An MCP server for training and evaluating models
Kojen's MCP server gives your own agents the whole dojo: create projects and datasets, tune and search models, run evaluations and red-team simulations, deploy endpoints and wait for runs to finish. It's the same set of tools the in-app Kojen Agent uses, and it only ever reaches your own data.
What it does
Sign in, no key to paste
Clients that support MCP authorization (Claude, ChatGPT, Cursor, VS Code, Claude Code) need only the URL: they open a Kojen sign-in and ask you to allow the connection. Disconnect any of them under Settings → Connected apps.
36 tools
From get_workspace_summary and create_dataset to train_model, evaluate_model, run_simulation, deploy_model, wait_for_run and ask_sensei. Every tool is also callable over REST at /api/v1/tools/{name}.
Scoped API keys
Clients without sign-in support take a kojen_… key in a header. Keys are stored as hashes, can be read-only, expire, and can be revoked at any time.
Listed in the MCP Registry
Kojen is published as ai.kojen/kojen, with discovery at /.well-known/mcp and OAuth metadata at the standard well-known paths.
How it works
- 1Add https://kojen.ai/api/mcp to your MCP client.
- 2Sign in to Kojen and allow the connection.
- 3Ask your agent to train, evaluate or deploy, and it calls Kojen's tools.
claude mcp add --transport http kojen https://kojen.ai/api/mcp
# then run /mcp in Claude Code and sign in to Kojen
// ~/.cursor/mcp.json (or .cursor/mcp.json in a project)
{
"mcpServers": {
"kojen": { "url": "https://kojen.ai/api/mcp" }
}
}Questions
- What is an MCP server?
- The Model Context Protocol is an open standard for giving AI assistants tools. An MCP server exposes tools over HTTP; a client like Claude or Cursor lists them and calls them when the conversation needs one.
- Can an agent change or delete my data?
- Only what the key or connection allows: tools run as you, with row-level security on every query. Give each agent its own key with the fewest scopes it needs; the in-app Kojen Agent asks before deleting anything.
- Does it cost anything?
- Connecting is free, and tools that run on hosted models count toward the same daily token allowance as the web app. Only GPU training uses paid credits.