T
MetaAgentAgent Engine
MetaAgent · Agent Engine

Your AI workforce,
orchestrated.

Build, schedule, and run autonomous agents that do real work — read your mail, talk to customers, log leads to your sheets, ship reports. One control plane. Every model. Every tool.

Sign in →Read the API docs
27+
Built-in tools
MCP
Native protocol
Claude
Sonnet · Haiku · Opus
Self-host
Your infra, your data

What you get

Everything you need to run agents in production, not just play with them in a notebook.

Orchestrator master agent

One agent at the front door. It picks the right specialist for every request, runs it, and reports back.

🛠

Tools that actually work

27+ built-in tools (Sheets, web search, mail, file I/O, HTTP, OpenAPI, …) plus any MCP server from Smithery's registry.

📬

Email triggers — in & out

Point a mailbox at an agent. Inbound mail fires it automatically (HMAC-signed webhook), and `mail.send` replies through MetaMail with the right From/Reply-To.

Cron, webhook, event triggers

Schedule an agent for 7am every weekday, fire it on each new lead via webhook, or subscribe to an engine event — all from one /triggers surface.

🧠

Conversational memory

Agents remember the full thread of tool calls, results, and human answers across resumes — no goldfish loops.

🔌

Native MCP

Both directions: install MCP servers as tool sources, AND expose this engine as an MCP server to other clients.

🛡

Human-in-the-loop

Agents ask the operator via /inbox when they need information only a human has — and resume cleanly once answered.

🕸

Visual workflow engine

Drag-and-drop multi-step flows on a canvas. Branch on conditions, chain agents, fan out parallel work, persist state at every step.

📜

Audited + observable

Every run is a row. Every step is a row. Every tool call costs USD. Full OTEL traces; full Postgres replay.

Workflow engine

When one agent isn't enough —
compose them.

Some work needs more than a single LLM turn: lead enrichment pipelines, daily-digest assembly, multi-step QA flows, ETL over a CSV. The Workflow Engine lets you wire agents, built-in tools, and human-approval steps together on a drag-and-drop canvas — and runs the result with full state persistence so a 15-step flow survives a worker restart.

  • Drag-and-drop canvas — React Flow, with a palette of agent, tool, condition, human-approval, and webhook step types.
  • Branch on outputs — conditional edges fire based on the previous step's result. Loop while a predicate holds.
  • Cron + webhook triggers can fire whole workflows, not just single agents (TAR-115).
  • State persisted at every step — restart a worker, and the run picks up exactly where it stopped.
  • Replay any run with one click — same input, optionally a different agent or model. Every old run stays untouched.
Build one →Workflow API

How it works

  1. 01

    Connect your models + tools

    Anthropic Claude is pre-seeded. Add OpenAI, Ollama, your Smithery MCPs, or your own Python skills.

  2. 02

    Talk to the Orchestrator

    Describe the work in chat. It builds a specialist agent for you, promotes it live, and runs it on the spot.

  3. 03

    Schedule + automate

    Cron triggers, webhook receivers, multi-step workflows — same primitives, same audit trail.

  4. 04

    Watch + steer

    Inbox, runs, queue, costs — every surface in the console. Cancel or force-complete any agent that wanders.

Ready to dispatch your first agent?

Sign in and ask the Orchestrator to set one up for you.

Sign in →