An AI Dev Team in Your Pocket: Multi-Agent Coding with AI Office
Beyond autocomplete: PocketCode's AI Office turns a brief into a team of specialized AI agents that plan, split, build and review your project — with real tools — from your Android phone.
Most "AI in your IDE" means autocomplete and a chat box. Useful — but it's still you doing the orchestration: deciding what to build, splitting it into steps, wiring the pieces, and checking the result.
PocketCode's AI Office flips that. You write a brief in plain language, and a team of specialized AI agents plans it, breaks it into tasks, executes them with real tools, and reviews its own work — while you watch. On your phone.
From a sentence to a working session
You start in the Lobby, a two-page wizard:
- What to build — write the brief (or tap an example, or pick one of the 13 templates) and, if you want, link an existing project so the agents read it first. Analyze reads that project in depth.
- Your team — name the project, switch roles on or off and press Start Office. Adjust details opens the Config sheet: language, platform, framework, code style, design system, workflow playbook, Quick Mode, QA reviews, advanced limits and the AI provider.
Hit Start Office and the office goes to work: the Project Manager may ask you three to five clarifying questions (Quick Mode skips them), rates the complexity, and produces a plan you review before the work runs.
The team: eight specialists
AI Office isn't one model pretending to be many. It's eight distinct roles, each with its own focus:
| Role | Focus |
|---|---|
| Project Manager | Planning, decomposition, prioritization (always on the team) |
| Architect | Architecture, system design |
| Frontend Dev | UI, layouts, animation |
| Backend Dev | APIs, business logic, the data layer |
| QA Engineer | Tests, code review (also the reviewer of each task) |
| DevOps | Builds, CI/CD, deployment files, infrastructure config |
| Tech Writer | Documentation, READMEs, comments |
| Data Engineer | Database schemas, migrations, queries |
You choose the team, from one agent to all eight. The default is five: Project Manager, Architect, Frontend Dev, Backend Dev and QA Engineer. The Project Manager can't be switched off: it's the one that turns your brief into a plan and coordinates everyone else. Reviews exist only if the team has a QA Engineer.
Workflow playbooks
A playbook tells the Project Manager how to plan and route the work. It shapes the plan; it doesn't add tools or approval steps. You pick one, or leave it on Auto and the office infers it from the words of your brief:
- Feature Delivery — plan, implement and verify a new capability
- Bug Fix — triage, patch the cause, regression test
- Refactor — map the structure, refactor, guard the behavior
- Performance Pass — measure, optimize, verify
- Security Review — audit, harden, verify
- Documentation — scope, write and review the docs
- Release — run checks, write notes, ship
- Auto — picks one from your brief and falls back to Feature Delivery
A real plan you approve
Before execution, the PM produces a sprint-aware plan: it rates the project from Trivial to Massive, then decomposes it into tasks — each with an assignee, dependencies, target files, acceptance criteria, a risk level, and an effort estimate. On the Plan Review screen you can edit a task's title and description, delete tasks, or send feedback to re-plan. Nothing runs until you press Approve & Start, with one exception: Quick Mode approves the plan by itself on Trivial and Simple projects. Complex and Massive projects also stop for your approval between sprints.
Watching the office work
Once running, four screens let you follow along (swipe between them or use the bottom bar):
- Activity — the team strip (one avatar per agent, with its state and an estimate of its tokens), the checklist of the current sprint and the live feed. When an agent can't go on without you, it stops and asks.
- Board — the tasks in seven foldable sections: Backlog, To Do, Needs input, In Progress, Review, Done and Failed. Cards carry the assignee, priority and risk; tap one to view its result, skip it, change its priority or retry it. It isn't drag and drop.
- Chat — the log of the office and your only channel to the orchestrator. Type a request and a model reads it as a re-plan, skip a task, change a priority, add a requirement, reassign a task, pause or resume. When it isn't sure, it does nothing.
- Results — tasks done and failed, progress, a live preview for web
projects, the documents the team produced and per-agent performance. Tokens are
always estimates (
~N tokens); there is no duration, no success rate and no amount of money.
Agents that actually do the work
This is the part that separates AI Office from a chat that only writes text. By
default the agents use the same tools as AI Chat: they read
and edit files, run terminal commands, use Git and more, limited to the project
folder and passing the same permission gates. The Office has no approval dialog
of its own: what needs your OK (a command, a git push, anything that leaves the
device) waits in AI Chat while the agent keeps working. With the default
Auto-edit access level, file edits run without asking and commands wait for
you. Tool commands show up in the feed as rows that start with a wrench.
It writes into a real project
Read this one twice: the Office writes your files live, as the agents work. There is no diff, no per-file accept or reject and no undo inside the Office, and pausing or cancelling doesn't roll anything back. Commit with Git, or copy the project, before you let it work on something you care about.
- With a linked project, it works inside that project's folder and edits its existing files.
- Without one, it creates a folder named like the project in the app's projects folder, and that folder becomes one more project in your list.
- Save writes a summary of the session (to
.pocketcode/office/in a linked project, to theREADME.mdotherwise) and the files from the task results. You don't need it to keep the files: they are already in the project.
If the app closes while an office runs, open it from My Offices: it comes back paused, with every finished task saved, and you press Resume.
Optional background analysis
After a session the Office can run seven analyses in the background, one after another. They are off by default: you turn them on in Settings → AI → AI Tool Permissions with Let the AI work on its own, and they run only while the phone is charging on Wi-Fi, capped per day. Security audit, test gaps, a documentation summary and a codebase map show up in Results, under Team insights. The optimization and refactor analyses run too, but their cards are never filled today.
What it costs — honestly
AI Office is a Pro module as a whole: on Free, opening it shows the paywall. The models behind it cost nothing extra from us, because every AI provider works with your own key or sign-in, and PocketCode sells no AI credits:
- OpenAI — an API key, or your ChatGPT account through the Codex CLI that ships with the app (it needs a paid ChatGPT plan). Gemini and Anthropic Claude — API key only. GitHub Copilot — your GitHub account (it needs a Copilot subscription).
- Any OpenAI-compatible endpoint via the Custom provider: OpenRouter, Groq, DeepSeek, LM Studio, Ollama or your own server. Plain HTTP works only for localhost; an Ollama or LM Studio on your network needs HTTPS or a tunnel.
So with a Pro plan (7-day free trial) you can hand a brief to a full AI dev team on the provider you like, and your provider bills you for the tokens. The regular AI chat works with every provider on the Free plan too; AI Office is what Pro adds. You can't set a token or time limit for a session: what bounds the spend is Pause and Cancel, a smaller team, Max retries per task, Max sprints and Quick Mode.
Working on something small? Quick Mode skips the clarifying questions and, on Trivial and Simple projects, approves the plan and skips the QA review, so you go from brief to build in one tap.
Autocomplete guesses your next line. AI Office plans the whole thing, splits the work across specialists, runs real terminal, git and file operations in your project, has QA review it, and hands you a project — from a phone.
PocketCode is heading to Google Play. Join the pre-registration to put an AI dev team in your pocket.
AI Assistant
Quick answers on this
Keep reading
- FeatureCortex: Stop Being the Index of Your Own AI ChatsA new PocketCode mode where you talk to a coordinator instead of an assistant: it splits your work across specialized chat sessions that stay alive for weeks, and hands back grouped decisions instead of scattered interruptions.14 min
- FeatureHow PocketCode's AI Works Under the HoodA technical look at how we designed PocketCode's AI assistant: inline suggestions, contextual chat, and tools that reach the app's modules, with your own provider.5 min
- TipsHow to Use AI to Generate Code FasterLearn techniques and best practices to maximize the integrated AI assistant in PocketCode.4 min

