AgentBridge

Summary

AgentBridge is a local CLI and daemon that connects a Claude Code session and a Codex TUI as peers in one working session. Either side can push a message into the other's running turn, so cross-review and task handoff need no human relay.

Captured 2026-09-08 · theagentsindex.com

From: Free
Free tier: YesThe whole tool is free under MIT with no vendor-imposed limits. Practical requirements are Bun as the runtime plus your own Claude Code and Codex CLIs and logins; macOS and Linux only, Windows is not officially supported.

Visit AgentBridge Compare Find alternatives

What it does

Writes code: NoAgentBridge relays messages between two coding agents; the code is authored by Claude Code and Codex, not by the bridge.A bridge that forwards messages between an MCP channel and the Codex app-server protocolgithub.com/raysonmeng/agent-bridge
Debugs and fixes: NoIts only diagnostic command inspects the bridge's own health, not your program; no failure diagnosis or fix landing.Read-only diagnosis: env, daemon health/readiness, build drift, artifact alignment, TUI attachment, logsgithub.com/raysonmeng/agent-bridge
Writes tests: NoNo test generation. The bridge never sees file changes or command execution, and the test count advertised on the homepage is AgentBridge's own suite, not a product feature.Only forwards agentMessage items, not intermediate commandExecution, fileChange, or similar eventsgithub.com/raysonmeng/agent-bridge
Reviews code: PartlyIt routes one agent's review of the other's diff into a running turn; the reviewing itself is done by Claude or Codex, not by AgentBridge.Codex implements; Claude reviews the diff in real time inside the same session and pushes change requests straight back into Codex's thread.raysonmeng.github.io/agent-bridge
Whole-codebase work: NoThe bridge carries messages, not edits. Repository-wide work is executed by the two connected agents, and scope is one pair per project directory.AgentBridge only relays messages between them on your machine; raw work never leaves the box.raysonmeng.github.io/agent-bridge
Ships it: NoNo deploy or release step: the command set (init, claude, codex, resume, pairs, doctor, budget, logs, kill) is local development plumbing only.A local developer tool for connecting Claude Code and Codex in one workflowgithub.com/raysonmeng/agent-bridge
Done-for-you service: NoSoftware you install and run yourself: a global npm package plus a local daemon. There is no delivery team and no managed service.A hosted service or multi-tenant systemgithub.com/raysonmeng/agent-bridge

How much it does unattended

Runs autonomously: PartlyBoth agents launch with permission prompts disabled so a pair can run unattended, but the vendor's own framing is that you steer while the agents coordinate.By default abg claude launches with --dangerously-skip-permissions and abg codex with --yolo. This is deliberate: an unattended agent pair can't stop…raysonmeng.github.io/agent-bridge
Multi-agent: YesThis is the entire product: two agents from different providers held live as persistent peers, each able to push into the other's running turn.AgentBridge is a local bridge that lets Claude Code and Codex collaborate in real time inside one working sessionraysonmeng.github.io/agent-bridge
Agent permissions: PartlyTwo states only: max-permission by default, or --safe to restore each CLI's normal prompts. The vendor states it is not a security boundary.Add --safe (or set AGENTBRIDGE_SAFE=1) to launch with normal prompts. AgentBridge is a local dev tool, not a hardened boundary between untrusted…raysonmeng.github.io/agent-bridge

Which models it runs on

Claude: YesClaude Code is one of the two supported sides; it is started as 'abg claude' with the AgentBridge channel enabled, using your own Claude login.Today the shipping in-session integration connects Claude Code and Codex.raysonmeng.github.io/agent-bridge
GPT: YesOpenAI's Codex CLI is the other side of the pair, started as 'abg codex' in a second terminal against the same bridge.Today the shipping in-session integration connects Claude Code and Codex.raysonmeng.github.io/agent-bridge
Open models: NoThe bridge selects no model at all, and the closed list of wired agents is the two proprietary CLIs; further adapters are a roadmap vote, not shipped.Today the shipping in-session integration connects Claude Code and Codex. Which agent comes next is a public vote.raysonmeng.github.io/agent-bridge
Model choice: PartlyAgentBridge picks no model and no boss role; whatever each CLI is configured to run is what runs. The choice is confined to the two supported CLIs.You use your existing Claude Code and Codex logins; AgentBridge only relays messages between them on your machineraysonmeng.github.io/agent-bridge

Where you use it

In your editor: NoBoth surfaces are terminal: 'abg claude' starts Claude Code with the channel and 'abg codex' starts the Codex TUI. No IDE or editor integration is offered.Start Codex TUI connected to AgentBridge daemon.github.com/raysonmeng/agent-bridge
On the command line: YesThe product is a CLI plus a background daemon; every capability is reached through terminal commands under 'agentbridge' or the 'abg' alias.All commands work with both agentbridge and the short alias abg.github.com/raysonmeng/agent-bridge
In your pipeline: NoNothing acts on a pull request or in CI. The enumerated command set is local-only, with no webhook, PR or pipeline surface.A local developer tool for connecting Claude Code and Codex in one workflowgithub.com/raysonmeng/agent-bridge
In a browser: NoNo hosted dashboard or web app; the GitHub Pages site is marketing and every runtime component runs on your own machine.AgentBridge is a local tool with zero runtime dependencies and no hosted service.raysonmeng.github.io/agent-bridge

Whose machine it runs on

Self-hosted: YesLocal-only by design: an npm-installed CLI and a background daemon on your machine, with no hosted component involved.AgentBridge only relays messages between them on your machine; raw work never leaves the box.raysonmeng.github.io/agent-bridge
Open source: YesMIT licensed, with the whole implementation public in the vendor's GitHub repository.Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to…github.com/raysonmeng/agent-bridge/blob…

What it costs to run

How it meters: NoNothing is metered or billed by the vendor: no seats, tokens, credits or tasks. Your only costs are the Claude Code and Codex subscriptions you already pay for.No. AgentBridge is a local tool with zero runtime dependencies and no hosted service.raysonmeng.github.io/agent-bridge
Free tier: YesFree of charge under the MIT licence; the whole tool is a public npm package you install and run yourself, with no call and no signup.Permission is hereby granted, free of charge, to any person obtaining a copy of this softwaregithub.com/raysonmeng/agent-bridge/blob…
API access: PartlyNo product API for external callers. The programmable surfaces are the MCP tools the plugin exposes to Claude, the local control WebSocket, and --json output from doctor and budget.Claude calls the reply tool -> bridge.ts -> control WS -> daemon.ts -> turn/start injects into the Codex threadgithub.com/raysonmeng/agent-bridge
MCP server: YesMCP is the transport by which the paired agent addresses it: the Claude Code plugin runs an MCP channel whose reply and get_messages tools carry the traffic. There is no remote MCP endpoint.the daemon intercepts Codex output and pushes it to Claude as channel notifications; Claude replies via the reply MCP toolgithub.com/raysonmeng/agent-bridge
Bring your own key: YesEntirely your own inference: it drives the Claude Code and Codex sessions already installed on your machine and needs no account or key of its own.You use your existing Claude Code and Codex logins; AgentBridge only relays messages between them on your machineraysonmeng.github.io/agent-bridge

Buying it for a team

Pooled budget: PartlyNo seat balances exist, but the daemon does track both agents' subscription quota windows jointly and reports drift between the two sides.The daemon polls both agents' account-level 5h and weekly quota. abg budget --json prints the live snapshot: both windows, drift between the two…raysonmeng.github.io/agent-bridge
Admin controls: PartlyConstraints exist but are per-machine developer settings (--safe, AGENTBRIDGE_* environment variables and .agentbridge/config.json), with no admin role and no centrally enforced policy.Set AGENTBRIDGE_FILTER_MODE=full (or the config equivalent) when you do want the unfiltered stream.github.com/raysonmeng/agent-bridge
Audit log: PartlyLocal daemon and Codex-wrapper logs only, tailed with 'abg logs' and stored in a platform state directory. There is no account-level audit trail, because there are no accounts.Tail this pair's daemon log (or the Codex wrapper log with --codex); -f follows, -n N sets the line count (default 100)github.com/raysonmeng/agent-bridge
Single sign-on: NoThere is no identity layer to federate: AgentBridge holds no accounts, and you authenticate to Claude Code and Codex separately as yourself.No. AgentBridge is a local tool with zero runtime dependencies and no hosted service. You use your existing Claude Code and Codex loginsraysonmeng.github.io/agent-bridge

What happens to your code

Opt out of training: PartlyThe bridge itself sends nothing off the machine, so it has nothing to opt out of; what Claude Code and Codex do with your code is governed by those vendors, not by AgentBridge.AgentBridge only relays messages between them on your machine; raw work never leaves the box.raysonmeng.github.io/agent-bridge
Data residency: YesResidency is wherever you run it: every bridge connection is loopback-only and runtime state lives in a local per-platform directory, with no vendor-side store.All connections are on 127.0.0.1 — no external network exposure by default.github.com/raysonmeng/agent-bridge/blob…
Getting out: YesAll state is plain local files you already own: .agentbridge/config.json, checkpoints and logs in a platform state directory. There are no credits to expire.Contents: daemon.pid, status.json, agentbridge.log, killed (sentinel), startup.lockgithub.com/raysonmeng/agent-bridge

32 sourced claims on this page. Checked 2026-09-02. How we check.

Filed under

ClaudeFree tierGPTMulti-agentOpen sourceSelf-hosted

Sources and updates

The pages we read4

3 of 4 independent of the vendor.

Updates2

2026-09-02Capabilities · Commercial terms · Free tier
2026-09-01Capabilities · Commercial terms · Free tier