Use denv from your agent (MCP)
Coming soon. What the denv MCP server will let Claude Code, Cursor and other coding agents do, and how it keeps you in control of spending.
Status: coming soon. The denv MCP server is not live yet. This page describes what it will offer. To hear when it ships, or to try it early, write to anup@denv.ai.
denv will ship a Model Context Protocol (MCP) server. With it, a coding agent such as Claude Code, Cursor or Codex can set up a denv cluster and run it for you. You sign in and add a card once. The agent can do the rest.
How you will connect
| Way in | What it is |
|---|---|
| Remote server | A hosted MCP endpoint over streamable HTTP. Nothing to install. |
| Local server | denv mcp in the denv CLI, over stdio, for an agent running on the GPU machine itself. |
| Directories | Listings in the official MCP Registry and in client directories, so agents can find denv. |
What your agent will be able to do
| Group | Tools | Needs sign-in |
|---|---|---|
| Catalog | Providers, GPU types, which models fit which GPU, pricing | No |
| See | Cluster status, sites, nodes, incidents and their explanations, endpoint metrics, job status and logs, uptime reports | Yes, read |
| Run workloads | Serve a model, run a job, scale, roll back, simulate a node failure | Yes, deploy |
| Infrastructure | Create a cluster, add a provider site, get the join command for a machine, set site limits, drain a node | Yes, infra |
| Destructive | Delete a cluster, remove a site | Yes, admin |
The agent can also subscribe to a cluster or an incident and watch a repair as it happens.
An example session
- On your GPU machine, ask your agent: "serve Llama 3.1 8B on this box with denv". It checks that the model fits the GPU.
- A browser window opens. You sign in, add a card the first time, and approve. This is the only step you do by hand.
- The agent creates a cluster.
- The agent runs the join command on the machine. It reports each step until the node is ready.
- The agent serves the model, prints the endpoint and key, and sends a test request.
You stay in control of spending
- Anything that can launch machines or delete things returns a plan with a cost estimate first. Nothing happens until you confirm.
- No single call launches machines.
- Agent tokens are separate from your keys. They are short-lived, scoped to one cluster and the scopes you approve, and carry a spend limit. Read and deploy are granted by default; infra and admin must be ticked on the consent screen.
- Every action is recorded as the agent acting for you and appears in the cluster timeline.
Machines your agent launches at a provider are billed by that provider. denv charges a platform fee per cluster plus a fee per GPU under management. See Pricing.