# AgentBridge — Local MIT bridge that connects Claude Code and Codex as live peers in one session. > Source: The Agents Index — https://theagentsindex.com/agentbridge (structured, researched, re-verified) > Facts last verified: 2026-10-06 AgentBridge is a local, MIT-licensed command-line bridge that connects a running Claude Code session and a running Codex session as persistent peers. Messages, task splits and review comments pass between the two CLIs without a human relaying text. | Fact | Value | | --- | --- | | Website | https://raysonmeng.github.io/agent-bridge | | Price | MIT·runs 100% locally·zero runtime deps·1,700+ tests | | Free tier | Yes | MIT-licensed and free in full; no account or API key is needed for AgentBridge itself. You supply your own Claude Code and Codex logins, and Windows is not an officially supported platform. ## What it does **Writes code** · No — The product is message plumbing, not an authoring agent: the connected Claude Code and Codex CLIs write the code while AgentBridge only forwards their messages. > A bridge that forwards messages between an MCP channel and the Codex app-server protocol > — https://github.com/raysonmeng/agent-bridge · README, 'What this project is / is not' section, 'This project is' list **Debugs and fixes** · No — No diagnosis or fix capability of its own is claimed anywhere; the bridge relays conclusions between two external coding CLIs that do the work. > A local developer tool for connecting Claude Code and Codex in one workflow > — https://github.com/raysonmeng/agent-bridge · README, 'What this project is / is not' section, 'This project is' list **Writes tests** · No — No 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 events > — https://github.com/raysonmeng/agent-bridge · README, Current Limitations, first bullet **Reviews code** · Partly — The bridge does not review anything itself, but its headline workflow routes one agent's diff to the other for review mid-turn, so review happens between the two connected CLIs rather than in the product. > Codex implements; Claude reviews the diff in real time inside the same session and pushes change requests straight back into Codex's thread. Two providers check each other's work without copy-paste. > — https://raysonmeng.github.io/agent-bridge/ · Three workflows section, 'Cross-review, inside the session' card **Whole-codebase work** · No — The 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. > — https://raysonmeng.github.io/agent-bridge/ · FAQ, 'Do I need an account or API keys for AgentBridge itself?' **Ships it** · No — Nothing in the documented CLI surface deploys anything: the commands cover init, launching the two agents, logs, doctor, budget and kill, and the tool is explicitly not a hosted system. > A hosted service or multi-tenant system > — https://github.com/raysonmeng/agent-bridge · README, 'What this project is / is not' section, 'This project is not' list **Done-for-you service** · No — There is no service engagement at all: you install an npm CLI and run it locally, and the vendor explicitly disclaims running anything hosted for you. > No. AgentBridge is a local tool with zero runtime dependencies and no hosted service. > — https://raysonmeng.github.io/agent-bridge/ · FAQ, answer to "Do I need an account or API keys for AgentBridge itself?" ## How much it does unattended **Runs autonomously** · Partly — The bridge is built for unattended pairs - permission prompts are skipped by default and quota relay resumes work automatically - but a human still steers and assigns the task. > `abg claude` launches with `--dangerously-skip-permissions` and `abg codex` launches with `--yolo` by default. > — https://github.com/raysonmeng/agent-bridge · Quick Start, WARNING callout under the five install steps **Multi-agent** · Yes — This is the whole product: two full agents from different providers stay live as peers, negotiate task splits and push messages into each other's running turns. > 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 working), not only at call boundaries. > — https://github.com/raysonmeng/agent-bridge · "Why not just…" section, one-way delegation plugin bullet **Agent permissions** · Partly — Both launchers inject maximum-permission flags by default so an unattended pair never stops to ask; a --safe flag or environment variable restores normal prompting, and the vendor warns to use it only in a trusted workspace. > Add --safe (or set AGENTBRIDGE_SAFE=1) to launch with normal permission prompts. AgentBridge runs entirely on your machine and is not a hardened boundary between untrusted tools. > — https://raysonmeng.github.io/agent-bridge/ · FAQ, answer to "Is it safe? What does --safe do?" (structured FAQ data) ## Which models it runs on **Claude** · Yes — Claude Code is one of the two agents the shipping integration connects, loaded through an AgentBridge plugin and an MCP channel in the live session. > Today the shipping in-session integration connects Claude Code and Codex. > — https://raysonmeng.github.io/agent-bridge/ · FAQ, answer to "Which agents does it support?" **GPT** · Yes — OpenAI's Codex CLI is the other half of the shipping pair; the daemon proxies the Codex app-server protocol and injects turns into its running thread. > Today the shipping in-session integration connects Claude Code and Codex. > — https://raysonmeng.github.io/agent-bridge/ · FAQ, answer to "Which agents does it support?" **Open models** · No — The vendor states the shipping integration as a closed pair of two proprietary vendor CLIs, with further adapters only a roadmap vote; no open-weight model or local-inference backend is wired in 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; OpenCode, Gemini CLI, and others are candidates. > — https://raysonmeng.github.io/agent-bridge/ · FAQ, 'Which agents does it support?' (structured FAQ answer) **Model choice** · Partly — You bring whichever models your Claude Code and Codex logins give you, but the bridge itself speaks to exactly two agent CLIs today, with further adapters put to a public vote. > A public adapter vote (issue #212) decides which agent is wired in next; OpenCode, Gemini CLI, and others are candidates. > — https://raysonmeng.github.io/agent-bridge/ · FAQ, answer to "Which agents does AgentBridge support?" (structured FAQ data) ## Where you use it **In your editor** · No — Both 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. > — https://github.com/raysonmeng/agent-bridge · README, CLI Reference table, 'abg codex [args...]' row **On the command line** · Yes — The entire product is a terminal CLI: a globally installed abg/agentbridge binary with init, claude, codex, resume, doctor, budget, logs and kill subcommands. > `abg` is a short alias for `agentbridge`. Both agents launch with max-permission defaults so an unattended pair doesn't stop for each prompt. > — https://raysonmeng.github.io/agent-bridge/ · Quick start section, note under the five numbered steps **In your pipeline** · No — Nothing 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 workflow > — https://github.com/raysonmeng/agent-bridge · README, 'What this project is / is not' — 'This project is' list, first bullet **In a browser** · No — There is no hosted surface of any kind - no account, no dashboard, no service - only a local daemon and two terminals on your own machine. > No. AgentBridge is a local tool with zero runtime dependencies and no hosted service. > — https://raysonmeng.github.io/agent-bridge/ · FAQ, answer to "Do I need an account or API keys for AgentBridge itself?" ## Whose machine it runs on **Self-hosted** · Yes — Self-hosting is the only mode: the daemon and bridge run as local processes on macOS or Linux, and the vendor says work never leaves the machine. > AgentBridge only relays messages between them on your machine; raw work never leaves the box. > — https://raysonmeng.github.io/agent-bridge/ · FAQ, answer to "Do I need an account or API keys for AgentBridge itself?" **Open source** · Yes — The full TypeScript source sits in a public GitHub repository under the MIT licence, with a documented local-development setup for modifying it. > If you want to modify AgentBridge source code, use the local development setup instead: > — https://github.com/raysonmeng/agent-bridge · README, "Install for local development" section (repo is MIT licensed) ## What it costs to run **How it meters** · Partly — The bridge itself meters nothing and charges nothing; your real cost is the two upstream subscriptions it relays between, whose 5h and weekly windows it only reports on. > No. AgentBridge is a local tool with zero runtime dependencies and no hosted service. You use your existing Claude Code and Codex logins > — https://raysonmeng.github.io/agent-bridge/ · FAQ, 'Do I need an account or API keys for AgentBridge itself?' **Free tier** · Yes — Free and MIT-licensed with no sales call, no account and no API key for the bridge itself - one npm install and it runs. > No. AgentBridge is a local tool with zero runtime dependencies and no hosted service. You use your existing Claude Code and Codex logins > — https://raysonmeng.github.io/agent-bridge/ · Structured FAQ data, "Do I need an account or API keys for AgentBridge itself?" **API access** · Partly — There is no hosted HTTP API, but the CLI exposes machine-readable output for scripting - doctor and budget both take --json - alongside local WebSocket control ports. > `abg budget [--json]` | Both agents' subscription quota snapshot (5h/weekly windows, drift, pause state) > — https://github.com/raysonmeng/agent-bridge · README, CLI Reference table, abg budget row **MCP server** · Yes — MCP is the Claude-side transport: the bridge is an MCP client loaded as a Claude Code plugin, and Claude answers the other agent through a reply MCP tool. > Claude replies via the `reply` MCP tool, and the bridge injects the reply into the Codex thread as a `turn/start`. > — https://github.com/raysonmeng/agent-bridge · README, Features list, "Bidirectional Claude ↔ Codex messaging" bullet **Bring your own key** · Yes — Bring-your-own everything: the bridge holds no credentials and uses the Claude Code and Codex logins already on your machine, with inference going through those accounts. > You use your existing Claude Code and Codex logins; AgentBridge only relays messages between them on your machine > — https://raysonmeng.github.io/agent-bridge/ · FAQ, answer to "Do I need an account or API keys for AgentBridge itself?" ## Buying it for a team **A company can buy it** · No — There 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 system > — https://github.com/raysonmeng/agent-bridge · README, 'What this project is / is not' section **Pooled budget** · Partly — There 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 sides, and pause state. > — https://raysonmeng.github.io/agent-bridge/ · Quota relay section, 'Both windows in one glance' card **Admin controls** · Partly — The operator constrains the pair locally through launch flags, environment variables and a per-project config file, but there is no central administrator console or org-wide policy. > Both launchers consume the wrapper flag `--safe` (it is never forwarded): it disables the max-permission defaults for that launch. > — https://github.com/raysonmeng/agent-bridge · README, "Owned flags" subsection **Audit log** · Partly — Per-pair daemon and Codex wrapper logs are written to a local state directory and tailed with abg logs, but these are diagnostic logs rather than tamper-evident administrative audit records. > `abg logs [--codex] [-f] [-n N]` | Tail this pair's daemon log (or the Codex wrapper log with `--codex`); `-f` follows, `-n N` sets the line count (default 100) > — https://github.com/raysonmeng/agent-bridge · README, CLI Reference table, abg logs row **Single sign-on** · No — There 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 logins > — https://raysonmeng.github.io/agent-bridge/ · FAQ, 'Do I need an account or API keys for AgentBridge itself?' ## What happens to your code **Opt out of training** · Partly — The bridge itself never transmits your work anywhere, so it adds no training exposure, but whether your code trains a model is governed by your Claude and Codex accounts, which the vendor does not speak for. > AgentBridge only relays messages between them on your machine; raw work never leaves the box. > — https://raysonmeng.github.io/agent-bridge/ · FAQ, answer to "Do I need an account or API keys for AgentBridge itself?" **Data residency** · Yes — All residency is your own machine: runtime state, logs and config live in platform state directories on macOS or Linux and no vendor cloud is involved. > AgentBridge runs entirely on your machine and is not a hardened boundary between untrusted tools. > — https://raysonmeng.github.io/agent-bridge/ · Structured FAQ data, "Is it safe? What does --safe do?" **Getting out** · Yes — Everything is already local plain files - a per-project config directory plus pid, status and log files in a documented state directory - so there is no export step and no credit to expire. > Contents: `daemon.pid`, `status.json`, `agentbridge.log`, `killed` (sentinel), `startup.lock` > — https://github.com/raysonmeng/agent-bridge · README, "State Directory" section, contents line _Not established: Certifications_ ## Provenance 3 of 4 sources behind the facts above are independent of the vendor. ## Coverage elsewhere None recorded. Every fact above comes from the vendor and is not independently corroborated yet. ## What changed - 2026-10-06 — Capabilities · Commercial terms · Free tier - 2026-10-05 — Capabilities · Commercial terms · Free tier - 2026-10-04 — Capabilities · Commercial terms · Free tier · Price - 2026-10-03 — Capabilities · Commercial terms · Free tier · Price - 2026-10-01 — Capabilities · Commercial terms · Free tier · Price - 2026-09-02 — Capabilities · Commercial terms · Free tier ## Correcting this record Anything marked "Not established" is a question we have not answered — it is not a No. If you represent AgentBridge, or anything here is wrong or out of date, you can supply it with a source: the operations this site accepts from an agent are listed at https://theagentsindex.com/.well-known/agents.json — that catalogue needs no key to read, and each entry states its own auth. Supplying evidence is free. Nothing that can be bought here changes what a field says, whether a gap is closed, or how this record is scored.