MCPJam is a testing and evaluation platform for Model Context Protocol servers, aimed at the developers, product managers, and engineering teams shipping MCP integrations into clients like ChatGPT, Claude, Cursor, Copilot, and VS Code. The pitch behind this launch is that users increasingly reach products through an MCP server rather than the product's own app, which leaves teams blind to what people are asking, how the agent is reasoning, and whether the server actually delivers value. MCPJam tries to close that loop by letting teams see how their server behaves across every major AI client before real users hit it.
The launch bundles several capabilities on top of MCPJam's existing open-source Inspector. Swarm runs configurable agent personas that prompt a server like real users across multiple clients so teams can watch where in the request flow value drops off. A User Testing mode lets internal QA or beta testers exercise the server inside a hosted client and share session feedback, and those prompts can be promoted into durable eval cases. The evals themselves run across clients and models with deterministic and non-deterministic checks, and slot into CI to gate changes before release. The broader product covers OAuth flow testing, LLM-powered evals, and CI/CD integration, with a CLI and SDK for running the same checks locally or in pipelines.
The launch is led by Prathmesh Patel, who in the announcement video describes leading MCP technical strategy at Asana before starting MCPJam and building the platform he wished had existed there. MCPJam has already shipped an open-source Inspector and a generally available evals product , and this release extends that foundation from local debugging into cross-client behavioral testing, which matters now because MCP surface area is expanding faster than the tooling to keep servers reliable across it.
Comments (7)
Sign in to join the discussion.
L
Linnea Kotani26d ago
The demo cuts are tight but the voiceover sounds like it was recorded in a walk-in closet. Fix the mic and this thing sings.
Y
yulia26d ago
Tagline rewrite, no charge: 'See how your MCP server actually behaves, before your users do.' Yours is doing too much heavy lifting with the word 'test'.
D
Devraj Mahesh26d ago
Naive question from a former intern: if my MCP server passes in Claude but fails in Cursor, whose fault is that in your report? Because I feel like every vendor is going to point at each other.
O
Obi Edozie26d ago
Before we buy: SOC2 Type II, SSO via Okta, and a DPA that our legal team won't rewrite for six weeks. Otherwise we're stuck piloting in a sandbox forever.
F
Farhan A.26d ago
Been telling people MCP tooling was the next gap for months and getting laughed at in group chats. Vindication tastes like a Tuesday launch tweet.
Q
quietbear26d ago
A server that never gets tested is just a rumor with endpoints.
M
Mireia de la Rosa26d ago
Curious about the eval methodology here. Are you scoring tool-call correctness, latency, or some composite, and is the rubric published anywhere I can actually poke at?
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
MCPJam
Test your MCP server in every major AI client.
Test your MCP server in every major AI client.
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
MCPJam is a testing and evaluation platform for Model Context Protocol servers, aimed at the developers, product managers, and engineering teams shipping MCP integrations into clients like ChatGPT, Claude, Cursor, Copilot, and VS Code. The pitch behind this launch is that users increasingly reach products through an MCP server rather than the product's own app, which leaves teams blind to what people are asking, how the agent is reasoning, and whether the server actually delivers value. MCPJam tries to close that loop by letting teams see how their server behaves across every major AI client before real users hit it. The launch bundles several capabilities on top of MCPJam's existing open-source Inspector. Swarm runs configurable agent personas that prompt a server like real users across multiple clients so teams can watch where in the request flow value drops off. A User Testing mode lets internal QA or beta testers exercise the server inside a hosted client and share session feedback, and those prompts can be promoted into durable eval cases. The evals themselves run across clients and models with deterministic and non-deterministic checks, and slot into CI to gate changes before release. The broader product covers OAuth flow testing, LLM-powered evals, and CI/CD integration, with a CLI and SDK for running the same checks locally or in pipelines. The launch is led by Prathmesh Patel, who in the announcement video describes leading MCP technical strategy at Asana before starting MCPJam and building the platform he wished had existed there. MCPJam has already shipped an open-source Inspector and a generally available evals product , and this release extends that foundation from local debugging into cross-client behavioral testing, which matters now because MCP surface area is expanding faster than the tooling to keep servers reliable across it.
Comments (7)
The demo cuts are tight but the voiceover sounds like it was recorded in a walk-in closet. Fix the mic and this thing sings.
Tagline rewrite, no charge: 'See how your MCP server actually behaves, before your users do.' Yours is doing too much heavy lifting with the word 'test'.
Naive question from a former intern: if my MCP server passes in Claude but fails in Cursor, whose fault is that in your report? Because I feel like every vendor is going to point at each other.
Before we buy: SOC2 Type II, SSO via Okta, and a DPA that our legal team won't rewrite for six weeks. Otherwise we're stuck piloting in a sandbox forever.
Been telling people MCP tooling was the next gap for months and getting laughed at in group chats. Vindication tastes like a Tuesday launch tweet.
A server that never gets tested is just a rumor with endpoints.
Curious about the eval methodology here. Are you scoring tool-call correctness, latency, or some composite, and is the rubric published anywhere I can actually poke at?