mcp-registry/com.completionkit/evals · v0.29.1
As of 2026-10-18 19:19 UTC, the Completionkit Evals MCP server is up with a Glimind reliability score of 93/100 (grade A). Over the last 30 days it answered 100% of Glimind's liveness checks. It had 6 outages in 90 days, with a median recovery of 2 minutes.
Prompt evals over MCP: run a prompt on your dataset, score each output 1-5 with an LLM judge.
Claude Code
claude mcp add --transport http completionkit-evals-mcp-server https://completionkit.com/mcp
Claude Desktop · Windsurf · Cline (mcpServers JSON)
{
"mcpServers": {
"completionkit-evals-mcp-server": {
"type": "http",
"url": "https://completionkit.com/mcp"
}
}
}
Third-party server, listed as discovered. Review its source before installing.
54 of 54 tool(s) carry no readOnlyHint/destructiveHint annotations.
Annotate read-only tools with {"readOnlyHint": true}. Agents and gateways can then run them without a confirmation prompt, and route around destructive ones. How to fix →
54 tools on one server. Tool-selection accuracy drops as the menu grows, and some clients (e.g. Cursor) cap the number of active tools.
Group tools by task into smaller servers or toolsets, and merge near-duplicates behind one tool with a mode parameter. How to fix →
111 parameter(s) across 43 tool(s) have no description.
Add a "description" to each parameter with its format and an example ("ISO 8601 date, e.g. 2026-10-17"). This is the commonest cause of wrong arguments on the first call. How to fix →
29 tool(s) look like they change state (create/send/delete/pay…) but declare no destructiveHint.
Set annotations {"readOnlyHint": false, "destructiveHint": true|false, "idempotentHint": …}. Clients use these to ask the user before a side effect and to auto-approve safe reads. How to fix →
17 tool(s) have descriptions under 20 characters.
Say when to use the tool and what it returns, not just its name restated ("Get weather" → "Current weather for a city: temperature, conditions, wind. Use for 'now', not forecasts."). How to fix →
Handshake + tools/list takes ~2,088 ms.
Serve tools/list from memory, keep the process warm, and host close to your users. Every new agent session pays this before its first call.
No tool declares an outputSchema.
Declare outputSchema and return structuredContent (MCP 2025-06-18+). Agents can then parse results without guessing, and chain tools reliably. How to fix →
In the last 30 days: 0 tool(s) removed and 3 schema change(s).
Add new tools or optional parameters instead of renaming or removing; deprecate first (say so in the description) and remove later. Agents cache tool contracts. How to fix →
Paste-ready definitions with the problems above fixed. Changes marked verify need a human look. 43 tool(s) still need parameter descriptions written by someone who knows the API. All of them as JSON: /v1/readiness/mcp-registry%2Fcom.completionkit%2Fevals/fixes
{
"name": "agreements_create",
"description": "Upsert an agreement for (run, response, metric, created_by). Verdict is one of agree, disagree, borderline. corrected_score (1..5) is required when verdict is 'disagree'.",
"inputSchema": {
"type": "object",
"required": [
"run_id",
"response_id",
"metric_id",
"verdict"
],
"properties": {
"note": {
"type": "string"
},
"run_id": {
"type": "integer"
},
"verdict": {
"enum": [
"agree",
"disagree",
"borderline"
],
"type": "string"
},
"metric_id": {
"type": "integer"
},
"created_by": {
"type": "string"
},
"response_id": {
"type": "integer"
},
"corrected_score": {
"type": "number"
}
}
},
"annotations": {
"readOnlyHint": false,
"destructiveHint": false,
"idempotentHint": false
}
}{
"name": "agreements_list",
"description": "List agreements. Filter by run_id, response_id, metric_id, or created_by.",
"inputSchema": {
"type": "object",
"required": [],
"properties": {
"run_id": {
"type": "integer"
},
"metric_id": {
"type": "integer"
},
"created_by": {
"type": "string"
},
"response_id": {
"type": "integer"
}
}
},
"annotations": {
"readOnlyHint": true
}
}{
"name": "datasets_create",
"description": "Create a dataset with CSV data. First row is the header. Two column names are recognized specially: \"expected_output\" is each row's answer key (ground truth) given to the judge and to checks that compare against the row's expected value, and \"actual_output\" is a pre-made output to score in a prompt-less run. Both are overridable per run (expected_column / output_column). Every column is also available to the prompt as a variable.",
"inputSchema": {
"type": "object",
"required": [
"name",
"csv_data"
],
"properties": {
"name": {
"type": "string"
},
"csv_data": {
"type": "string"
},
"tag_names": {
"type": "array",
"items": {
"type": "string"
}
}
}
},
"annotations": {
"readOnlyHint": false,
"destructiveHint": false,
"idempotentHint": false
}
}Get an email or webhook when this server's readiness drops by 10+ points, tool-poisoning text appears, or a tool breaks:
curl -X POST https://glimind.com/v1/subscriptions -H 'content-type: application/json' -d '{"tools":["mcp-registry/com.completionkit/evals"],"email":"[email protected]"}'Agents: watch_tool({"toolId": "mcp-registry/com.completionkit/evals"}) over MCP.
metrics_update ~628 tokensmetrics_create ~622 tokensruns_create ~561 tokensruns_update ~440 tokensresponses_list ~430 tokensStatic checks of the advertised tools/list (schema, descriptions, annotations, tool-poisoning patterns, names, size) plus observed protocol, uptime, latency and contract changes. Read-only: no tool is ever called. Weights: schema 15, docs 20, safety 15, efficiency 15, compatibility 15, protocol 5, reliability 10, stability 5. Token counts are estimates (~3.6 characters per token).
Operator? Fix these, then re-check live. Embed the readiness badge:
[](https://glimind.com/server/mcp-registry%2Fcom.completionkit%2Fevals?ref=badge)
<a href="https://glimind.com/server/mcp-registry%2Fcom.completionkit%2Fevals?ref=badge"><img src="https://glimind.com/v1/readiness/badge/mcp-registry%2Fcom.completionkit%2Fevals.svg" alt="MCP agent-readiness"></a>
.. image:: https://glimind.com/v1/readiness/badge/mcp-registry%2Fcom.completionkit%2Fevals.svg :target: https://glimind.com/server/mcp-registry%2Fcom.completionkit%2Fevals?ref=badge :alt: MCP agent-readiness
Glimind records every version of this server's tool definitions in a signed, append-only transparency log. Current manifest sha256:f98b426f7e0176ee…, first seen 2026-10-17, included at leaf 14814 of 23,166. Version history · verify what your client received.
Compare: most reliable ml servers · most reliable data servers · most reliable other servers · most reliable knowledge servers · most reliable security servers
Glimind tells agents whether a server is up right now and how to call it, then fails over to a live alternative when it isn't.
Claude Code
claude mcp add --transport http glimind https://glimind.com/mcp
Claude Desktop · Windsurf · Cline (mcpServers JSON)
{
"mcpServers": {
"glimind": {
"type": "http",
"url": "https://glimind.com/mcp"
}
}
}
As of 2026-10-18 19:19 UTC, the Completionkit Evals MCP server is up with a Glimind reliability score of 93/100 (grade A). Over the last 30 days it answered 100% of Glimind's liveness checks. It had 6 outages in 90 days, with a median recovery of 2 minutes.
Completionkit Evals MCP server advertises 54 tools, including agreements_create, agreements_list, datasets_create, datasets_create_from_url, datasets_delete.
Point your MCP client at this streamable-http endpoint: https://completionkit.com/mcp
Glimind scanned all 54 tool definitions for tool-poisoning patterns — hidden or invisible characters, instructions that override the model, concealment from the user, requests to read credential or MCP config files, conversation exfiltration and cross-tool shadowing — and found none. This checks what the server tells the model, not what its code does; review the source before granting access to sensitive data.
Status page: 90-day uptime, every outage and how long it lasted.
Claim this listing to get the Verified operator marker, a remediation checklist and owner analytics, or embed the live badge from any of the tool pages above. Maintainers: alerts, deploy attestation and monthly reports.
Status is an automated estimate from Glimind's liveness checks (handshake and tool listing only; no tool is ever called) and may lag or differ from your network. Glimind is not affiliated with this server or its operator. Operator? Request a correction or opt out at https://glimind.com/opt-out.