AgentBridge

What it does

Writes code

NoAgentBridge itself authors nothing; it is a message relay between an MCP channel and the Codex app-server, and the code is written by the two CLIs it connects.A bridge that forwards messages between an MCP channel and the Codex app-server protocolgithub.com/raysonmeng/agent-bridge ↗

Debugs and fixes

NoNo diagnosis or fix capability of its own: the vendor describes a local bridge forwarding messages, with any debugging done by Claude Code or Codex on their own.A local developer tool for connecting Claude Code and Codex in one workflowgithub.com/raysonmeng/agent-bridge ↗

Writes tests

NoThe product generates no tests; its own 1,700+ test suite is the project's internal coverage, not a capability it offers users.A bridge that forwards messages between an MCP channel and the Codex app-server protocolgithub.com/raysonmeng/agent-bridge ↗

Reviews code

PartlyIt enables cross-review rather than performing it: Codex implements and Claude reviews the diff mid-session, with AgentBridge carrying the change requests back into the other agent's thread.Codex implements; Claude reviews the diff inside the same session and pushes change requests straight back into Codex's thread. Two providers check…github.com/raysonmeng/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 path is offered; the vendor's complete CLI reference covers daemon, pair, log and budget commands only, and shipping is left to whatever the connected agents do.Only forwards `agentMessage` items, not intermediate `commandExecution`, `fileChange`, or similar eventsgithub.com/raysonmeng/agent-bridge ↗

Done-for-you service

NoNobody runs it for you: it is software you install with npm and run on your own machine, with the vendor explicitly ruling out a hosted or multi-tenant offering.No. AgentBridge is a local tool with zero runtime dependencies and no hosted service.raysonmeng.github.io/agent-bridge ↗

How much it does unattended

Runs autonomously

YesDesigned for unattended pairs: both agents launch with permission prompts skipped by default so an overnight run keeps going, and the vendor frames the human as a steerer rather than an approver.By default `abg claude` launches with `--dangerously-skip-permissions` and `abg codex` with `--yolo`. This is deliberate: an unattended agent pair…raysonmeng.github.io/agent-bridge ↗

Multi-agent

YesMulti-agent coordination is the entire product: two full CLIs stay live as persistent peers, negotiate task splits between themselves and push messages into each other mid-turn.AgentBridge keeps both agents live as persistent peers, and either side can push a message mid-turn (a review comment lands while the other is still…github.com/raysonmeng/agent-bridge ↗

Agent permissions

YesPermission posture is explicit and switchable: maximum permissions are the default for unattended pairs, and --safe or AGENTBRIDGE_SAFE=1 restores normal per-action prompts.Add `--safe` (or set `AGENTBRIDGE_SAFE=1`) to launch with normal permission prompts. AgentBridge runs entirely on your machine and is not a hardened…raysonmeng.github.io/agent-bridge ↗

Which models it runs on

Claude

YesClaude Code is one of the two supported sides, wired in as a Claude Code plugin whose MCP channel carries replies into the bridge.Today the shipping in-session integration connects Claude Code and Codex.raysonmeng.github.io/agent-bridge ↗

GPT

YesOpenAI's Codex CLI is the second supported side, reached through the Codex app-server protocol, so GPT-family models arrive via that CLI's own login.Today the shipping in-session integration connects Claude Code and Codex.raysonmeng.github.io/agent-bridge ↗

Open models

NoThe vendor states the shipping integration as a closed pair, Claude Code and Codex, with other runtimes still at the voting stage, so no open-weight path exists today.Today the shipping in-session integration connects Claude Code and Codex. A public adapter vote (issue #212) decides which agent is wired in next;…raysonmeng.github.io/agent-bridge ↗

Model choice

PartlyYou choose models inside each CLI, not in AgentBridge, and the bridge is agnostic about which side plans or executes; but the two runtimes it supports are fixed at Claude Code and Codex.which agent plans vs executes is your call — the bridge doesn't hard-code a boss; you can just as well have Codex drive Clauderaysonmeng.github.io/agent-bridge ↗

Where you use it

In your editor

NoBoth surfaces are terminal programs: a Claude Code session and a Codex TUI started from shells, with no editor or IDE extension in the vendor's install path or command list.`abg codex [args...]` | Start Codex TUI connected to AgentBridge daemongithub.com/raysonmeng/agent-bridge ↗

On the command line

YesIt is a CLI first: the agentbridge/abg binary covers init, launch, resume, pairs, doctor, budget, logs and kill, all run from a terminal.All commands work with both `agentbridge` and the short alias `abg`.github.com/raysonmeng/agent-bridge ↗

In your pipeline

NoNo pull-request or CI surface: the documented command inventory is local and interactive, and even quota resume is defined as happening inside the original interactive TUI.Automatic resume — when the paused side's window refreshes, the bridge resumes the task in the original interactive TUIgithub.com/raysonmeng/agent-bridge ↗

In a browser

NoThere is no hosted web application; the vendor rules out a hosted service outright and the only processes are a local foreground client and a local daemon.A hosted service or multi-tenant systemgithub.com/raysonmeng/agent-bridge ↗

Whose machine it runs on

Self-hosted

YesSelf-hosting is the only mode: the daemon, proxy and plugin server all run on your machine over loopback ports, and the vendor states work never leaves the box.AgentBridge only relays messages between them on your machine; raw work never leaves the box.raysonmeng.github.io/agent-bridge ↗

Open source

YesFull source is public on GitHub under the MIT licence, including the daemon, CLI, plugin and Docker harnesses, and local development builds are documented.MIT·runs 100% locally·zero runtime depsraysonmeng.github.io/agent-bridge ↗

What it costs to run

How it meters

NoAgentBridge meters nothing and charges nothing; your cost is whatever your existing Claude Code and Codex subscriptions cost, which the bridge only observes for quota relay.You use your existing Claude Code and Codex logins; AgentBridge only relays messages between them on your machine.raysonmeng.github.io/agent-bridge ↗

Free tier

YesInstall and use need no account, key or sales call: one npm global install of an MIT-licensed package, with no paid tier published anywhere on the site.No. AgentBridge is a local tool with zero runtime dependencies and no hosted service.raysonmeng.github.io/agent-bridge ↗

API access

PartlyNo public HTTP API, but it is drivable programmatically: local control and proxy WebSocket ports are configurable and doctor and budget emit --json for scripting.`abg budget [--json]` | Both agents' subscription quota snapshot (5h/weekly windows, drift, pause state)github.com/raysonmeng/agent-bridge ↗

MCP server

PartlyMCP is how the Claude side is wired, with a reply tool and a get_messages fallback queue, but these serve the bridge's own plugin channel rather than exposing the product to arbitrary MCP clients.Claude replies via the `reply` MCP tool, and the bridge injects the reply into the Codex thread as a `turn/start`.github.com/raysonmeng/agent-bridge ↗

Bring your own key

YesYou bring your own agent accounts and your own machine: AgentBridge adds no credentials of its own and runs entirely on local infrastructure you control.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

A company can buy it

NoThere is nothing to purchase and no vendor relationship to sign: it is an MIT npm package each developer installs on their own machine, with no hosted or multi-tenant offering.This project is not: - A hosted service or multi-tenant systemgithub.com/raysonmeng/agent-bridge ↗

Pooled budget

PartlyThere is no billing balance to pool, but the daemon does treat the two agents' subscription quotas as one coordinated budget, polling both windows and reporting drift between them in a single snapshot.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-launch and local, not organisational: --safe or explicit permission flags suppress the max-permission defaults, and owned flags are rejected outright.Both launchers consume the wrapper flag `--safe` (it is never forwarded): it disables the max-permission defaults for that launch.github.com/raysonmeng/agent-bridge ↗

Audit log

PartlyLocal daemon and Codex wrapper logs are kept and tailable with abg logs, but these are operational diagnostics on one machine, not an administrative audit trail.`abg logs [--codex] [-f] [-n N]` | Tail this pair's daemon log (or the Codex wrapper log with `--codex`)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

PartlyAgentBridge sends your code nowhere itself, since it relays only locally and never ships raw work off the machine; but training policy for the Claude and Codex accounts it fronts stays with those vendors.AgentBridge is a local tool with zero runtime dependencies and no hosted service.raysonmeng.github.io/agent-bridge ↗

Data residency

YesResidency is your own disk: project config lands in .agentbridge/ and runtime state in a platform directory under your home, with the bridge talking over 127.0.0.1 ports.State directory for pid, status, logs (macOS: `~/Library/Application Support/agentbridge/`, Linux: `$XDG_STATE_HOME/agentbridge/`)github.com/raysonmeng/agent-bridge ↗

Getting out

YesEverything it writes is a plain local file you already own — project config, collaboration contract, logs and checkpoints — and there are no credits that could expire.`config.json` | Machine-readable project config (Codex ports, turn coordination, idle shutdown)github.com/raysonmeng/agent-bridge ↗

Not established: Certifications