iOS app
PocketMai
Native conversations, voice, documents, OCR, browser handoff, and background replies.
Download on the App Store ↗PocketMai · open source · powered by MaiCore
↗
PocketMai, from phone to shell.
A native iOS app and an agent CLI for macOS, Linux, and Android. Choose the model, decide which tools it can use, steer every subagent, and keep your conversations on your terms.
Open source on GitHub ↗Two interfaces
The same MaiCore engine in the app and the terminal.
macOS · Linux · Android
pmai CLI
Terminal-native chat, tools, memory, skills, protocols, and full agent supervision.
curl -fsSL https://raw.githubusercontent.com/trufae/pocketmai/main/www/install.sh | sh
One shared coreiOS appmacOS CLILinux CLIAndroid CLI
Why PocketMai
Agents you can inspect, steer, and stop.
See the work, change direction, and set clear boundaries.
Full subagent control
See and steer every subagent.
Follow the whole agent tree and each child’s output, transcript, status, and usage. Send a message, focus one agent, pause and continue it, or stop a branch.
- Per-agent model, prompt, and tools
- Fair queues for busy slots
- Depth, count, turn, tool, token, and time limits
Editable context
Change the conversation when you need to.
Edit, undo, trim, archive, search, or compact with a focus. Linked tool calls stay valid.
Live steering
Add instructions while it is working.
New guidance joins at the next safe boundary without throwing away progress.
Clear permissions
Decide what each tool may do.
Allow once, always allow, edit arguments, deny, or cancel. Each agent sees only what you enabled.
Your stack
Use the models you already trust.
Apple Intelligence, OpenAI-compatible providers, local models, skills, plugins, MCP, and ACP.
Local-first
Local models and offline conversations.
Keep the model, documents, memory, and useful tools close to you.
Local conversations are first-class.
With an on-device or local provider selected, normal conversations, project memory, attachments, OCR, and local tools can stay off the internet.
No cloud accountNo remote transcriptNo connection
Run the model on your hardware.
Use Apple Intelligence on device, local MLX models on supported iPhones, or connect to Ollama, llama.cpp, and other local endpoints.
Spend context on the task, not tool schemas.
Native calls when available; JSON, XML, and text fallbacks when they are not. The hybrid tool proxy keeps the common tools direct and the long tail compact.
Read documents without uploading them.
Read images and scanned documents with Apple Vision, or use a local Tesseract installation from the CLI—without uploading the page.
Your local workspace remembers.
Chats, system prompts, project memory, todos, and context controls remain available locally between sessions.