Build agents. Skip the infrastructure.
Hosted runtime, computer use, 150+ integrations, scheduling, and memory. Ship agents that work like people — not just APIs.
Stop building agent infrastructure
Everything an agent needs to run in production. None of the setup.
Hosted agent runtime
Agents run in isolated sandboxes. You ship the workflow, not the infra.
Computer use
Agents control a real desktop. Browse the web, fill forms, run scripts, use any GUI app. Works for anything — not just apps with an API.
Human-in-the-loop
Require approval before sensitive actions. Keep the speed without giving up control.
150+ managed integrations
Gmail, Slack, Notion, HubSpot, Google Ads, Linear. OAuth and token refresh handled.
Scheduled runs
Set the cadence once. Daily, weekly, or hourly runs happen automatically.
Email inbox
Teammates get a unique @m8tes.ai address to trigger runs from inbound email. Agents also have their own inbox to read mid-run — 2FA codes, verification links, confirmations.
Agent orchestration
Agents spin up other agents, inspect live runs, answer approvals, and set their own schedules mid-run.
Agent memory
Agents remember past conversations and build on them. Per-user scoping for multi-tenant apps.
File handling
Agents generate reports and spreadsheets, downloadable through the API.
From zero to deployed agent in minutes
Create a teammate, schedule it, give it an inbox. Everything runs in a hosted sandbox.
pip install m8tesFrom weeks of engineering to minutes
Other SDKs give you building blocks. We give you deployed agents.
With OpenAI Agents SDK, Claude Code SDK, etc.
- Build and maintain sandboxed execution
- Implement OAuth for every app you connect
- Write your own scheduling and trigger system
- Build approval flows so agents don't go rogue
- Design memory that persists across executions
- Set up real-time streaming for your UI
- Handle file output and delivery
- Build webhook infrastructure for agent events
- Isolate user data for multi-tenant apps
- Give agents their own email inbox
- Build desktop automation with screenshot feedback loops
- Write orchestration so agents can manage other agents
- Keep up with fast-moving AI models and APIs
Weeks of engineering before your first agent runs.
With the m8tes SDK
- Hosted sandbox, no servers to manage
- 150+ integrations with managed OAuth
- Scheduling, webhook triggers, and email inbox built in
- Human-in-the-loop so you stay in control
- Persistent memory across executions
- Real-time streaming out of the box
- Generated files, downloadable via the API
- Outbound webhooks for agent events
- Per-user isolation for multi-tenant apps
- Enable an @m8tes.ai inbox per teammate
- Computer use — agents control a full desktop, no driver setup
- Agents manage themselves and each other via built-in task tools
- Always up to date as AI evolves. Zero maintenance.
Live in minutes. Not weeks.
Integrate agents into any workflow
Run agents on demand, on a schedule, from an email, or from any external event.
Email inbox
Enable a teammate inbox, then email it to kick off work.
Schedule
Daily, weekly, or hourly. Set a cadence and your agent runs automatically.
Webhook
POST when a Stripe charge fails, a GitHub issue opens, or a form submits. Any event, any system.
On demand
Trigger runs from your code or the API.
Works with
Gmail
Slack
Notion
Linear
HubSpot
Stripe
Google Sheets
Google Adsand 150+ other apps
m8tes vs. LangChain, CrewAI, and other SDKs
LangChain, CrewAI, OpenAI Agents SDK, etc.
- Orchestrate LLM calls locally. No runtime included.
- You build and host the execution environment
- You implement OAuth for every integration
- You write scheduling, triggers, and memory from scratch
- You add human-in-the-loop and approval flows yourself
m8tes
- Hosted runtime. No infra to manage.
- 150+ integrations with managed OAuth
- Scheduling, webhooks, and email triggers built in
- Persistent memory out of the box
- Human-in-the-loop built in. Agents ask before acting.
- Computer use — agents operate any desktop app or website
Build for your users. Automate for your team.
Build AI features for your users
Customer-facing assistants
Give each user their own agent with isolated memory, tools, and permissions. Multi-tenant without custom plumbing.
Automated onboarding
New signup triggers a run that provisions accounts, sends welcome emails, and schedules a check-in.
In-app research agent
Users describe what they need. The agent searches, synthesizes, and returns structured results in your UI.
Automate internal workflows
Revenue reporting
Pull MRR from Stripe, update the sheet, and post the summary to Slack. No more manual Monday reporting.
Support triage
Classify inbound tickets, draft replies, escalate blockers to the right Slack channel. Runs 24/7 on a schedule.
Ad spend monitoring
Check Google Ads weekly. Pause ads burning budget on low-converting keywords. Alert the team.
Go deeper on AI agent development
Architecture, patterns, and production setup.
Computer use and desktop automation
When sandbox execution is enabled, every run gets a full Linux desktop alongside your normal tools. The agent opens a browser, navigates to any URL, fills forms, runs commands, and uses any GUI app. After every action it sees a screenshot and plans the next step.
Computer use tools and MCP integrations run side-by-side. An agent can pull data from an API, open a browser to check a dashboard, take a screenshot, and post it to Slack — in a single run.
Teammate, task, run
The SDK is built around three concepts. A teammate is a reusable agent with a name, instructions, and connected apps. A task defines what the agent should do. A run is a single execution of that task.
You configure a teammate once. Then trigger runs on demand, on a schedule, or from a webhook. Each run streams events in real time. Show progress in your UI or pipe results downstream.
Hosted sandboxed execution
Every run executes in an isolated sandbox with its own filesystem and network. The runtime handles tool calls, OAuth token refresh, rate limiting, and retries. You never touch infrastructure.
Streaming and real-time events
The SDK uses server-sent events (SSE) to stream run progress. You get events for tool calls, text output, file creation, and completion. Build real-time UIs or process events in a backend pipeline.
Set user_id on any run to scope it to a specific end user. Memory, task history, and tool access are strictly isolated. One user never sees another user's data.
User isolation with user_id
Set user_id on any run to scope it to a specific end user. Memory, task history, and tool access are strictly isolated. One user never sees another user's data.
Per-user memory and tool scoping
Each end user gets their own memory context. The agent remembers past interactions with that user and builds on them. Scope tool permissions per user for fine-grained access control.
Embedding AI agents in your product
Every user gets their own agent experience inside your app. Isolated runs, memory, and tool access. You control the instructions and capabilities. Your users just see results.
Three modes. Autonomous runs without asking. Approval asks before each sensitive action. Plan proposes first; you approve, then it executes. Start with approval. Move to autonomous when you're ready.
Permission modes for safe execution
Three modes. Autonomous runs without asking. Approval asks before each sensitive action. Plan proposes first; you approve, then it executes. Start with approval. Move to autonomous when you're ready.
You can change the mode after a run starts. Use client.runs.update_permission_mode(run_id, permission_mode=PermissionMode.APPROVAL) when you need to tighten control mid-run. Switching back to autonomous resumes pending tool approvals on paused runs, but questions still wait for an answer.
Scheduling, webhooks, and email triggers
Cron schedules handle recurring work. Webhooks let you trigger runs from any external event (Stripe charge failed, GitHub issue opened, form submitted). Email triggers let users forward messages directly to an agent.
Monitoring and output validation
Every run produces structured logs showing each tool call, its inputs, and its outputs. Use them to debug, audit, and improve. Output files (CSV, PDF, reports) are downloadable through the API.
Track account-level consumption too. client.auth.get_usage() gives you runs, spend, limits, and billing period boundaries so you can surface usage in your own dashboards or alerts.
FAQ
Common questions about building AI agents with the SDK.
Explore more
Build your first AI agent today
Free to start. No infrastructure to manage. Deploy in minutes.