Docs/AI Office

AI Office

Free (Gemini BYOK)Pro for other providers

Turns a natural-language brief into a work session where multiple specialized agents (PM, Architect, Frontend Dev, Backend Dev, QA, DevOps, Tech Writer, Data Engineer) plan, decompose, execute and review code collaboratively, following the workflow you pick: feature delivery, bug-fix, refactor, performance, security, docs, release or SPARC.

Your account, your model (BYOK / OAuth)

AI Office runs the agents using the AI account you connect: API key (Gemini, OpenAI, Anthropic, Copilot, GitHub Models) or OAuth login through the official Claude Code, Codex and Gemini CLIs. Pocket Code does not provide AI credits and does not proxy requests — each call goes directly from the device to the provider, who bills you under their usage policy.

4-step Lobby

Wizard to pick template, brief, team (toggle per role) and workflow + Quick Mode.

Project Research

Link an existing project and the office analyzes techStack, architecture, modules and patterns before planning.

Reviewable plan

Sprint-aware decomposition with dependencies, priority and acceptance criteria. Edit, add, remove or re-plan with feedback.

Observable execution

Office Floor (visual map), Task Board (Kanban), Chat Log (inter-agent feed) and Results (summary).

7 post-session workers

Audit, test gaps, documentation, memory consolidation, dependency map, optimization and refactor.

Persistent memory

Short-term, long-term and episodic learnings injected into future sessions on the same project.

Agent roles

Eight specialized roles, each with predefined skills. PROJECT_MANAGER is always forced in every session.

Project Manager

Planning, decomposition, prioritisation

Architect

Architecture, system-design, diagramming

Frontend Dev

Compose-ui, xml-layout, animation, theming

Backend Dev

API, repository, use-case, networking

QA Engineer

Testing, code-review, debugging

DevOps

Build, deploy, terminal, git

Tech Writer

Documentation, readme, comments

Data Engineer

Database, sql, schema, migration

Office Tools — agent capabilities

Each agent can invoke 10 tools during execution. Every call is persisted as a trace.

TERMINAL

FILE_READ

FILE_CREATE

FILE_EDIT

DATABASE_QUERY

DATABASE_SCHEMA

DATABASE_LIST_TABLES

HTTP_REQUEST

SSH_EXEC

GIT

Workflows — orchestration modes

Each workflow defines the phases, default skills and routing topology between agents. Pick it in Lobby Step 3 or let it be auto-inferred from brief keywords in Auto mode.

Auto

auto

The orchestrator infers the workflow from brief keywords ("sparc", "bug", "refactor", "performance", "security", "docs", "release"). Default: Feature.

Feature

feature

End-to-end delivery of a new feature.

Plan → Implement → Verify

review gates · max 2 parallel tasks

Bug-Fix

bug-fix

Reproduction, isolation and root-cause patch with regression check.

Triage → Patch → Regression

review gates

Refactor

refactor

Improve structure while preserving behavior. Regression tests required.

Map → Change → Guard

mandatory tests

Performance

performance

Measure before and after, optimize, verify concrete improvement.

Measure → Optimize → Verify

DevOps + QA + Architect + Backend

Security Review

security-review

Audit exploitable risk, unsafe defaults, data exposure. Harden and verify.

Audit → Harden → Verify

mandatory review

Documentation

docs

Create or update developer guides and docs. No blocking gates.

Scope → Write → Review

Tech Writer + PM + QA

Release

release

Run checks, prepare release notes and organize delivery artifacts.

Check → Notes → Ship

DevOps + QA + Tech Writer

SPARC

sparcStructured methodology

Five sequential phases with a mandatory review gate at Architecture.

Specification

Pseudocode

Architect

3gate

Architecture

Architect + Data Eng

Refinement

Backend + Frontend (parallel)

Completion

QA + DevOps + Tech Writer (parallel)

Lobby — 4-step wizard

Step 0 · Basics

Template (5 available), project name, optional link to existing project.

Step 1 · Description

Free textarea with the brief in natural language.

Step 2 · Team

Toggles for the 7 selectable roles (PROJECT_MANAGER is forced).

Step 3 · Advanced

Workflow (9 options), Quick Mode, max parallel tasks, max sprints, max retries, token budget, provider + model.

Planning & Plan Review

The PM generates clarifying questions (optional in Quick Mode), analyzes complexity (TRIVIAL → MASSIVE) and proposes a sprint-aware decomposition. The plan is fully editable before approval.

Clarifying questions

Questions with options the user answers before decomposition.

Complexity analysis

TRIVIAL / SIMPLE / MODERATE / COMPLEX / MASSIVE

Sprint-aware decomposition

Tasks with dependencies, priority, target files, risk and acceptance criteria.

Live edit

Edit / add / remove tasks or Re-plan with feedback.

Execution screens

Office Floor

Visual team map: grid showing each agent, status (IDLE / THINKING / WORKING / REVIEWING / DONE), avatar and current activity. Animated pulse when working.

Task Board

6-column Kanban (Backlog / Todo / In Progress / In Review / Done / Failed) with sprint filter and cards showing avatar, priority and risk level.

Chat Log

Unified feed of messages from agents, orchestrator, system and user, interleaved with tool execution traces. Input to send directives (replan, skip, change priority, reassign…).

Results

Summary with stats (tasks completed, tokens used, duration, success rate), agent roster and outcome per task. Button to save as project with README + files.

Project Research

When you link an existing project, the office analyzes it before planning: tech stack, architecture, modules, existing patterns, key dependencies and naming conventions. The result is injected into the PM's context.

7 post-session workers

When a session is archived, the scheduler chains seven workers automatically in background (Wi-Fi only). Their findings are persisted as memory entries for future sessions.

HIGH

Audit

Security audit — hardcoded secrets, injection, insecure API, path traversal, data exposure.

NORMAL

Test Gaps

Detects functions / classes without tests and complex logic without coverage.

NORMAL

Documentation

Generates README-style summary of output (goal, architecture, files, limitations).

LOW

Consolidate Memory

Deduplicates and compacts memory entries (keeps audit findings, test gaps and documentation intact).

—

Map

Dependency map of generated code (file → imports / exports).

—

Optimize

Performance optimization suggestions.

—

Refactor

Structural refactor suggestions.

Quick Mode

Toggle in Lobby Step 3 for fast iterations on simple projects. When active:

•Clarifying questions are skipped for TRIVIAL and SIMPLE
•Plan auto-approves (Plan Review is skipped)
•QA Engineer per-task review is disabled
•max parallel tasks bumps to 2 automatically
•PM produces a consolidated plan with fewer tasks

Not recommended for COMPLEX or MASSIVE complexity.

Project templates

Five templates auto-fill target platform, language, framework, code style and suggested roles:

android_app

Jetpack Compose · Clean Architecture · PM + Architect + Frontend + Backend + QA

rest_api

Ktor · Clean Architecture · PM + Architect + Backend + Data Engineer + QA

fullstack

Compose Multiplatform · Clean Architecture · 7 roles (all except DevOps + Tech Writer)

library

Kotlin Multiplatform · Clean Architecture · PM + Architect + Backend + QA + Tech Writer

game

Compose Canvas · MVVM · PM + Architect + Frontend + Backend + QA

Memory and persistence

Short-term

Temporary memory of the current session. Discarded on close.

Long-term

Learnings that persist across sessions of the same project.

Episodic

Relevant punctual events (incidents, decisions) with timestamp.

Up to 10 memory entries are injected into future task prompts. Sessions are persisted on-device; a checkpoint manager saves sprint progress to disk for crash recovery.

Free vs Pro plan

Module access, the 9 workflows, the 10 office tools, background workers and persistence are fully Free. The effective gate operates at the AI provider level: Gemini is available on Free; OpenAI, Claude, Copilot and other providers unlock on Pro. Pocket Code does not provide AI credits — you always connect your own account or API key with the provider.

Feature	Plan
Create and save AI Office sessions	Free
Lobby, team, workflow, Quick Mode	Free
Project Research on linked project	Free
Plan Review, task editing, re-plan	Free
Office Floor / Task Board / Chat Log / Results	Free
9 workflows (Auto + 7 templates + SPARC)	Free
7 post-session workers	Free
10 office tools (terminal, file, db, http, ssh, git)	Free
Run with Gemini	Free
Run with OpenAI / Claude / Copilot / GitHub Models	Pro
Mock Server, Performance Testing, Collection Runner	Pro
Cloud Sync, Live Collaboration, Export APK	Pro

Module statistics

Agent roles

Office tools

Workflows

UI screens

Post-session workers

Project templates

Complexity levels

Workflows