Band 1
Free Conversation
Greetings, smalltalk, generic questions get a natural answer. Off-topic is welcomed; MH AI politely notes that its primary expertise is Macro Handler.
AI Assistant
MH AI — Macro Handler's built-in assistant
MH AI is a smart assistant that can chat naturally, knows the app, and designs macros for you through an approval gate. Pick from cloud providers such as Gemini, OpenAI, Claude, Groq, DeepSeek, Qwen, and more as the brain (BYO API key); on-device local models are on the roadmap.
Macro assistant on the website
The floating MH AI robot on public pages lets you use your own Gemini, OpenAI, Claude, OpenRouter or another supported provider API key to learn from guides and draft Lua. Model access, charges and quotas belong to your provider account; a chat subscription does not mean API access. Some providers block direct browser connections. In that case, copy the prompt with references into your own assistant. The web assistant does not control your device; review and test the draft in the app.
Paste your API key first. Provider-specific formats are detected in your browser; for shared key formats, you select the provider that issued the key. After you approve the connection, its model catalog loads and you can choose a model. If the provider does not offer a catalog, use an exact model ID available to your account. Detection never sends trial requests to other services.
Connections and the model team are managed in a separate view inside the panel. Add multiple keys. In the composer, tap the options icon to the left of the send actions. Choose the connection, model, workflow mode and supported reasoning effort in the small window just above the button. Your messages appear on the right and MH AI replies on the left; code actions and expandable references stay inside each reply. The fullscreen icon in the header expands the panel; the same button restores the compact view without losing your conversation. When the panel is closed, drag the robot head with its animated eyes to move it, or click to open the assistant. Opening motion respects reduced-motion settings and the site’s pause-animations control.
Simple mode uses one request for a quick response. Advanced, Developer and Ultra modes add bounded drafting, planning and review steps. When you consent to teamwork, selected models use the same version-specific official API reference and can review outputs from earlier stages. Check the request count and your own provider account terms before sending. Reasoning effort is adjustable only when the model declares support; other connections use the provider default.
Choose Automatic failover, Sequential collaboration or Collective collaboration under Model execution. Set your team and priority order in Connections. Automatic mode switches to your next approved connection on quota or service failure. Sequential mode passes drafts to the next model for review; collective mode prepares independent drafts in parallel, then synthesizes them. Replies identify actual contributors. Sharing and billing consent is required; cancelled or incomplete replies are not replayed through another account. The app prepared for Android 1.0.51 offers the same modes in MH AI connections; device actions remain on one approved execution path.
Generated Lua drafts are checked for API names, call forms and syntax. These checks do not guarantee success on a real device. Verify coordinates, images and stop conditions, then test the macro safely in the app. The web assistant does not act on your device; app and device actions on Android retain their existing approval gates.
AI connections added while signed in are stored encrypted in this browser. Refreshing or signing out and back into the same account does not require entering your keys again. Removing a connection also deletes its saved key. Guest keys and chat are not persisted; connections are not synced to other devices.
MH AI works in three bands
Depending on your message, MH AI picks a different prompt family. General chat is light; an app question is grounded; an automation request is structured and approval-gated.
Band 2
App-Grounded Question
Questions like "how to use Region" or "OCR in no-code" are answered from the app's own API and block reference; a proposal that names an unknown API is rejected.
Band 3
Automation Proposal
"Build me a macro" requests produce a structured JSON proposal. Validator → risk → preview → YOUR approval → apply → rollback if needed.
- Plans can branch with typed, nested IF/ELSE: up to 32 steps and four condition levels, with a preview of each branch before you approve.
- Counted repeats: proposals that repeat a group of steps a set number of times (1–1000). Newv1.0.51
Agent mode: step by step on your device
Switch the panel chip from Chat to Agent and describe a task. MH AI breaks it into device actions and asks you before each one.
What the agent can do
- Open an app
- Read the screen
- Tap
- Swipe
- Type text
- Back
- Home
- Each action opens an approval card that names the action and the on-screen target the agent intends; choose Approve or Deny. Unanswered requests expire.
- The Google Play build has no auto-approval (Full Access): every step waits for you.
- Agent mode needs a brain that supports tool calls; with other brains MH AI stays in chat. STOP ends the run at any time.
Use it by voice
Speak instead of typing with the microphone button or the /voice command. While Agent mode is on, MH AI also reads its replies aloud with your device's text-to-speech voice.
Speech recognition and text-to-speech use your Android device's system services.
Newv1.0.51 The My Agent Routine block is a separate feature from MH-AI agent mode: a routine does not start while MH-AI autopilot is running a task, and MH-AI Stop also stops a running Agent. My Agent Routine guide
Slash commands
Type / in the panel and the list filters as you type. Command names are the same in every language. Commands marked "Approval" go through an approval card.
Input from the screen
/region- Select a target region on screen.
/point- Select a target point on screen.
/screenimage- Capture an image template from the screen.
/screencolor- Sample a color from the screen.
/image- Attach an image from the device.
/voice- Speak your request.
Build and review
/macroApproval- Generate a new macro.
/codeApproval- Write a script into the code editor.
/nocodeApproval- Add blocks to the no-code workspace.
/audit- Audit the current macro.
Watch, take over, and device
/watch- Watch me and learn.
/takeoverApproval- Take over this task and continue.
/ocr- Read text from the screen with OCR.
/androidApproval- Run an Android action.
Libraries
/gallery- Open the gallery.
/template- Open the template browser.
/color- Open the color palette.
/replay- List saved replays.
Settings and safety
/settings- Open MH AI settings.
/model- Choose the AI brain and model.
/language- Change the reply language.
/help- Show the available commands.
/stop- Emergency stop.
You choose which brain to use
In the current release path, you can choose from 15 cloud providers with your own API key. The on-device local model catalog is on the roadmap; once release-ready, it will use the same capability registry, safety rules, and approval flow.
Google Gemini
Cloud — high capacity, multimodal. API key is BYO (bring your own).
OpenAI GPT
Cloud — structured macro proposal via function-calling. BYO key.
Anthropic Claude
Cloud — stable across long contexts, safety-aware. BYO key.
Groq
Cloud — low-latency high-speed inference. BYO key.
DeepSeek
Cloud — code-oriented models. BYO key.
Qwen
Cloud — Alibaba Qwen models (OpenAI-compatible endpoint). BYO key.
xAI Grok
Cloud — xAI Grok models. BYO key.
OpenRouter
Cloud — multi-provider routing behind one key. BYO key.
Cerebras
Cloud — high-speed inference. BYO key.
Fireworks
Cloud — open-weight model hosting. BYO key.
Perplexity
Cloud — search-grounded answers. BYO key.
Mistral
Cloud — Mistral models. BYO key.
Together
Cloud — Together AI model hosting. BYO key.
Cohere
Cloud — Cohere Command models. BYO key.
NVIDIA NIM
Cloud — reviewed NVIDIA NIM models. BYO key; NVIDIA trial quotas and service terms apply.
Gemma / Phi / LlamaOn the roadmap
On-device local models are not available yet. They are planned to run on the device without an internet connection; until then MH AI works with the cloud providers above.
What you can ask
MH AI auto-classifies your message; each class routes to the right pipeline.
General Chat
Greetings, smalltalk, off-topic — natural answer; Macro Handler primary expertise is politely noted.
Macro Handler Question
Questions about the API, blocks, the no-code workspace and the code editor. Answers are grounded in the app's own API and block reference.
Macro Planning
Natural-language macro brief → structured JSON proposal. Only registered capability IDs are used.
Code / No-Code Generation
Lua script for the code editor or block chain for the no-code workspace. Everything flows through approval.
Audit / Fix
Inspects your active macro, surfaces issues, proposes fixes — you approve before anything applies.
Observe-Learn / Live Takeover
AI learns while you act or assists live. Still routed through the approval gate.
Safety rules
Generative AI can return incorrect or incomplete output. MH-AI never treats provider output as authority: capability validation, risk rules, preview, and user approval reduce the chance that an incorrect or dangerous proposal is applied. Critical safety boundaries are protected by regression tests.
Per-action approval in the Play release
In the Google Play release build, MH-AI cannot auto-approve device tools. Every write to the code editor, the no-code workspace or the Logic workspace, and every Android device action, waits for your approval.
Rejects hallucinated capabilities
If a proposal references an unknown capability id, the safety gate rejects it; after a single retry, a clarifying question is shown.
Dangerous operations blocked
Denylisted actions (payment, account deletion, root, wipe) cannot be proposed by any provider.
"I applied/saved" claim blocked
If the provider claims it already applied or ran the macro, that approval-bypass false-claim is rejected.
Low confidence → clarifying question
When the model reports low confidence, MH AI asks what you actually want instead of auto-applying.
Approval flow
When AI produces a proposal, this is the path it takes — nothing moves without you.
- 1Message
- 2Intent classifier
- 3Context + capabilities
- 4Plan / proposal (JSON)
- 5Validator + Safety gate
- 6Preview
- 7Your approval
- 8Apply + Rollback
Where you use it
Open MH AI with the MH AI button on a macro in the app’s home list, or with the MH AI button on the floating control panel; the assistant opens for that macro.
In the no-code workspace
Open the MH AI panel and type something like "find Y in app X, tap it, wait 1 second, repeat". AI proposes the right block chain; you approve, and it lands in the no-code workspace.
Go to no-code →In the code editor
Get a Lua draft from a sentence like "find a template in Region, tap if found, wait 500 ms". AI writes code with the documented Macro Handler Lua API, and the sandbox boundary stays intact.
Go to code editor →In the Logic workspace
Ask for a change while the Logic workspace is open; the proposal you approve is applied to the Logic project captured at that moment. If the project changed in the meantime, the stale approval is rejected instead of overwriting your newer edits.
Go to the Logic guide →Install Macro Handler — MH AI is with you
After installing, the MH AI panel is one tap away on every screen. Add your key, verify provider and model settings, then prepare and review your macro draft.
