Martha

Martha is an AI-powered workflow orchestration platform built on FastAPI, PostgreSQL, and Temporal. It connects LLM-driven chat sessions with declarative workflow definitions, function calling, and human-in-the-loop capabilities.

New here? Start with the Quickstart — install the CLI, configure a profile, run your first command in under five minutes.

Building an agent? Read Agent integration for the JSON-contract discipline, exit-code semantics, and skill bundle.

Key Capabilities

• Multi-provider LLM integration — Claude, OpenAI, and any LiteLLM-compatible model with structured output support
• Declarative function definitions — Define API functions in the database, auto-sync from OpenAPI specs
• Workflow orchestration — Multi-step workflows with branching, iteration, retries, and conditions via Temporal
• Human-in-the-loop — Pause workflows to wait for external events (approvals, payments, reviews)
• Document intelligence — Upload, parse, chunk, and embed documents with hybrid search (keyword + semantic) for agents
• Multi-tenant — API key-based authentication with per-client function access control

Guides

• CLI — Command-line interface for managing definitions and running workflows
• Citations — Formal source references from agents and workflows to document chunks
• Document Ingestion — Automatic parsing, chunking, and embedding of uploaded documents
• R2 Folder Sync — Automatic ingestion from Cloudflare R2 bucket uploads
• Document Tools — Give agents access to list, read, and search documents
• Embeddable Chat — Install Martha chat in third-party apps using the hosted widget bundle
• Agent Memory — Recall past chat messages and tool outputs from earlier in the session
• Iteration Nodes — For-each, filter, and reduce nodes for array processing in workflows
• Structured Outputs — Request JSON Schema-constrained responses from any LLM provider
• Human-in-the-Loop Workflows — Pause workflows for approvals, reviews, and external events
• Event-Driven Triggers — Automatically start workflows when documents are uploaded, ingested, or events occur

Reference

• Embed API — Manifest, token, upload, and origin policy for embedded Martha chat
• Webhook API — Send events to waiting workflows

Architecture

• Overview — System components and design decisions