MCP Channel home

MCP analytics and evaluations

Your first useful report starts with one tool call.

Bring an existing MCP and evaluate it privately, understand calls through MCP Channel, or improve a published server. These guides are public. An account is required to run checks; a marketplace listing is not required for private evaluations.

Connect your MCPChoose a setup guide

Choose your analytics guide

Guide: evaluate a private MCP

Before you start: use a public HTTPS Streamable HTTP MCP endpoint on port 443 and a tool you are authorized to test. Localhost, private networks, stdio, legacy SSE-only endpoints, redirects, and automatic OAuth sign-in are not supported. Put credentials in headers, never in the URL; query strings and fragments are not accepted.

  1. Sign in and save a private connection

    Open Observe & Evaluate. Under Connect a private server, enter a connection name and your public HTTPS MCP endpoint, such as https://your-server.example/mcp. If needed, supply Authentication headers as JSON using a dedicated test credential, for example {"Authorization":"Bearer YOUR_TEST_TOKEN"}. Select Save private connection. Nothing is published.

  2. Inspect tools for free

    Select your connection and choose Inspect tools (free). Read the static schema and description checks. Discovery contacts your server but does not invoke its tools; a clean static report is not evidence that a call works.

  3. Activate the trial or use your existing Pro allowance

    Before executing a tool, choose Start 7-day analytics trial, or use the allowance already included in MCP Channel Pro. You do not need a second subscription if you are already on Pro. Check evaluation calls and Shared AI today in the plan panel. AI grading, analytics assistance, and Dev Studio all spend the same daily tokens.

  4. Approve and run one safe test

    Under Explicit evaluation, choose a read-only tool you control. Read its displayed schema and enter an Arguments (JSON object) value that matches it. Add optional Expected text to check the output. Confirm the call approval, then select Run one evaluation. Start without AI grading; opt in separately only if you are comfortable sending a redacted sample to the evaluator.

  5. Read the result, not just the execution status

    Look for Successful tool call when the call returns useful output and the requested deterministic checks pass. Otherwise read the failure or checks-needing-attention report before deciding whether to retry. Review duration and any optional AI assessment, then find the saved report under Recent synthetic evaluations. A single success is not a guarantee of future correctness.

Start your private evaluation

Guide: understand your gateway calls

Basic observability is free. It only covers traffic through MCP Channel, not calls made directly to an external server.

  1. Sign in, subscribe to a marketplace listing, and use Keys to scope your client connection to the subscriptions it needs.
  2. Make one permitted tool call through your connected client or the Playground. The listing's own pricing and permissions still apply.
  3. Open Observability. Choose a Window and Calls from key, optionally filter by Listing slug, Tool name, or Error category, then select Show calls.
  4. Read success and failure, latency, and recent calls. If the report is empty, check the selected key, scope, time window, and whether the call actually passed through MCP Channel. From your MCP client, you can also call mcpchannel.account/call_statistics.
Open free observability

Guide: improve a published MCP

  1. Open My listings, select a listing you own, and find Usage analytics. Reports need calls through MCP Channel; direct traffic to your server is not included.
  2. Choose a time window. Compare call quality, failures, and latency with the preceding window. Investigate a failed call or inspect schema and description issues with Server doctor.
  3. Make a targeted improvement, run an explicitly approved test if needed, and compare subsequent results. Basic reports and rule-based answers remain free; AI assistance and active evaluations use your analytics trial or Pro allowance.

No listing? Use the private evaluation guide instead; you do not need to publish just to test your server.

Open publisher analytics

Start free. Upgrade only when you need more.

Shared AI tokens reset at midnight UTC without rollover or automatic AI overages. Runs without AI grading spend evaluation calls but no platform AI tokens. Evaluation-call limits stay separate from the shared AI pool and hosted calls. Dev Studio Pro includes 5,000 hosted calls/month, then $0.005 per additional hosted call. Marketplace subscriptions and upstream provider charges are separate.

View your trial and Pro plan

If your first test does not work

My connection or discovery fails

Check the exact Streamable HTTP endpoint, public HTTPS certificate, port 443, and test headers. Redirects and private-network targets are rejected. Discovery is bounded to 200 tools with an overall 20-second operation deadline.

Run one evaluation is disabled

Inspect tools first, select a tool, activate your trial or Pro allowance, and approve this particular call. Approval resets after dispatch and when you change the selected tool or arguments.

My AI allowance is exhausted

Dev Studio and analytics share the same daily tokens. Check Shared AI today and wait for midnight UTC, or upgrade from Free to Pro for a larger shared pool. You can still run an approved test without AI grading if evaluation calls remain and your trial or Pro access is active.

The call returned, but the checks need attention

Execution success is not the same as useful output. Read the report for empty output, tool errors, or a failed Expected text match. AI scores are advisory and do not replace deterministic checks.

Read the analytics FAQ