Know your MCP server works before your users do.

In the pre-production layer between build and production, and we're SDK and framework agnostic, so we work with any MCP server however you built it. We don't instrument your code or your live traffic. We exercise your server the way real AI clients do, during dev, QA, beta, and CI/CD, because that's the window where every failure mode is still visible and cheap to fix. Once it ships into an external agent it's a black box, so the highest-value reliability work happens just before that line, which is exactly where we live.

Agent eval and observability tools (Datadog, Braintrust, LangSmith, Arize) measure the agent you built, in a system you control that already sees the user's prompt, context, and tool calls. MCP evals measure the other side of the handshake: how your software behaves when an external agent you don't control (ChatGPT, Claude, Copilot, Cursor) decides whether to call it, with what arguments, and how it uses the result. MCPJam sits outside your system and confirms your software is production-ready for every external agent, before your users ever interact with it.

The core is open source and free, forever: the client, Inspector, core CLI and SDK, local evals, and conformance checks, to run locally or in your own CI/CD.

Paid plans are for when your team is ready to go further:

• Live client matrix: test against continuously maintained emulations of every AI client, kept current for you so a host change never quietly breaks your tests.
• Swarm: turn loose AI agent personas that acceptance-test your software at scale.
• Chatboxes: hosted, shareable UAT environments that capture and replay every human tester session.
• AI insights: root-cause diagnosis and fix suggestions, drawn from reliability patterns across thousands of MCP servers.
• Reporting & history: team dashboards and trends across every run and release.
• Enterprise governance: SSO, audit logs, DPA, and SOC 2 (Type 1 in-progress).

No, we are strictly pre-production. There may be an ability in the future to import traces of production traffic to bolster session data and provide better insights for you, but for now we're strictly pre-production, where we believe you can get more reliable telemetry anyway given external agents unreliably offer you production user insights.

No, that's the point. ChatGPT, Claude, Gemini, Cursor, Slack and the rest each support different things and change constantly. MCPJam maintains the current behavior of all of them, so your tests reflect what your users experience today without your team tracking a single client.

With evaluation, not assertions. MCPJam scores whether the agent selected the right tool, sent the right arguments, and completed the job across runs and clients, then diagnoses why a score dropped and what to change, so erratic behavior becomes a metric you can gate on.

Acceptance testing run by AI agents that act like your users. You define personas; a swarm runs multi-turn journeys through your server across every client and surfaces where it fails the job, continuously and before launch.

Secure, isolated UAT environments: a shareable web client that mirrors the major AI hosts, so internal QA and beta testers can break things safely, and their real sessions become regression tests automatically.