MCP
Connect an AI agent to work with your Tensorant projects, data, training runs and models.
Use Tensorant from Claude Code, Cursor, Codex, VS Code, Claude, ChatGPT or another compatible MCP client. An agent works as you in one organization, follows your role and plan limits, and records its changes in the organization's audit trail. Every plan, including Free, supports agents.
Start here
| What you want to do | Guide |
|---|---|
| Connect your first agent and ask about your projects | MCP quickstart |
| Set up your app with browser sign-in or a personal token | Connect your app |
| See available tools, file uploads and spending previews | Tools |
| Choose permissions, create tokens or revoke access | Access and tokens |
Connect your agent to Tensorant's MCP server:
https://mcp.tensorant.io/mcpYou can also copy this address from Settings → Agents.
Start with a request such as “Which projects do I have, and what does my plan allow?” The agent can inspect your account and projects before making changes.
Choose the access it needs
| Access | What the agent can do |
|---|---|
| Read | Inspect projects, data, runs, evaluations, models and activity |
| Build | Import, generate and review data, create versions, evaluate existing endpoints and ask running models |
| Full | Start training, trials and deployments after a spending preview; owners can also manage published models |
Viewers can grant Read access. Members and owners can grant more, but an agent never gains permissions its user does not have. Agents cannot manage connections, billing or organization memberships. Connect again to change an agent's access level, then revoke the old connection.
Spending and credentials
Training and deployment launches need Full access and a spending preview. Build access can still incur endpoint charges when generating examples or running evaluations, and calling a wake-on-request model can start a billed Serverless worker. See Spending.
Personal tokens are shown once and expire after 30, 90 or 365 days. Revoke your own agents under Settings → Agents; owners can revoke any agent in their organization. Treat tokens as secrets and keep them out of prompts and shared files.
For an application that only needs answers from a published model, use the application API and a model API key. MCP tokens and model API keys authenticate different services.
Need a hand? Visit troubleshooting