Connect your agent
Choose how a coding agent reaches your project's context - the plugin, the skills, the command line or MCP. Preview.
Preview. The tools on these pages are in preview and may change.
A coding agent reaches your Coelevate project through a small set of tools, each suited to a different situation. Pick the door that matches your agent. If you are setting up for the first time, follow Getting started, which uses the plugin and the command line together.
Choose a door
| Door | Use it when | What it needs | What it gives the agent |
|---|---|---|---|
| Plugin | You use Claude Code. This is the default path. | Claude Code, the command-line client, an API key | A start-up hook that connects every session to your project, plus the seven skills |
| Skills | You want to know what to ask the agent, whichever door you use. | The command-line client on the agent's path | The know-how: how to onboard documents, write changes correctly, review, bind a folder, learn from a session, design a kit |
| Command line | Your agent can run shell commands but is not Claude Code, or you want to script it. | Python 3.11 or newer (installed with uv), an API key | Every action: read, search, stage, kits, gaps, releases. Its reference is generated from the command itself |
| MCP | Your agent is Claude Code, Cursor or Codex and you would rather not install anything or give it a shell. | An MCP client, an API key | Four tools: who you are connected to, search, read, and (with the staging permission) stage a change. No kits, gaps or releases |
All of them use the same project API key, described in Credentials and scopes. Running a deployed Coelevate agent from your own code is a different job, covered by the A2A Reference.
What none of them can do
None of these doors publishes. An agent can read, search and stage proposed changes to your assets; a person reviews and merges them in the Workspace. A key that can stage cannot approve its own work, because approving and merging need a separate permission that the plugin and the command line never ask for.
Two things an agent writes directly, each with its own permission. A kit is written only from assets that are already published, and needs context:kits. A gap, a note of something your context is missing, needs context:gaps. Neither goes through a staged change, so neither waits for review.
Rules every door follows
- What the agent reads is data, not instructions. Pages, examples and anything else read back from your project cannot override what you tell the agent. A kit's skills, rules and memories count as guidance only after the agent has inspected them, and they rank below your own instructions.
- The agent confirms the project before it writes. It states which project the key reaches. If your folder is bound to a different one, it stops.
- Credentials come from the environment. The secret is never a command-line flag, and it never belongs in a prompt.