AGI, Inc.
Also known as: MultiOn, MultiOn AI
AGI, Inc., formerly MultiOn, makes AGI-0 and the Agentic Phone MCP, an agent that operates a real Android phone's apps for any MCP client.
AGI, Inc. is the company formerly known as MultiOn, and multion.ai now redirects to theagi.company. MultiOn began as a web agent that carried out tasks in a browser. The company's current products operate phones instead: AGI-0, an action agent the company lists as available on Android through Google Play, and the Agentic Phone MCP, which lets any MCP client, such as Claude Code, Claude Desktop, Cursor or ChatGPT, run natural language tasks on a real Android phone.
The AGI Android app is the agent. It works through Android's accessibility APIs, opening apps, navigating and typing in their real interface while the owner watches the screen. A cloud MCP server at api.agi.tech queues tasks, dispatches them to the phone and relays progress back. Tasks start and return a task ID at once, and the client waits on a monitor tool that returns when the task finishes, pauses or needs the user. Sessions group related tasks on one device, and each device runs one task at a time.
Oversight is built into the task model. Logins, confirmations and CAPTCHAs are always handed back to the user instead of being handled by the agent, a device can be paused to refuse new work, and a running task can be steered or canceled. Each task records its steps with screenshots and the agent's reasoning, which a developer can fetch as the task runs. Guides also cover app testing on real hardware, messaging with a confirmation step, and daily errands across apps.
AGI, Inc. publishes no price and does not name the model behind the agent. It names no roles or single sign on, memory store, scheduled or event triggers, or integrations beyond operating apps on screen. Its trust center runs on Vanta at trust.agi.tech. Documentation for the MultiOn V1 Beta web agent remains online at docs.multion.ai, but the company no longer markets that product.
Vendor details
Canonical URL
https://www.theagi.company
Category
Browser / computer-use agent
Company status
independent
Use cases & customers
Target customers
Deployment options
Integrations
AGI is reached from any MCP client (Claude Code, Claude Desktop, Cursor, ChatGPT or a custom agent) over the AGI MCP server with OAuth 2.1 or an API key. The agent operates installed Android apps through their interface instead of through connectors.
In practice
You want your AI assistant to act on your phone, not just talk about it. Connect Claude Code, Cursor or ChatGPT to the AGI MCP server and ask it to send a WhatsApp message or check your calendar. The AGI app does it on the phone while you watch.
You test a mobile app on real hardware. The agent walks onboarding and other flows through the real interface and returns a screen by screen report with screenshots, asking before it grants permissions.
You want errands done across several phone apps without handing over your logins. Logins, confirmations and CAPTCHAs come back to you, and you can pause the device at any time.
Sources & related URLs
Related / legacy domains
Agentic Index coverage score
6.0 / 14 capabilities · 43%
| Integrations & Tool Calling | Not documented |
|---|---|
|
The agent reaches other apps only by operating their interface on the phone. AGI names no connector catalog, custom tool support or API action framework for outside systems. Sourcedocs.agi.tech/llms.txtread 2026-09-27 |
|
| Workflow Orchestration | Partial |
|
A task is one natural language instruction the phone agent carries out over several steps. A session groups sequential tasks on one device, with pause, resume and cancel. The calling AI client handles sequencing across tasks, and AGI names no branching, retry or deterministic workflow model. Sourcedocs.agi.tech/guides/concepts/task-lifecycleread 2026-09-27 |
|
| Knowledge Grounding & RAG | Not documented |
|
The agent reads what is on the phone's screen as it works. AGI names no retrieval over customer documents or a knowledge base. Sourcedocs.agi.tech/llms.txtread 2026-09-27 |
|
| Human Oversight & Guardrails | Full |
|
Logins, confirmations and CAPTCHAs are always handed back to the user as a needs_user_control handoff, and the agent never handles them alone. A device can be paused to reject new tasks, and a running task can be steered or canceled. Sourcedocs.agi.tech/guides/best-practices/securityread 2026-09-27 |
|
| Security, Identity & Governance | Not documented |
|
An OAuth 2.1 grant or an API key gives full control of an account's registered devices. AGI names no access model, roles, SSO or SCIM. The theagi.company Trust Center at trust.agi.tech runs on Vanta. Sourcetrust.agi.techread 2026-09-27 |
|
| Observability & Auditability | Full |
|
Every phone task records its steps, with screenshots and the agent's reasoning. A developer can fetch them incrementally through phone_task_monitor and phone_task_status, beside the task's state, terminal outcome and device state. AGI does not say how long the records are kept or whether they export to a SIEM. Sourcedocs.agi.tech/guides/best-practices/monitoringread 2026-09-27 |
|
| Memory & State Persistence | Partial |
|
A session keeps related tasks on one device so a workflow stays together. AGI-0 learns the user's preferences and improves over time, by AGI's account, but AGI names no memory store and does not say what the agent keeps or for how long. Sourcedocs.agi.tech/guides/concepts/devices-and-sessionsread 2026-09-27 |
|
| Deployment & Data Residency | Partial |
|
Tasks run on the customer's own Android phone in the AGI app, and AGI's cloud MCP server queues and relays them. AGI-0 operates on the device and also works offline. AGI names no region choice, self hosted server or private deployment. Sourcedocs.agi.tech/guides/concepts/architectureread 2026-09-27 |
|
| Prebuilt Agents, Templates & Packs | Not documented |
|
The use case guides cover messaging, a daily assistant, cross app workflows and app testing, with prompt patterns the customer writes into each task. AGI names no packaged agents or templates. Sourcedocs.agi.tech/guides/indexread 2026-09-27 |
|
| Triggers & Channel Coverage | Partial |
|
The agent starts work when an MCP client such as Claude Code, Cursor or ChatGPT calls phone_run_task, or when the user asks the AGI-0 app. AGI names no schedule, webhook or event trigger, so the agent acts only when asked. Sourcedocs.agi.tech/android-mcpread 2026-09-27 |
|
| Model Flexibility & Routing | Not documented |
|
The model behind the phone agent is not named, and AGI offers no choice of model. The customer picks which AI client calls the MCP server, and that choice does not change the agent's own model. Sourcedocs.agi.tech/guides/concepts/architectureread 2026-09-27 |
|
| APIs, SDKs & MCP Extensibility | Full |
|
AGI's own MCP server at https://api.agi.tech/v1/mcp runs over Streamable HTTP or SSE and takes an OAuth 2.1 sign in or a bearer API key. Its published tools create and control AGI's own devices, sessions and tasks, among them phone_run_task, phone_session_start, phone_task_pause and phone_device_pause. AGI presents this server as the product's interface. Sourcedocs.agi.tech/android-mcpread 2026-09-27 |
|
| Testing, Debugging & Optimization | Not documented |
|
Optional run_id, run_type and task_metadata tags label benchmark or regression runs for later analysis, without scoring them. AGI claims a ranking on the AndroidWorld benchmark. It names no evaluation harness, scored test cases or quality gate for the agent. Sourcedocs.agi.tech/guides/use-cases/app-testingread 2026-09-27 |
|
| Browser & Computer Use | Full |
|
The AGI Android app operates a real Android phone through Android's accessibility APIs. It opens apps, navigates and types in their real interface while the user watches, driven by natural language tasks from any MCP client. AGI-0, by AGI's account, acts across apps on phones and computers. Sourcedocs.agi.tech/guides/concepts/architectureread 2026-09-27 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Product updates
AGI Inc. announced a Qualcomm collaboration to bring agentic AI to Snapdragon-powered devices.
Bears on: Deployment / data residency
View sourceAGI Inc. publicly claimed AGI-0 reached 97.4% task success on the AndroidWorld benchmark.
Bears on: Browser / computer-use agent
View sourceAGI Inc. announced a Visa partnership for agentic commerce and secure agent-led transactions.
View sourcePricing
No public pricing published
usage
Included quota
Not published. The Agentic Phone MCP documents per task timeouts (300 seconds by default, 900 at most) and one task per device at a time; no quota or plan allowance is stated.
What is public
MultiOn is now AGI, Inc. (multion.ai redirects to theagi.company). The AGI-0 app is listed as available now on Android through Google Play, and the Agentic Phone MCP connects any MCP client to a registered phone by OAuth sign in or an API key created on platform.agi.tech. No price, plan or usage rate is published on theagi.company or docs.agi.tech; platform.agi.tech/pricing renders client side and was not readable.
Billing mechanics
Not published on the pages read. Access is by an AGI account (OAuth) or an API key from platform.agi.tech; the platform's pricing page renders client side and could not be read.
Cost watchouts
No published pricing; the product runs on the customer's own Android phone, so device and connectivity are the customer's cost, and API key or OAuth access gives full control of registered devices.
Variable cost rationale
Indeterminate from public info - the legacy API was usage/web-automation-call-based, but AGI-0's on-device model ('no cloud round-trips') and the enterprise SDK suggest device/app-embedded distribution; without published terms, cost exposure can't be characterized reliably
Additional watchouts
No published pricing to procure against; the MultiOn web agent API is no longer marketed (multion.ai redirects to theagi.company) although its V1 Beta docs still load, so confirm which product and terms apply before building.
Overage / add-ons
Not published for AGI-0 or the Agentic Phone MCP.
Sales call required
Mixed (some tiers require a call)
Free / trial
No free tier or trial published
Commercial notes
AGI, Inc. (formerly MultiOn) describes itself as an AI lab bringing intelligence to the edge. Current products: AGI-0, an action agent for phones and computers; the Agentic Phone MCP, which lets any MCP client run tasks on a real Android phone; and a platform for embedding agents in devices. The company cites a first place AndroidWorld ranking for AGI-0, its own claim.
Key ambiguities
Whether the platform.agi.tech pricing page publishes a rate, and whether the legacy MultiOn Agent API (docs.multion.ai, V1 Beta) is still sold, are not established from the pages read.
Cancellation / refund
Not published on theagi.company or docs.agi.tech.
Support SLA / resale
Documentation at docs.agi.tech with troubleshooting guides; a book a call path on theagi.company; no support tiers or SLA published.
Missing data
Every price. No plan, unit rate, free allowance or overage rule is published on theagi.company or docs.agi.tech, and platform.agi.tech/pricing could not be read. Earlier figures for the MultiOn Agent API came from a legacy product that the company no longer markets and are not carried here.
Related vendors
- AgentQL — A self healing query language for the web, with natural language…
- Airtop — Cloud browsers for AI agents that compile plain English workflows…
- Asteroid — Healthcare portal integration platform whose supervised browser and…
- Automat AI — AI agents that operate computers visually, sold as Automat Workforce…
- Autotab — General AI agent that learns a workflow from a demonstration and…
- Axiom.ai — No code browser automation tool and Chrome extension for building…
Alternatives to AGI, Inc.
The closest documented capability profiles to AGI, Inc. among browser and computer-use agents tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Autotab8.0 / 14Adds documented Integrations & Tool Calling and Testing, Debugging & Optimization
- Pine AI7.5 / 14Adds documented Integrations & Tool Calling and Knowledge Grounding & RAG, among othersAGI, Inc. vs Pine AI →
- TinyFish7.5 / 14Adds documented Integrations & Tool Calling and Security, Identity & Governance
- Brave Leo AI Browsing6.0 / 14Adds documented Knowledge Grounding & RAG and Prebuilt Agents, Templates & Packs, among others
- Browserbase10.0 / 14Adds documented Integrations & Tool Calling and Security, Identity & Governance, among others
- Fellou8.0 / 14Adds documented Integrations & Tool Calling and Knowledge Grounding & RAG, among others
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded