Choose an agent

On this page

DevCapsule supports three coding agents. Select one or more for your checkout; DevCapsule installs them on the next launch and preserves their local state and settings across container restarts.

Agent Add it to your checkout Start it in the IDE terminal Download trust channel
Claude Code devcapsule project config add claude-code-agent claude claude-code (claude-code-download)
OpenAI Codex devcapsule project config add codex-agent codex npm package (@openai/codex, SHA-256 verified)
Antigravity CLI devcapsule project config add antigravity-agent antigravity antigravity-cli (antigravity-download)

Run the table's config add command in a host terminal from your project directory (or request add from the running session's web console). It updates your local checkout record and per-checkout component pins without changing .devcapsule/devcapsule.toml or the shared lock. If the project already offers the agent, use config restore when it is omitted. Required agents are already selected. The console records an accepted add request immediately; the running environment stays unchanged. Save your work, stop the capsule, and run devcapsule project run to start with the updated selection.

Claude Code and Antigravity CLI are proprietary binaries downloaded on your machine from pinned sources and verified by SHA-256. Review the origin shown in the trust prompt. For Antigravity, the URL alone does not prove Google owns the deployment. Trust a channel once across all projects in your configuration home (devcapsule home trust accept claude-code or devcapsule home trust accept antigravity-cli), or approve it for this checkout only (devcapsule project config authorize claude-code-download true / antigravity-download true) after reviewing the vendor's terms; see Vendor software and trust.

Bring your own provider account, subscription, or API key—DevCapsule does not bundle an AI plan and never imports credentials from your host automatically. See Sign in and defaults for authentication and seeded settings.

Why agents run without step-by-step prompts here

DevCapsule seeds each agent to run without interactive tool-approval prompts so it can read code, edit files, and run builds and tests autonomously. That posture relies on the container boundary rather than per-command dialogs:

  • Your host home directory, SSH keys, browser profile, other repositories, and host desktop are unmounted by default (The boundary).
  • Inside the capsule, the agent has full write access to the mounted repository (/workspace/<slug>, including .git) and to this checkout's persistent state. Push clean commits to a remote or keep an independent backup before large autonomous runs, because local Git history inside a writable .git directory can be modified by any process in the capsule.
  • Prefer step-by-step confirmation? Change the permission mode in the agent's own settings file inside the capsule; DevCapsule preserves your explicit setting across launches.

A first request

Start with a small, verifiable task in the IDE terminal:

Explain how this project builds and runs its tests, then propose one small improvement and how to verify it. Wait for my go-ahead before editing files.

Inspect the diff, run the test suite, and commit what you want to keep. Give every agent durable repository context with an instructions file at the project root; see Instructions for agents.

Using multiple agents or switching later

Give each agent the project instructions as described in Instructions for agents. They share the working tree and keep settings in separate directories (~/.claude, ~/.codex, ~/.gemini). These directories do not isolate agents from each other: they run as the same capsule user. You can install two or all three in the same checkout—for example, using one to implement and another to review—or switch later without losing saved sign-ins; see Review and change agents.

Read the source on GitHub Back to top ↑