Agent capabilities
Which AI agents the Coord Runner supports and how to set them up.
The Coord Runner can execute jobs using multiple AI agents. Each agent has different strengths and platform requirements.
Supported agents
Claude Code
Anthropic's coding agent. Claude Code is the primary agent for Coord and works on all supported platforms.
To install: Follow the instructions at Claude Code's documentation.
Codex
OpenAI's coding agent.
Platform support:
- macOS — supported
- Linux — supported
- Windows — not available
To install: Follow the instructions at OpenAI's Codex documentation.
Gemini CLI
Google's coding agent. Gemini CLI works on all supported platforms.
To install: Follow the instructions at Gemini CLI documentation.
Agent detection
The Coord Runner automatically scans your machine for installed agents when it starts. For each agent, the runner shows:
- Detected — the agent is installed and ready to use.
- Not found — the agent is not detected on your machine. The runner provides instructions for installing it.
- Unsupported — the agent is not available on your platform (e.g., Codex on Windows).
To re-detect agents after installing a new one, click Rescan Capabilities & Tools in the runner settings.
Enabling and disabling agents
In the runner settings, you can toggle each agent on or off. Only enabled agents are used to execute jobs. If multiple agents are enabled, the runner uses the one best suited for the job based on configuration and availability.
Tool approvals
When an agent needs to perform an action that could affect your system — such as running a command or modifying a file outside the project directory — the runner displays an approval prompt showing the requested action and its details. You can approve or deny it directly in the runner.
Actions within the project directory are handled automatically based on your permission settings. Certain dangerous operations (such as writing to system directories or running destructive commands) are automatically denied.
Session continuity
When a job is revised after a review, the agent can resume from its previous session rather than starting fresh. The runner automatically saves session data during execution, so the agent has full context of its earlier work when picking up a revision.
Execution timeout
An agent that has been idle for one hour — producing no session activity — is automatically timed out. Agents that are actively working and returning session data will not time out, regardless of how long the job takes. Time spent waiting for your input — such as tool approval prompts or questions from the agent — does not count as idle time.
Next steps
- Configuration — adjust runner settings
- Set up a runner — full setup guide