MarthaBuild agents and workflows, ship them everywhere.

A platform for production-grade AI agents — durable workflow orchestration, plug-and-play tools, embeddable delivery, and a CLI that's first-class for both humans and other LLMs.

Quickstart →

Concepts

GitHub

Multi-step workflows that actually finish

Build flows out of LLM calls, function calls, branches, loops, parallel paths, and human approvals. Each step retries independently. Long-running pauses don't lose state.

How workflows work

Any REST API becomes a tool

Drop in an OpenAPI spec and every endpoint shows up as an agent-callable function. Add domain context once; agents do the rest. Connections store credentials without baking them in.

Import an OpenAPI spec

Documents in, retrieval out

Upload PDFs, markdown, HTML, images. Martha parses, chunks, embeds, and exposes hybrid keyword + semantic search. Agents cite sources back automatically. R2 folders auto-ingest.

Document pipeline

Embeddable chat, two lines of code

Drop the chat widget into any web app with React, Vue, or vanilla JS. Auth handled via short-lived tokens you mint server-side. Origin allowlist enforced.

Embed chat

Reach users on their channel

Browser chat, SMS, WhatsApp, voice — same agent, multiple front doors. Send notifications via email, Slack, or webhook from any workflow node.

Notification channels

Pause for humans without breaking flow

Workflows can wait days for an approval, a payment confirmation, or any signed webhook event. State persists; no polling. Agents can request approvals too.

Human-in-the-loop

Memory that scales with sessions

Agents recall past chat messages, tool outputs, and persistent facts across sessions. Hybrid retrieval (semantic + keyword + recency) with tenant isolation.

Agent memory

Async work via task queues

Queue work for agents to claim and complete. Built-in lifecycle (open → claimed → running → done) with heartbeats. Bidirectional sync with Linear, GitHub, and GitLab.

CLI tasks reference

A CLI built for humans and AI alike

Every command supports JSON output and predictable exit codes. Ships a bundled skill reference so agent runtimes (Claude Code, CrewAI, Pydantic AI, ork) integrate in minutes.

Agent integration

Why Martha

The orchestration, integration, and delivery layers around an LLM are 80% of what makes an agent production-ready. Martha builds those once, well — so the part you actually own is the prompts, the functions that touch your systems, and the workflow shapes for your specific use cases.

Trade-off:

You don't build	You do build
Conversation engine, streaming, tool calls, citations	Prompts
Workflow runtime, retries, branches, pauses	The shape of your flows
Document parse + chunk + embed + search	Which docs to upload
Embed widget, SMS/WhatsApp adapters, voice	Front-door rollout strategy
Credential store, per-client allowlists	Which integrations to grant
Task queue, heartbeat semantics	What work to dispatch

Concepts walks the primitives. Quickstart gets a CLI configured and a first command running in under five minutes.