Skip to content
Upstate Docs
Log in Open Upstate

Operate with confidence

Coding-agent access

You can let a coding agent inspect the same production evidence you would read in Upstate. The hosted MCP connection provides bounded, read-only tools; it is not permission to run shell commands, deploy applications or repair production.

Connect a supported client

Use this endpoint:

https://upstate.cloud/mcp

For Codex:

codex mcp add upstate --url https://upstate.cloud/mcp
codex mcp login upstate

For Claude Code:

claude mcp add --transport http upstate https://upstate.cloud/mcp

In Claude Code, open /mcp and complete browser authorization. Other clients that support remote HTTP MCP and browser OAuth can use the same endpoint. You do not need to install the Upstate CLI or copy observer credentials for this connection.

Review the approval

The browser shows the client-provided name and the accounts you may grant. A client name alone does not prove who operates the client. The current account is initially selected; additional accounts require explicit selection. A newly created account is not silently added to an old approval.

A grant only reads the accounts you approved and still belong to. Removing account membership reduces access; disconnecting the grant in Local tools revokes that connection. Human CLI sessions, MCP approvals and host observer credentials are separate.

What an agent can investigate

The tools can enumerate approved accounts, projects, components and hosts; read fleet and coverage summaries; inspect deployment receipts and comparisons; and read bounded alert, health, activity and host-posture evidence. It can wait for a receipt's verification result, but that operation only reads server state.

For example, ask it to identify which host stopped reporting, explain a receipt that needs attention, or compare production errors before and after a deploy. Its answer is only as good as the collected evidence and retention window; unknown or stale information remains unknown or stale.

Deploy with an agent

A coding agent that has your separate local shell/repository permissions may run Kamal through those separate permissions after your approval. The MCP tools do not grant the ability to do so. A safe workflow is:

  1. Let the agent inspect the local source and Upstate's read-only production evidence.
  2. Review its deployment proposal, including migrations, destination and version.
  3. Run upstate kamal deploy in the authorized local environment.
  4. Inspect upstate receipt --watch or ask the MCP tool to wait for the receipt.
  5. Review the independent result and any suggested next steps before approving further changes.

The hosted receipt wait holds a request briefly and may need repeated calls until a terminal state is reached. It does not guarantee the agent's local deployment finished.

Older MCP clients

upstate mcp --help
upstate mcp

The optional stdio bridge requires an existing CLI login, forwards the hosted tool catalog and is intended for clients without remote HTTP support. It is not required for modern Codex or Claude Code integrations. For terminal operations, see CLI command reference.