Connect coding agents

wheeljack launches coding agents already installed on your machine. It does not bundle subscriptions, provider credentials, model access, or provider billing.

Supported adapters

Adapter	Executable	Structured mode	Authentication outside wheeljack
Claude Code	`claude`	Yes	Install and authenticate Claude Code.
Codex CLI	`codex`	Yes	Run `codex login` or use the CLI’s normal sign-in flow.
OpenCode	`opencode`	Yes	Run `opencode auth login` or configure its normal provider credentials.
Pi	`pi`	Yes	Install `@earendil-works/pi-coding-agent`, then run `pi /login`.
Generic shell	Platform shell	No	No provider authentication.

Provider installation commands change independently of wheeljack. Use each provider’s official documentation, then confirm that the CLI starts successfully in an ordinary terminal before diagnosing it inside wheeljack.

The installation method does not matter. wheeljack resolves adapter executables from the desktop process and, on macOS, the user’s login-shell PATH, so Homebrew, MacPorts, Nix, npm, Bun, pnpm, nvm/fnm, Volta, mise/asdf, Cargo, and standalone installations work when they expose the documented executable. Standard manager locations remain available if the login shell cannot be queried; Settings → Agents shows which search-path source was used.

Detect and verify

Install and authenticate the CLI outside wheeljack.
Restart wheeljack if the executable was added to PATH while wheeljack was already running.
Open Settings → Agents.
Choose Scan all to detect and probe every adapter.
Choose Verify all to run one real non-mutating turn through every installed, signed-in adapter, or Verify selected for only the current adapter.
Choose the adapter as the default or select it when creating a new agent pane.

Verification checks executable resolution, the adapter’s structured protocol, and the launch profile wheeljack will use. A changed executable or protocol profile invalidates stale verification instead of trusting an old result. Verification is explicit because it may contact each CLI’s configured provider and consume one minimal turn.

Update installed adapters

Choose Update all to inspect the active executable for each installed adapter. wheeljack proves ownership through npm, pnpm, Bun, Yarn, Homebrew, WinGet, Scoop, Chocolatey, or a documented native self-updater before offering an update. The confirmation dialog shows every exact command; confirmed updates run sequentially, then wheeljack scans again and verifies successfully updated adapters.

Ambiguous, custom, Nix-declared, version-manager-shim, app-bundled, and unknown standalone installs are skipped with a reason. wheeljack never guesses an installer, runs a composed shell command, changes declarative package configuration, or updates an adapter while one of its sessions is active.

Structured sessions versus shells

A generic shell displays terminal output only. A structured adapter additionally exposes chat messages, reasoning, tool activity, approvals, questions, cancellation, model selection, Plan routing, autonomy requests, and review evidence.

wheeljack fails a structured launch when the required persistent-session contract is unavailable. It does not silently downgrade a Plan task to a one-shot or unstructured agent.

Choose project access

Agent default preserves the CLI’s normal sandbox and approval policy.
Full access explicitly asks the CLI to allow internet access and local files without its ordinary approval boundary.

Managed agents can request bounded coordination actions such as messaging a peer, starting a child, handing off work, or requesting review. Configure each action as Allow automatically, Ask every time, or Deny in agent settings. wheeljack enforces the selected policy and records the request and result in Autonomy history. New profiles allow agent discovery automatically; messaging, spawning children, handoff, review, and conflict resolution default to Ask every time because approved actions can create provider usage.