Start in Claude
you → claude
Collaborate with Codex. Review this plan and challenge its assumptions.
Claude ⇄ Codex
Claude Code ⇄ Codex Desktop
Open source
Two models catch what one misses.
Claude and Codex were trained differently, so they make different mistakes. Agents Bridge puts the two chats you already have open in one conversation, so together they catch and fix each other’s errors.
Interactive panel preview
Set the limits. You approve before the chats begin.
Suggested by Claude · just now
Waiting for your approval
Review the error handling in
src/api and propose fixes.Codex chat
codex-my-app-3f
Codex can edit and commit
Local edits and commits only; no push, pull requests or merges.
Messages
Time
Change
-
CompletedPlan the caching layer, then implement it phase by phase.Claude ⇄ Codex · 31/60 messages · 1 h ago
-
PausedInvestigate why the sync test is flaky.Claude ⇄ Codex · 12/24 messages · yesterday
The collaboration panel, as it opens on your Mac. Try the switch, the steppers, and the language and theme buttons.
Two real chats. One shared conversation.
Not a delegated job.
A conversation.
Your Claude chat. Your Codex chat. Each keeps its context, questions the other, and answers back. No sessions to manage. No copying between windows.
Start in Codex
you → codex
Collaborate with Claude. Review this plan and challenge its assumptions.
Codex ⇄ Claude
Both chats stay in the conversation.
Compare the approachesWhy two models
Different models.
Different blind spots.
One model reviewing its own work tends to agree with itself. A second model, trained differently, questions the assumptions the first one never thought to check.
-
One finds what the other misses
Each model has its own habits and biases. When Codex reviews Claude's plan, or Claude reviews Codex's code, mistakes that slipped past the author tend to surface.
-
One fixes what the other breaks
Findings go straight back to the chat that wrote the code, with the context to fix them. No copying output between windows.
-
Disagreement is the signal
Where they still disagree is exactly where you should look. The bridge keeps it on the record instead of forcing a consensus.
-
Reviews with real context
The reviewer isn't a fresh run starting from zero. It's your chat, with its history and the files it has already read.
How it works
No sessions to manage.
Just ask in plain words.
Ask either chat. From Claude or from Codex, describe what you want them to do together and the limits you have in mind.
you → claudeCollaborate with Codex on this: review the error handling in src/api and propose fixes. 20 messages, 20 minutes.
Answer three questions. Nothing is sent until you say yes, and the bridge reads your answers from the app itself.
claude asksMay Codex edit files and make local commits?› Yes
How many messages and how much time at most?› 20 messages · 20 min
Start with a new Codex chat?› Yes
Watch them work. The two chats message each other directly. Pause, finish or export from the panel at any time.
claude → codexReview task: the retry path in src/api/client.ts. Draft v1 attached.
codex → claudeTwo defects: the timeout is swallowed in fetchUser, and retries ignore 429.
6 of 20 messages · 14 min left
Read the result. Every result and review says who wrote it and which version it refers to. Where they still disagree stays on the record.
result v2 · by claudeFix the 429 retry; surface the fetchUser timeout.
review of v2 · by codexAgree on the retry. Disagree on the timeout: it is handled upstream.
Disposition: proposal, with disagreements
You stay in control.
The agents don't.
The bridge reads your answers from the app itself, never from what an agent says. Text from the other agent carries no authority, whatever it says.
- Starting
- Only when you answer Yes or press Start
- Editing
- Off until you allow it, within the scope you write
- Push, PRs, merges
- Excluded unless your scope names them
- Limits
- Messages and time, enforced both ways
- Delivery
- Uncertain messages are shown, never resent
- Data
- Local sockets and a private file on your Mac. No model calls.
Compare
Different ways to work together.
codex-plugin-cc lets Claude Code hand reviews and tasks to Codex, resume those jobs, or transfer the session into a Codex thread. It's a good fit for delegating. Agents Bridge connects two chats that are already open, for an exchange over many steps.
| Aspect | Delegation plugins | Agents Bridge |
|---|---|---|
| The idea | Delegation pluginsDelegate reviews and tasks from Claude Code; resume jobs or transfer context | Agents BridgeTwo live chats in an ongoing exchange |
| Who starts | Delegation pluginsClaude Code, with slash commands | Agents BridgeEither chat, in plain words |
| Shape | Delegation pluginsA job and its result | Agents BridgeA back-and-forth with message and time limits |
| Your controls | Delegation pluginsSlash commands and job status | Agents BridgeThree questions, then the panel: pause, finish, export |
| The record | Delegation pluginsJob results | Agents BridgeAuthored, versioned results and reviews, disagreements kept |
Install
Two commands per app.
Then just ask.
Installs as agents-bridge.
For macOS. You need Claude Code 2.1.224 or later, in the Code tab of Claude Desktop or as claude in a terminal; Codex Desktop running (support for the Codex CLI is coming); Bun 1.4 or later; and Git.
-
01 Install the Codex side
codex plugin marketplace add jaimebernalm/agents-bridgecodex plugin add agents-bridge@jaimebernalmIn Codex, open Agents Bridge in the plugins and trust both hooks. Then open a new Codex chat.
-
02 Install the Claude side
claude plugin marketplace add jaimebernalm/agents-bridgeclaude plugin install agents-bridge@jaimebernalmThen open a new Code chat in Claude Desktop, or start
claudein a terminal in the project folder. -
03 Check that they see each other
Open the same project folder in both apps and ask either chat: “Use Agents Bridge to list the chats in this folder.” Not listed? See Troubleshooting.