Cross-agent delegation, on your machine

Your favorite agents.
Now your subagents.

Let the agent you’re working with hand a task to another.
Each subagent uses its own tools, model, and account.

Open-source desktop app · macOS preview · No Bridge account

Subagent BridgeINTERACTIVE PREVIEW

Who should do what?

A rule for the agent you already use.

Try a local example
Rule 1

Click any underlined choice. Example models only; this demo never starts an agent.

From coding agents to personal agents.

Codex Pi OpenCode DeepSeek Claude Code Grok Build HermesPlanned Antigravity
HOW IT WORKS

Your agent leads.
Another agent lends a hand.

  1. 01

    Pick your team.

    A task, an agent, a model.

    CodexPi
  2. 02

    Apply the rule.

    Tell your agent who can help.

    can useshould use
  3. 03

    Keep working.

    Your subagent takes it from here.

    The result comes back.

How the handoff works

CURRENT COVERAGE

Know what connects.

Codex, Pi, OpenCode, and DeepSeek can orchestrate. A catalog of 27 agents helps Bridge discover more subagents on your machine.

Integration details
Tested versions & connection methods
Coding agents, tested versions, and integration roles
Agent / tested versionAs orchestratorAs subagent
Codex0.160.1Native hooksACP adapter
Pi1.1.0ExtensionACP adapter
OpenCode1.18.34PluginBuilt-in ACP
DeepSeekDesktop-bundled 0.2.0-rc.2Native jobsACP profile
Claude CodeNot live-testedDeferredUnverified adapter

Catalog entries are candidates, not 27 verified integrations. Settings distinguishes installed adapters from successful ACP discovery; a live task is still needed to confirm provider access. Claude Code is deferred, and Hermes is planned. Codex and Pi need separate worker adapters. Setup guide ↗

Bring your agents together.

Install notes

Preview 0.3.2· MIT license

Requires Node.js 22.13+ and configured coding agents. The app and CLI share the same rules and configuration.

Developer ID signed · Apple notarized

Open normally after moving the app to Applications. See the Mac guide for setup and worker adapters.

Windows & Linux installers · Coming soon

THE DETAILS

Before you begin.

Which account pays for the model?

The subagent uses its harness’s existing account, API key, or provider. Subscription allowances and API charges follow that setup. Bridge does not provide model access or sell another subscription.

What stays on my computer?

Bridge’s rules and coordination run locally. Your agents still send prompts and code to their configured model providers. A shared project folder does not copy the whole chat: the orchestrator supplies the context each task needs.

How do results and follow-ups work?

Your main agent (the orchestrator) sends a task and the context it needs. The subagent works in the same project folder using its own harness, tools, and account. Independent work can continue in the background. First-time setup may need a reload or hook approval. Supported integrations return results to the session that started the task. Codex hooks deliver during active turns or when a turn stops; an idle chat receives pending results next turn. Follow-ups reuse the same worker conversation while its process is alive. Restarting the harness ends that continuity.

What do “can use” and “should use” mean?

“Can use” lets the orchestrator decide whether to delegate. “Should use” instructs it to delegate matching tasks; it cannot guarantee a model will obey. Multiple agents or models are alternatives: the orchestrator picks one of each per task. Bridge enforces those allowed choices when a task runs.

BUILD IT WITH ME

Help connect the next agent.

Subagent Bridge is open source under the MIT license, with a macOS preview you can download today. Found a bug or have an idea? Tag @am_tarachand on X, or connect on LinkedIn.