> ## Documentation Index > Fetch the complete documentation index at: https://docs.mcpjam.com/llms.txt > Use this file to discover all available pages before exploring further. # MCPJam in Claude > Connect Claude to your MCPJam account through the hosted MCP server — projects, saved servers, evals, and user testing MCPJam runs a hosted MCP server that gives Claude access to your MCPJam account: your projects, the MCP servers saved in them, eval suites and runs, Swarms, and user-testing studies. You add it to Claude once, sign in with your MCPJam account, and then ask Claude to do the work in a conversation. | | | | - | - | | Server URL | `https://mcp.mcpjam.com/mcp` | | Transport | Streamable HTTP | | Sign-in | OAuth with your MCPJam account (`login.mcpjam.com`) | | Acts as | You. Every call uses your own MCPJam identity and your own project access | ## Before you start * **An MCPJam account.** Sign up at [app.mcpjam.com](https://app.mcpjam.com). The free plan works. * **A project.** Claude can list your projects and create a new one, and every project tool falls back to your most recently updated project when you don't name one. * **MCP servers to work with** (optional). Servers you already saved in the MCPJam app are available right away. Claude can also connect a new server by URL; when that server needs OAuth, Claude gives you a sign-in link to open. * **Plan and feature access.** Some features depend on your plan, your project role, or features enabled for your account. Ask Claude what you can do in a project. It reads that from your account, and a tool you can't use is refused with a clear error. ## Add it to Claude ### Claude (web and desktop) Any Claude plan can add a connector by URL. 1. Open [**Customize → Connectors**](https://claude.ai/customize/connectors) and add a custom connector. 2. Name it **MCPJam** and enter `https://mcp.mcpjam.com/mcp`. 3. Click **Connect** and sign in with your MCPJam account. The connector shows **Connected** under **Your connectors** when sign-in finished. 4. In a chat, turn it on from **+ → Connectors**. On Team and Enterprise plans, an owner may need to add the connector for the organization first. See Anthropic's [guide to adding a connector by URL](https://claude.com/docs/connectors/custom/add-unlisted#add-a-connector-by-url). ### Claude Code ```bash theme={"theme":"css-variables"} claude mcp add --transport http mcpjam https://mcp.mcpjam.com/mcp ``` Then run `/mcp`, choose **mcpjam**, and sign in. ## Sign-in and what Claude can reach The connector is an OAuth 2.0 protected resource. Claude discovers MCPJam's sign-in page from the server, you sign in with your MCPJam account, and Claude receives a token for this server. Any request without a valid token is refused with `401`, which is what starts sign-in. * **Your access, no more.** Each tool call is forwarded to MCPJam's API with your token, so you see exactly the organizations and projects your account can see, and writes need the same role they need in the app. * **Nothing stored by the connector.** The server is stateless. It keeps no session and no copy of your token between requests. * **Credentials you save stay write-only.** Server headers, environment values and client secrets you save are encrypted at rest. No tool returns a stored secret's value. * **Disconnecting.** Remove the connector in Claude's connector settings to stop Claude from calling MCPJam. ## Tool permissions Each tool tells Claude what kind of action it is: * **Read-only** tools only read your account: listing projects, reading a run's results, fetching a trace. * Tools that **create** something new, such as a project, an eval suite or a run, are not read-only. * Tools that **change, replace, cancel, revoke or delete** something that already exists are marked destructive, and Claude asks before running them. So are `call_server_tool` and `send_chat_message`, because they run tools on your own MCP servers, and those tools can do anything. You can review every tool and set its permission on the connector's page in Claude, under **Tool permissions**. ## Costs Most tools are free to call. These spend your organization's MCPJam credits, or the model-provider keys your organization configured, and their descriptions say so: | Tool | Spends on | | - | - | | `run_eval_suite`, `run_eval_case` | Model calls for every iteration of every case | | `request_eval_run_judge`, `backtest_eval_run_judge` | LLM-as-judge grading | | `send_chat_message` | Each Playground message | | `launch_goal_run` | Every Swarms session the run fans out to | | `set_study_guest_execution` | What anonymous study visitors may run on your account | These are included with MCPJam and are not charged to you, but they count against per-minute and daily usage limits: `generate_eval_cases`, `import_eval_cases`, `generate_personas`, `generate_goals`, `request_swarm_run_insights`, and `request_study_insights`. A tool that hits a limit is refused with `RATE_LIMITED` and the time it can be called again. Adding credits does not lift these limits. ## Examples **See where things stand** > Which MCPJam projects do I have, and which servers are in the most recent one? Claude calls `list_projects`, then `list_project_servers`. **Check a server** > Diagnose the "billing" server in my Payments project. It stopped connecting. Claude calls `diagnose_server`, which probes the URL, connects, and reports what failed. **Run and read evals** > Run the "checkout flows" eval suite in Payments and tell me which cases fail. Claude calls `run_eval_suite` (it asks first, because the run spends credits), then `get_eval_run` until the run finishes, then `get_eval_run_steps` for a failing case. **Did a change break anything?** > Compare the latest "checkout flows" run with the one before it. Claude calls `compare_eval_run`, which classifies each case as regressed, fixed, new, removed, changed, or unchanged. **Add a server** > Connect [https://example.com/mcp](https://example.com/mcp) to my Payments project. Claude calls `connect_project_server`. If the server needs OAuth, Claude gives you a sign-in link to open, then calls `get_project_server_connection_status` until the connection is ready. ## Troubleshooting **Claude says the connector needs you to sign in again.** Your MCPJam sign-in expired or was revoked. Reconnect the connector in Claude's connector settings (in Claude Code, run `/mcp`, choose **mcpjam**, and re-authenticate). **You see a different account's projects.** You signed in with another MCPJam account. Disconnect the connector, connect again, and choose the account you meant on the MCPJam sign-in page. `get_me` shows which account Claude is using. **`NOT_FOUND` for a project, server, suite or run.** The name or id did not match anything your account can see. Project tools default to your most recently updated project, so name the project when you have several. **`403 FORBIDDEN`.** Your role or plan does not allow that action, or the feature is not enabled for your account. A refusal for a feature carries `details.code: "FEATURE_UNAVAILABLE"` and names the feature. **`RATE_LIMITED`.** A usage limit was reached. The error says when the tool can be called again. **`TARGET_REQUIRED` from `run_eval_suite`.** The suite is attached to several environments or clients. The error lists them; tell Claude which one to run. **A tool your teammate mentioned is missing.** Features in beta are not offered through this connector yet, even when they are enabled for you in the app. **Anything else.** Email [support@mcpjam.com](mailto:support@mcpjam.com) with the request id from the error message when there is one. ## Privacy and terms The connector processes your MCPJam data under MCPJam's [Privacy Policy](https://www.mcpjam.com/privacy-policy) and [Terms of Use](https://www.mcpjam.com/terms-of-use). * Tools that run models (evals, judges, Playground messages, Swarms, and the generation and insight tools) send the content they work on to the model providers MCPJam uses. `get_eval_run_disclosure` lists, before a run, which models it calls, where they route, and what is retained. * `call_server_tool` and `send_chat_message` send their inputs to your own MCP servers. * `send_feedback` sends the text you give it to the MCPJam team, where it is kept for 180 days. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.