Coordinate the coding agents you already use

Terminal interop overlay listing detected coding agents with their connection status and a preview of the configuration Clio would add. — /interop lists the coding agents Clio detected and the entry that connecting would add. Terminal · v0.5.7

On this page

Before you start See which peers Clio can use Choose managed work or an interactive handoff Give one bounded assignment Keep worker authority separate Separate changes with a worktree Review before combining

You can use Clio Coder to coordinate the coding agents you already have installed. Give a peer a focused assignment, follow a managed run where the connector supports it, and review its output before accepting the change.

Before you start

Clio Coder with one working model connection. The first session covers it.
A peer's CLI installed and signed in under its own account.
An assignment whose answer you can check against the repository.

See which peers Clio can use

Shell

clio-coder interop inspect --json
clio-coder configure --interop

The first command reports what Clio can find; the second configures interoperability. In a terminal conversation, /interop shows the available connections. The interoperability guide covers the mode for each peer.

Terminal report of detected coding agents with versions and the next step needed to use each one. — What Clio detected on a machine with three peer CLIs installed but not signed in.

Choose managed work or an interactive handoff

The three commands answer different needs. Pick by whether you want a recorded run or a place to work with the peer yourself.

Command	What happens	What Clio records
`/delegate <peer> <task>`	A configured ACP peer runs the task	The result and a receipt
`/run --target <target> <agent> <task>`	A configured headless target runs it	The result and a receipt
`/peer <peer> <brief>`	A Herdr pane opens for you to work in	No managed receipt

Do not infer that a pane completed a task because it opened. Do not infer that a managed receipt contains every internal reasoning step or mediates every operation inside a peer-owned loop.

Give one bounded assignment

Assign to a peer

Inspect the parser's handling of empty input. Identify the existing tests and propose a regression test. Do not edit files.

That last sentence is a task instruction, not an enforced boundary. Where supported, select the actual read-only restriction on the dispatch and inspect the admission preview. A connector that cannot enforce a restriction is not equivalent to one that can.

Keep worker authority separate

Dispatched workers and external peers run at default, even when the main session uses yolo. Read-only is a restriction on one dispatch; Clio does not carry the main session's unrestricted setting into every worker.

Each peer enforces read-only through its own mechanism. Check the mode you select instead of assuming one contract for all of them.

Peer	Read-only headless run in v0.5.7
Codex	Its read-only sandbox mode
Claude Code	Plan mode with read tools
Antigravity	Its plan and sandbox route
Pi	Read tools
OpenCode	Refused before launch

ACP peers may own their tools. Explicitly permitting peer-owned tools changes what Clio can observe and mediate, so read the guide's permission limits before using such a peer on sensitive work. Clio's coordination is not an operating-system sandbox.

Separate changes with a worktree

For supported headless dispatch, --worktree keeps a Git task branch for review. That helps you compare two proposed changes without mixing their working files.

A worktree is an organizational boundary. It does not confine a peer's filesystem or network access. Keep assignments narrow, inspect the branch and diff, and run the repository's checks before merging a result.

Review before combining

What to look for

A useful result may be a diagnosis, a patch, or a failed attempt that reveals a missing dependency. Look for the files it touched, the checks it ran, and the limits of its report. A claimed test result and an observed check are different evidence.

Clio Coder terminal: a verifier worker on Gemma-4-E2B-it-GGUF finished with 0 tool calls and reported that the test suite passed; its receipt reads verdict pass and quality inferred: validation claimed, none observed. — A worker on a small model reports a passing suite after no tool calls; its receipt records the claim as unobserved. Terminal · v0.5.7

The worker above ran on a small model, made no tool calls, and still reported a passing suite. Its receipt keeps the difference: the verdict repeats the claim, and the quality line says no validation was observed.

When comparing agents, give them the same task and evaluate the actual output. Record the model, connector, setup, check results, and any costs you can substantiate. Clio coordinating two peers does not make either one better.

Boundaries

Supported modes differ by peer, and some cannot receive live steering.
A pane is an interactive handoff; it produces no managed receipt.
Peer accounts, subscriptions, and usage limits stay with each peer's own provider.