Connect a coding agent

Add another coding tool, and its first session starts with the same instructions and skills as your other tools. Its changes must pass the same gate: your project's own commands, such as its linter and tests. You don't teach the project again, because what your agents have learned lives in the project.

discerndiscern↗A tool that lets you hand real work to coding agents and still decide what joins your project. supports Claude Code, Codex, Gemini, Cursor, and GitHub Copilot. Platforms and providers lists what each needs.

Say your project uses Claude Code and Codex, and you want to try Cursor too. Install Cursor first, then ask the agent you already use:

"Set this project up for Cursor alongside our other tools, keeping our instructions and skills. Tell me about any trust steps I need to take."

Any other supported tool works the same way.

You don't need to run any of these commands yourself.

It works in its own worktree, a separate copy of the project on its own branch, so the change stays off your shared branch until you land it.

It updates the list of tools. discern.toml, your project's discern configuration, lists the tools discern sets up, and a tool missing from the list isn't set up. So the agent adds Cursor and keeps the tools you still use:

[project]
agents = ["claude_code", "codex", "cursor"]

It refreshes the tool files. For each tool on the list, discern refresh sets up its instructions, skills, hooks, and Model Context Protocol (MCP) registration. MCP is the connection that lets an agent call discern's tools directly. Where a file also holds your project's own settings, refresh changes only discern's entries.

It checks the setup and brings the change to you. discern doctor checks the installation and names a fix for anything wrong. The agent then commits and runs the gateGate↗The full set of checks your project requires before a change counts as finished.. The change lands only when you say so, as Finish and land a change describes.

Once the change lands, open a new Cursor session in your main checkout, your original project folder. The new tool finds discern only in a checkout that contains the change.

Cursor then asks you to trust the folder and to approve discern's tools the first time it uses them. Those steps belong to the tool, so only you can take them. discern doctor names them for every tool on your list. For Cursor, it reports:

`cursor`: Committed integration configuration and first-use tool calls require workspace approval. Trust the workspace, then approve the discern MCP server tools on first use; use the headless bypass only when explicitly intended (flag `--approve-mcps`).

Then ask the new agent:

"Check that you can reach discern from this session, and tell me what it says to do next."

The agent calls discern_status, discern's status tool. In a clean main checkout, a connected session answers:

`main` is clean in the main checkout.

Some tools need a new conversation, a window reload, or a restart before they load the connection. Only a successful status call shows that a session has loaded it.

Ask your agent to find out why, and whether a trust step or restart needs you. The tools don't appear in the agent's session covers the causes and fixes.

If you decide against Cursor, tell your agent which tools to keep. It updates the list, refreshes, and runs the gate as before.

Refresh leaves a dropped tool's files where they are, so ask your agent to remove the ones no remaining tool uses. For Cursor, those are discern's entries in .cursor/mcp.json and .cursor/hooks.json, while AGENTS.md stays, because Codex reads it too.

• The tool list names every tool your project uses.
• The change passed the gate and landed.
• A new, trusted session of the new tool calls discern's status tool successfully.

To change what every tool knows, edit your project instructions once.