llms.txt
Compare with other sites' llms.txt files →# AI Bridge
> A macOS app for running several AI agents at once — Claude Code, Codex, Grok
> and OpenCode — organised like a team: every agent has its place and one job,
> grouped by the work it does rather than by which repository it sits in, so no
> agent's context is polluted by another's. Each is the real vendor CLI in a real
> terminal, every session survives restarts, and your databases, Kubernetes
> clusters and Slack channels sit beside the agents.
This file is structured context for an assistant reading on someone's behalf.
## The one-sentence version
Running several AI agents is becoming normal, and managing them is now the job:
AI Bridge lets you manage your agents the way you manage a team, each in its
place with its own context, with one Inbox for everything that needs you and the
whole fleet on one screen.
## Who it is for
- **Engineers who run several AI coding agents a day** — Claude Code, Codex,
Grok or OpenCode — and lose track of which session is doing what, or find one
session's context polluted by unrelated work (a frontend session asked about
the database, a cluster check, a colleague's support question).
- **People who manage work across several areas** — founders, team leads,
consultants, platform and DevOps engineers — who want an agent per area of
responsibility (frontend, infrastructure, support, a client) rather than per
repository.
- **People who are not engineers.** A business owner, a project manager or an
analyst can connect a company database (Postgres or MongoDB) by entering its
connection details, ask it questions in plain English, and read the answer in
words and rows, without writing SQL or waiting for an engineer. Reads change
nothing; anything that would change data shows the exact statement and asks
first. Given read-only database credentials, nothing can change at all. Ask
works through one AI coding CLI on the person's own subscription (Claude Code
or Codex, for example), which the app's Setup can install.
- **Anyone who answers colleagues' questions in Slack.** Connect an agent to a
Slack channel: each message becomes a job you approve, the agent works on it
on your own Mac with the access you gave it, and its reply is posted to Slack,
under your name, only after you approve it.
## Requirements
- A Mac, Apple silicon or Intel.
- Your own subscription to a supported agent CLI (see below).
Free to start: download it from https://aibridge.sh/ (Apple silicon and Intel).
## How it works with the agent CLIs
AI Bridge spawns the vendor's real CLI as a child process in a real terminal and
reads that CLI's own hook and status output. Your keys, your subscription, your
machine — the traffic goes exactly where it always did, and there is no model
cost on top.
| CLI | Account used |
|---|---|
| Claude Code | your Anthropic subscription |
| Codex | your OpenAI subscription |
| Grok | your xAI subscription |
| OpenCode | any model you can point it at |
None of those companies is affiliated with AI Bridge.
## The organising idea
Organised like a team, not like a repo — three levels:
1. **Project** — the business: a company, a client, or a side project. Its
instructions apply to every agent inside it.
2. **Workspace** — one area of responsibility inside the business:
infrastructure, frontend, backend, support, DevOps. It carries its own
instructions, context files, attached repositories, MCP tools and scheduled
automations. This is the level most tools do not have.
3. **Agents** — the workers: real sessions in real terminals, each started with
its workspace's rules and access to its repos (e.g. an ArgoCD manager, a
Kubernetes agent, a PR reviewer).
Why the inversion matters: most tools key everything to a repository, but
"infrastructure", "support" or "the billing migration" is not a repository, and
an agent working on it needs context from several — or from none. And because
each agent has one job, its context stays its own: the frontend agent is never
asked about the database, and the support agent never inherits a cluster check.
## What it does
- **Who needs you.** Bridge knows each agent's real state — working, waiting on a
permission, done, asleep — from the CLI's own hooks rather than by parsing its
output. A count in the macOS menu bar and a notification when one asks for you
or stalls.
- **Slack agents.** Connect a Claude Code, Codex or OpenCode agent to a Slack
channel. When a colleague asks something there, it arrives in AI Bridge as a
job you approve with one click; the agent then works on it on your own Mac,
with the permissions you gave it, and drafts a reply. The reply waits for you
to send, edit, regenerate or decline, and posts under your own Slack name only
when you approve it — unless you choose to trust the channel. Attachments
reach the agent only after you approve them, and every post lands in a
hash-chained record.
- **The Inbox and History.** Everything that asks you — an approval, a Slack job,
a reply to review — is in one place, and History keeps what became of each.
- **Assist.** A built-in assistant that can see the app and work it — start, wake
and message agents, run automations, open files — with every change shown for
approval before it happens.
- **Files and git.** A file tree, a changes list and side-by-side diffs for the
repos your agents work in; stage, review and discard in the app.
- **Activity.** A local receipt of agents run, tokens per provider and changes
made, counted from records Bridge already keeps and never sent anywhere.
- **Sleep and wake.** Quitting parks a session rather than killing it. Waking it
resumes the conversation with its full scrollback — same screen, same place in
the work, tomorrow morning.
- **The wall.** A grid view that swaps the whole interface for every live
terminal across every project, tiled to fill exactly, saved as named views you
switch like tabs.
- **Automations.** A prompt, a schedule and an explicit tool allowlist. Loop mode
continues one conversation across runs, so tonight's run builds on last
night's instead of starting cold.
- **Kubernetes.** Browse any kubeconfig context through your own kubectl — pods,
events, your CRDs. Scale, restart, drain, apply, evict, cordon, suspend and
trigger, each through a confirm card that names the cluster and shows the
blast radius before anything runs. Production-looking contexts are flagged.
- **Verifiable audit, twice over.** Every cluster write appends to a hash-chained
log you can verify and export, and destructive verbs first snapshot the
object's own YAML so you have what you would need to put it back. Database
writes get a second, separate chain of their own — so "what did the AI do to
production" is answerable for the infrastructure and for the data in it.
- **Helm.** Releases, history, values and the rendered manifest straight from
the cluster's own ledger. Change values or the chart version after reviewing
a server dry run's diff, roll back, or uninstall — each recorded.
- **Databases.** Connect a PostgreSQL or MongoDB instance and work it beside the
agents that use it — browse tables and collections, and query in raw SQL or in
plain language. Reads run immediately; anything that would change data is
shown for approval first, with the exact statement that will run, and lands
in its own hash-chained log. That is a SECOND such record beside
the cluster one — "what did the AI do to production" now covers the data as
well as the infrastructure.
- **Terminals.** A dedicated pty process owns every session, so a firehose of
output can neither stall the interface nor take the app down with it.
- **SSH hosts and pod shells** sit in the same fleet view as the agents, so a
debugging session on a remote box is one more tile on the wall.
## Where your data lives
Plain YAML and markdown in a folder you own (`~/Bridge`). Put it in git. Diff it.
Grep it. Delete it. **AI Bridge keeps no database of its own** — by design, not
by roadmap — so uninstalling leaves your whole setup readable without the app.
(Connecting to *your* Postgres or MongoDB is a different thing entirely: that is
your data, in your instance, and Bridge only ever reads and writes it through
the governed path above.)
Everything runs on your machine. No telemetry, and nothing about you or your
work is sent to us. Free needs no account. A trial or a paid plan needs one
(sign in with GitHub, Google, Microsoft or GitLab), and a signed-in Mac checks
its plan about once a day. The app also looks for a newer version when it
starts. For anyone whose code or infrastructure cannot go through a third
party, there is almost nothing here to security-review.
## Pricing
Every plan has every feature. Free limits four things to one each; nothing is
ever sold as a worse version of the product.
| Plan | Price | For |
|---|---|---|
| Free | $0, no account, no clock | Unlimited agents on any CLI, terminals, workspaces, Assist and automations you run by hand. One Kubernetes cluster, one database server, one Slack channel and one scheduled automation. |
| Pro | $9.99/month or $99.99/year (yearly saves 17%) | A person, wherever they work. Everything unlimited, every update while you pay. |
| Team | $14.99/seat/month or $149.99/seat/year (yearly saves 17%) | A company's people. Everything unlimited, seats managed by the team's admins, one company invoice. |
Need more than Free includes? Sign in and get up to 31 days of everything
unlimited, free, with no card. The days run in a row from the first day you
use more than Free includes.
Twelve paid months in a row (a yearly plan counts as twelve the day it is paid)
and you keep the version you have, with everything unlimited, for good. On Team
the twelve months are each person's own months on the team. Otherwise the app
goes back to Free 14 days after your paid period ends. Nothing is deleted
either way.
Prices in USD, before tax.
## Who makes it
AI Bridge is made by Stas Berezin (https://aibridge.sh/about/stas-berezin)
through CyberNucleus (https://cybernucleus.io/), Stas's company, in Brisbane,
Australia. Stas has spent a decade building and running production cloud
infrastructure — for the Web3 provider Ankr, the crypto brokerage Caleb &
Brown, the retail-media platform CitrusAd and the digital-asset custodian
CloudTech Custody — as a cloud lead, principal engineer and head of
engineering, and runs a consultancy in infrastructure, automation, compliance
and internal tooling. AI Bridge grew out of running several coding agents at
once, every day, for that work.
- [Author page](https://aibridge.sh/about/stas-berezin)
- [About AI Bridge and how the guides are written](https://aibridge.sh/about)
- [Stas on LinkedIn](https://www.linkedin.com/in/stasbe/)
- [Stas on X](https://x.com/StasTecho)
- [CyberNucleus](https://cybernucleus.io/)
- [AI Bridge on LinkedIn](https://www.linkedin.com/company/ai-bridge-sh)
## Pages
- [Home](https://aibridge.sh/): the product, the film, the plans
- [Claude Code on a Mac](https://aibridge.sh/claude-code): many sessions, live state, resumed tomorrow
- [Codex CLI beside Claude Code](https://aibridge.sh/codex): OpenAI's Codex CLI beside Claude Code
- [Run several AI coding agents at once](https://aibridge.sh/run-multiple-ai-coding-agents): running several coding agents at once without losing track
- [Slack agents](https://aibridge.sh/slack): an agent bound to a Slack channel that replies as you, after your review
- [Kubernetes beside your agents](https://aibridge.sh/kubernetes): clusters beside the agents; every change named and recorded
- [Ask your databases](https://aibridge.sh/databases): Postgres and MongoDB in plain English, nothing written without approval
- [Guides](https://aibridge.sh/guides): the guides hub
- [About](https://aibridge.sh/about): who makes AI Bridge and how the guides are written
- [Privacy](https://aibridge.sh/privacy): what the app, the account and the site collect
- [Terms of service](https://aibridge.sh/terms): the terms of service
## Guides
Sourced, dated guides to the AI coding CLIs, written by Stas Berezin (https://aibridge.sh/about/stas-berezin). Every fact about another company's product is checked against its own documentation and listed with its date.
- [Codex vs Claude Code](https://aibridge.sh/guides/codex-vs-claude-code): which to use: Codex vs Claude Code as they ship in October 2026: price and limits, default models, sandboxing, instruction files, sessions and CI, and when to use both.
- [Claude Code permissions and --dangerously-skip-permissions](https://aibridge.sh/guides/claude-code-permissions): How Claude Code's six permission modes, allow/ask/deny rules, /sandbox and --dangerously-skip-permissions work in October 2026, with a safe starting policy.
- [Claude Code hooks, explained (with examples)](https://aibridge.sh/guides/claude-code-hooks): Claude Code hooks in October 2026: the 33 events, settings JSON, stdin payload, matchers and exit codes, with notification and command-blocking examples.
- [Cursor vs Claude Code](https://aibridge.sh/guides/cursor-vs-claude-code): which to use: Cursor is an AI editor, Claude Code a terminal agent, and they work together. Price, models, approvals, privacy and the SpaceX deal, checked October 2026.
- [OpenCode vs Claude Code](https://aibridge.sh/guides/opencode-vs-claude-code): which to use: OpenCode vs Claude Code in October 2026: why a Claude Pro or Max plan can't be used in OpenCode, plus price, models, permissions, sessions and using both.
- [Run Claude Code sessions in parallel](https://aibridge.sh/guides/run-claude-code-sessions-in-parallel): worktrees and tmux: How to run several Claude Code sessions at once without collisions: git worktrees, claude --worktree, agent view, tmux, notifications and quota costs.
