BENCHYARD GUIDES
Coding agents in Benchyard
Benchyard gives your coding agent a workspace that keeps running. The agents share one interface, but each signs in its own way and can do different things.
Last reviewed . What is available depends on your installation and account.
Which agents and sign-ins are supported?
| Agent | Supported sign-in | Important detail |
|---|---|---|
| Claude Code | Claude subscription or API key | Use the subscription or key that provides your agent access. |
| Codex | ChatGPT sign-in with a one-time code, or API key | Signing in with a code must be allowed for your ChatGPT account. If it is not, use an API key. |
| Cursor | Cursor account or API key | Account sign-in may need renewing after approximately 60 days. |
| Gemini CLI | API key | Subscription sign-in is not supported in Benchyard. |
How do agents, models and providers differ?
An agent is the program that reads files, runs tools and carries out a coding task. A model is the AI behind it, and a provider is the company that runs the model. Supporting an agent does not mean every model it can use has been tested.
A team can send its agents' requests through a gateway of its own (a service the team runs between its agents and the provider) for Claude Code, Codex and Gemini CLI, but not for Cursor. Whether a particular gateway and model work together is for the team to check; Benchyard does not guarantee it.
More agents and models are planned, including Kimi Code, GLM Code and MiniMax. They are not available today.
Do all agents use the same approval controls?
No. Benchyard shows the approvals each agent asks for, and agents differ in what they ask. Cursor does not ask before each edit in the way Claude Code and Codex do in Ask first mode.
A session's mode is Ask first, Accept edits, Full auto or Plan, where the agent has it. While an approval waits, the action that needs it does not happen. A notification opens the session, and you decide there, signed in.
Can I change direction while an agent is working?
Codex can take a new instruction while it is working. With the other agents you stop what it is doing and send a new message.
One device drives a session at a time, and your other signed-in devices can watch and approve. Sending a message from another device takes the session over. Separate sessions, each with its own copy of the code if you like, keep tasks apart. Agents in different sessions do not coordinate with each other by themselves.
Where do the credentials come from?
Use your own subscription or API key for the agent, or a sign-in your team provides. Benchyard does not include an agent subscription.
Agent sign-ins for Claude Code, Codex and Cursor, API keys for the services Benchyard knows, and secrets you tie to a service's address are held by credential keepers outside the workspace, so the agent uses them without seeing them. They are stored encrypted. Benchyard is not end-to-end encrypted; the security guide says what that means before you choose how to provide secrets.