OpenMark MCP server

• Addresshttps://openmark.ai/mcpnot in the official registry: answered when asked

No entry in the official MCP registry names this server. It answered on openmark.ai when the Atlas asked at the address servers are usually mounted on, so it is the vendor's own. Clients and catalogues that list servers from the registry do not show it.

• Stateanswerschecked 2026-10-27
• Tools9
• Server calls itselfopenmark
• Server version1.0.0
• Protocol version2025-06-18

On 2026-10-27 the server answered initialize and tools/list without a sign-in. First answer in 186 ms.

• Whether its tools do the product's job has not been judged yet.
• Checked on 1 of the 7 days it takes to judge reliability.

Said by rules from the measurements above. The likely first call is a model's pick from the tool list; read its description before relying on it.

• INITE MCP scoreafter a weeka score needs 7 daily checks: 1 so far
• Answeredmeasuring1 of the 7 daily checks it takes to print a share
• Median first answer249 ms
• Space in an agent's context≈ 2,172 tokensthe definitions of all its tools (9), which an agent carries on every turn
• Tool definitionsB · 78 of 100
• Typed results0 of 9 toolsdeclared outputSchema: results come typed, read once

• Parameters without a description: in 3 of 9
• Not saying whether it only reads or can destroy: 9 of 9

A program checks the server once a day: initialize and tools/list, no model. Tokens are the definitions' length at about four characters a token. The grade averages six checks of every tool's name, description, schema and annotations; it says how well the tools are described, not whether they work. The score weighs what was measured and nothing else: answering (25%), doing the product's job (25%), definitions (20%), context (10%), use (10%) and how readily a client connects (10%). A part that was not measured is left out, and a hidden instruction caps the score at 40. The state of vendors' MCP servers: method and numbers →

• get_task_format

  How to write an OpenMark task: the YAML structure, all 18 scoring modes, and tips. Call this before writing a task. Pass a topic for the full detail on one subject (the app's Task Guide): fields, scoring, composite, chaining, documents…

• validate_task

  Check a task YAML the way OpenMark will before running it: structure, scoring settings, and the test limit of the user's plan. Free, saves nothing. Returns the problems to fix and warnings worth a look.

• list_models

  Current OpenMark model ids with list prices ($ per 1M tokens), as text and as a structured list. Use these ids in run_benchmark. Append @low, @medium or @high to a reasoning model id to set its reasoning effort.

• prepare_attachments

  Get an upload_uid and a one-hour upload link (no API key needed) for images, PDFs or text documents, with the exact curl command. Upload the files, reference them by file name in each test's attachments list, then pass the upload_uid to…

• estimate_cost

  Credits a run_benchmark call would use, before spending anything: the expected total, the split per model, the balance needed to start, the worst case, and the user's balance. Same inputs and run settings as run_benchmark. Use it to drop…

• run_benchmark

  Save a task (YAML, see get_task_format) to the user's OpenMark account and run it on the given models with real API calls. Returns a job_id and the expected cost; poll get_results. Spends the user's credits; pass max_credits to only start…

• get_results

  Status of a benchmark job, and once completed the ranked results (score, stability, cost per run, latency), the tests each model missed with its answer, a recommendation for the chosen objective and quick picks for the others.

• get_account

  The user's plan, credit balance and limits, and what a credit pack, Pro or Expert would unlock. Use it before large runs, or when a run is refused for credits or limits.

• watch_task

  Watch the task of a completed job: when new or cheaper models ship, OpenMark re-tests up to 4 of them next to the user's pick and emails the user only when one beats it. Each check uses the user's credits at the normal rate.