Rohan Paul, who writes about AI on X, described a problem most engineering teams know too well. When he pings a teammate for context, he wrote, the answer is often “already sitting in their coding agent’s session.” Alook features posts on its homepage, and they land because the workaround is so clumsy. The teammate becomes a relay: paste the question into Claude Code, wait, copy the reply back into Slack, repeat.
alook.ai is an open-source platform built to remove that relay. It gives local agents like Claude Code, Codex, Cursor, and OpenCode a handle and an inbox, then seats them in shared channels next to people. You ask the agent directly, and the answer shows up in a thread everyone can read.
That convenience brings a shift that’s easy to miss. An agent in a shared room still runs on one person’s laptop, inside that person’s repo, with that person’s credentials. Once teammates can message it, they can steer it. So the first question to ask about any human-agent collaboration tool is what the agent will do when someone other than its owner asks.
What Alook Is, in Plain Terms
Every agent on Alook gets a persistent identity: a handle, an inbox, and memberships. Teammates reach it in servers, channels, and DMs the same way they’d reach a person. A small daemon on your computer keeps the agent connected to Alook’s app over a WebSocket. The agent does its actual work in your local project folder.
Alook supplies no models of its own. Your agent keeps whatever model and setup you already gave it. The code is Apache-2.0 on GitHub, and a single npx @alook/app onboard command sets up a self-hosted copy that serves locally.
| Piece | Where it lives |
| Your agent (Claude Code, Codex, Cursor, OpenCode or Pi) | Your computer |
| Alook daemon | Your computer, connecting out to Alook |
| Rooms, messages and files shared in them | Alook’s hosted app, or your server if you self-host |
| Models | Whatever your agent already uses |
Two details from Alook’s FAQ matter later. Files on your machine don’t sync automatically; Alook stores only what someone shares in a conversation. And the agent keeps working after you close the browser tab, as long as its computer stays on and online.
The Message Becomes the Instruction
Here’s a made-up but very plausible Tuesday. Marco shares his Codex agent in the #checkout channel so the team can ask it about the payment flow. A product manager types: “@marco-codex can you remove the old discount flags so we can ship Friday?”
The agent is helpful. It opens Marco’s checkout, deletes the flags, runs the tests, and reports back with a tidy summary. Nobody in the thread knew that two of those flags still guarded a live partner promo. Nothing malicious happened. A teammate made a reasonable request, and the agent carried it out with Marco’s write access.
The second risk is quieter. Someone pastes a vendor’s error log into the channel and asks the agent to diagnose it. Every line of that log is now input to software with a shell. Prompt injection, where text hidden inside data gets read as a command, sits at the top of OWASP’s list of LLM application risks. Models still struggle to tell content they should read from instructions they should follow.
In a solo terminal, you are the only person feeding your agent. In a shared room, everyone with access feeds it, along with anything they paste.
Where Alook Draws the Line, and Where You Have To
Alook doesn’t ignore this. Its FAQ says an agent may pause for your input or permission, and you can check its updates and reply from your phone. Its own tagline is about sharing agents “with people you trust.” Read that literally, because trust is the access model.
Below Alook sits your agent’s own permission settings, and those still decide what actually happens on your machine. Claude Code asks before editing files or running commands unless you launch it with –dangerously-skip-permissions. Codex has a full-auto mode with the same trade-off. Whatever freedom you gave the agent for your own convenience, a channel message inherits.
A sensible default before you share anything:
| What teammates ask for | How to set it up |
| Explanations, code questions, summaries of recent changes | Let the agent answer freely |
| Edits on a feature branch | Owner approves each write |
| Pushes, deploys, secrets, payments, customer data | Keep these out of shared agents entirely |
Teams that tier approvals by risk across their agent workflows tend to land on the same split. One extra habit helps a lot: point the shared agent at a separate clone or worktree with a narrowly scoped token, rather than the checkout holding your unpushed work.
Why Shared Rooms Still Beat Solo Terminals
None of this argues for keeping agents locked away. The single-player setup has real costs too. Context dies when a terminal closes, and the thing your agent figured out at 6 p.m. never reaches the person who needs it at 9 a.m.
In a shared room, that history stays put. A developer can ask one agent to draft an implementation and tag a second agent to review the diff, while a product manager asks a scoping question in the same thread. The next morning, the review notes are waiting in the channel. Nobody has to forward a log or retell the plan.
Slack and Discord support bots too, and Slack supports AI agents. Alook’s FAQ draws its own line elsewhere: the agent runs on your computer, with your projects and tools already configured. Cloud agent services hand you a remote machine instead. The local approach keeps your codebase where it lives. The catch is that the host computer has to stay on and online for the agent to answer anyone.
Who Should Try It, and Who Should Wait
Alook fits teams already running two or three coding agents and tired of shuttling messages between them. Solo builders juggling Claude Code and Codex get one inbox instead of three windows. Open-source maintainers can let trusted contributors query an agent that already knows the repo.
Wait if your laptops sleep half the day, since a sleeping host means a silent agent. Teams under strict code-handling rules should start with the self-hosted version and read exactly what gets stored, because shared messages and files live on whichever server runs the rooms.
Getting Started Without Lending Out Your Whole Machine
You can sign up on the hosted service and pair a machine, or run the onboarding command to self-host. Alook offers a free plan, plus paid Studio and House plans that scale with active bots and online machines. Release notes and new runtime support show up first on @alook_ai.
Start small. Share one agent in one channel, limit it to read-only questions for a week, and watch what people ask. The requests teammates send in that first week will show you where the approval gates belong far better than any policy written in advance.
When 14 agents traded messages at a New York founders’ dinner last month, one of them dismissed the exercise as “protocol theater.” None of them had the authority to decide anything. Agents in an Alook room are different: they can edit real code on a real machine. That’s the whole reason to bring them in. It’s also why the first @mention should come after you’ve decided how much of your laptop you’re willing to lend.
Related: 8 Best Wiz Code Alternatives for AI-Powered Code Security in 2026
