Loom Agent

What it is

Loom is a single FastAPI backend that turns any of 400+ LLMs (via OpenRouter) into a tool-using agent: one that searches the web, scrapes pages, and runs calculations autonomously in a streaming, multi-step loop. On top of that agent sits a visual workflow builder: a drag-and-drop canvas where tool calls become nodes in a DAG, so multi-step automations can be built without writing code.

The same backend powers three different clients, proving out one API across three very different runtimes: a React + TypeScript + Vite web app, a Go + Bubble Tea terminal UI, and a VS Code extension.

Architecture

Highlights

• Agentic tool-calling loop: the model decides when to search, scrape, or calculate, reconstructing streamed tool-call deltas server-side and looping until it's done (up to 15 iterations).
• Visual workflow builder: a React Flow canvas plus a Python topological-sort executor that lets you chain tools into a pipeline with node-to-node data references ({{node_1.result}}), like a lightweight Zapier for your own tools.
• Auto-discovering tool registry: drop a decorated Python function into backend/tools/, and it's live for the agent on next restart. No registration step.
• 400+ models, one interface: GPT, Claude, Gemini, Llama, Mistral, and image-generation models, all through a single OpenRouter integration with automatic model-type detection.
• Persistent history: every conversation, message, and tool call is stored in PostgreSQL with full analytics (execution time, status, tags).
• One backend, three clients: the web app, TUI, and VS Code extension all talk to the same /api/chat SSE stream.

Screenshots

Workflow builder canvas

Visual workflow builder: chain tools into a DAG, wire node outputs into downstream inputs with {{node_id.field}} references, and execute the whole pipeline in one click.

Chat with live tool calls

Agentic chat: the model calls web_search and scrape_webpage on its own, streaming each tool call and its output inline.

Chat with syntax-highlighted code explanation

Streamed code explanations: full syntax highlighting and structured, numbered breakdowns for code walkthroughs.

Tech stack