and let Percival read them inside Patronus.
Setup instructions for agent clients do not exist. The README's 430 lines contain no client configuration at all, and its install step still reads git clone https://github.com/yourusername/patronus-mcp-server.git, an unedited template placeholder. The only documented invocation is python src/patronus_mcp/server.py --api-key ....
What connecting costs before the first question
A context window is the amount of text a model can hold at once, and connecting uses some of it before the agent asks anything. Every tool the server registers sends its name, description and input schema to the model when the session starts. The tool list is charged whether or not a single tool is called.
Six tools cost little because the list is short. Grafana's default set costs roughly six to nine times as much, depending on the column, and answers questions this server cannot reach.
Five of the six tools carry no description at all. Only custom_evaluate has one. A description is how an agent decides which tool to call. The low figure below is therefore partly the cost of definitions an agent cannot use with confidence.
| Tool set | Tools | Tokens (GPT-5.6, GPT-5.5) | Tokens (GPT-4, GPT-3.5) | Tokens (Claude Haiku 4.5 to Opus 5) | Share of a 200,000-token window |
|---|---|---|---|---|---|
| Every tool, as registered | 6 | 1,802 to 3,344 | 1,700 to 3,342 | 3,097 to 3,875 | 0.9% to 1.9% |
Encodings: GPT-5.6, GPT-5.5 o200k_base; GPT-4, GPT-3.5 cl100k_base; Claude Haiku 4.5 to Opus 5 count_tokens. Counted 19 August 2026 from tools/list on anwarchk/patronus-mcp-server. It was run locally with a placeholder key.
Registration is static, so no account was involved. Figures cover name, description and input schema. The two models in column one share the o200k_base encoding, so they get one count. The spread is formatting, about 86%. Schemas are nearly the whole payload. Reasoning tokens are output, spent per request. This count is input, charged once at connection.
The Claude figures are measured rather than encoded. Anthropic publishes no local tokenizer. The definitions were sent to the API's count_tokens endpoint, and a no-tools baseline was subtracted. The range covers two tokenizer generations rather than two serializations. Haiku 4.5 and Opus 4.6 sit at the low end. Opus 5 and Sonnet 5 sit at the high end. The gap between them is about a quarter.
Can Patronus AI watch your own AI agents?
Patronus does read agent traces, and not through this connector. Percival’s parser is built on OpenTelemetry and OpenInference, a set of OpenTelemetry conventions for LLM and agent traces. The signals below describe what the six tools reach, not what the product can do. Read from src/patronus_mcp/server.py on 19 August 2026.
| Signal | Covered | Notes |
|---|---|---|
| Trace and session tracking | No | None of the six tools reads a trace, a span or a past run. That gap is the reason this page sends you to OpenTelemetry |
| Eval scores on recorded traffic | Partial | evaluate, batch_evaluate and custom_evaluate start an evaluation. No tool reads a finished score back |
| Prompt version tracking | No | No tool stores or compares a prompt version |
| Offline experiments | Partial | run_experiment starts one, and its results are read inside Patronus rather than returned to the agent |
| Token and cost tracking | No | No tool returns a token count or a cost figure |
Do you need the MCP server at all?
No. The working integration is OpenTelemetry. It is the route Patronus builds for. Patronus ingests traces over OTLP, the OpenTelemetry wire protocol, and Percival's parser is built on OpenTelemetry and OpenInference, a set of OpenTelemetry conventions for LLM and agent traces.
Documented framework support covers smolagents, Pydantic AI, OpenAI Agents SDK, LangChain and CrewAI, plus custom OpenAI and Anthropic clients through OpenAIInstrumentor and AnthropicInstrumentor.
Percival detects more than 20 failure modes in agent traces, clusters errors, suggests prompt fixes, and scores traces 1 to 5 on reliability and security dimensions.
Percival Chat is a hosted conversational interface over your traces, so the interactive path exists; it runs inside Patronus rather than inside your coding agent.
The REST API is fully documented, roughly 160 pages with per-endpoint samples in cURL, JavaScript, Go and Python.
If you already emit OpenTelemetry traces, the integration is already done. A path for pointing Claude Code at Patronus and asking questions is not built.
Troubleshooting
- The README has no client configuration
- None exists. Its 430 lines contain no client setup at all, and the install step still reads
git clone https://github.com/yourusername/patronus-mcp-server.git, an unedited template placeholder. - The only documented invocation is
python src/patronus_mcp/server.py --api-key ..., run by hand. - `pip install patronus-mcp` finds nothing
- It was never published. PyPI returns 404 for
patronus-mcp,patronus-mcp-serverandpatronus-mcp-py, and the official MCP registry returns zero results. - You want Patronus to read your agent traces
- Send them over OpenTelemetry. Patronus ingests traces over OTLP, and Percival's parser is built on OpenTelemetry and OpenInference. That is the route Patronus builds for.
Patronus AI MCP server: Reference
| Item | Value |
|---|---|
| Official server | patronus-ai/patronus-mcp-server, in the vendor's own organization |
| Last commit | 26 March 2025 |
| Commits | 25, all landed in one week, all by one author |
| External contributors | 0 |
| PyPI | 404 for patronus-mcp, patronus-mcp-server and patronus-mcp-py |
| Official MCP registry results | 0 |
| Hosted endpoint | None |
| Tools | 6 |
| Tool names | evaluate · run_experiment · batch_evaluate · list_evaluator_info · create_criteria · custom_evaluate |
| Tools with a description | 1 at tools/list. Only custom_evaluate. create_criteria and list_evaluator_info are written with docstrings, but registered through lambdas, which carry none |
| Tools that read a trace | 0 |
| README length | 430 lines, with no client configuration |
| Documented invocation | python src/patronus_mcp/server.py --api-key ... |
| Working route instead | OpenTelemetry over OTLP, read by Percival |
| Inventory source | src/patronus_mcp/server.py, read 19 August 2026 |
What engineers report
Engineers have published almost nothing about Patronus, so this section reports counts instead of quotes.
| What was checked | What it shows |
|---|---|
| Hacker News, exact phrase Patronus AI, all time | 3 hits |
| Stack Overflow | 0 questions |
| Mostly vendor posts and job listings |
The low counts are not a judgment on the product, which is well documented. They mean you will be evaluating Patronus without the benefit of other people's experience, and should budget time for that. For scale within the same category, the same exact-phrase search finds DeepEval 60 times, Traceloop 49 and Komodor 27. Counts retrieved from the Hacker News Algolia API on 13 August 2026.
Findings to weigh before you commit
The company's attention has moved to a new mission.
The Patronus homepage reads: "Simulating the World's Intelligence... We are a frontier lab training the first Digital World Models." The evaluation platform is a navigation item beneath that.
The public repositories say the same thing. None of the six public repositories was pushed in the 90 days before verification. The most recent, mdlm_world_modeling, last moved on 22 May 2026. The rest are research and benchmark projects older still. A $50M Series B closed on 25 June 2026, with Datadog among the investors, so the company is funded. The open question is whether the evaluation product stays a priority.
What to verify: Ask Patronus directly for a written roadmap commitment on the evaluation platform before you build on it.
The pricing page contains an unedited template block.
It renders a default reading "Individual / Base / $25 / 600 pages", which is not a Patronus price. Real per-call rates are published for some evaluators, enterprise is contact-sales, and no price is published for Percival on any plan.
What to verify: Get a written quote covering Percival before you commit.
Should you connect an agent to Patronus AI?
No. A server exists and is not usable. Best for OpenTelemetry-instrumented teams that want agent-quality evaluation and can live without a coding-agent connection.
- Strongest fit
- Teams that want deep evaluators (hallucination, PII, context relevance) and an agent debugger. The connection is OpenTelemetry; the results live in Patronus rather than in your coding agent.
- Main advantage
- Percival is a distinct product: a trace-reading debugger that names failure modes and suggests prompt fixes.
- Main weakness
- The MCP server is abandoned and has no trace-query tools, so your coding agent cannot investigate through Patronus.
- Operational risk
- The company's attention has visibly moved to world models. Confirm roadmap commitment before building on the platform.
From Oodle
One platform for agent traces and infrastructure
Agent traces usually sit in a different product from the rest of your telemetry, so when a slow database makes an agent slow the symptom is in one tool and the cause is in another. Oodle keeps both in one query surface, from $10 per million spans.
See agent observabilityFrequently asked questions
Can I connect Claude Code to Patronus?
No, not usefully. The MCP server exposes only evaluation tools, has no trace queries, is not on PyPI or the MCP registry, and ships no client setup instructions. Use OpenTelemetry instead.
Is the MCP server abandoned?
Patronus has not said so, but the repo shows no commits since 26 March 2025, no releases, no external contributors, no publication to PyPI, and no entry in the official MCP registry.
Does Patronus monitor infrastructure?
No. It evaluates LLM and agent output and ingests agent traces. It has no host, container or database telemetry.
