Local-first · files as truth
LatchAI runs on your machine — a headless engine on :7777 and a workbench in front of it — pointed at whatever models you want. Everything it makes is a plain file in a folder you own.
Most tools give you one of these and an integration story for the other two. LatchAI runs them in a single process, which is what makes the seams between them possible.
Editor
CodeMirror 6 with an optional Vim mode, a workspace file tree, a git panel, and diagnostics from your project's own typechecker. Mount a repo and it becomes a top-level folder every tool can see.
The file you're viewing is passed to chat as context.
Canvas
Workflows are a graph you can drag, wire, and run with ⌘⏎ — or hand-write as JSON, or emit from the code-first SDK. Every view projects the same document, so there is nothing to sync.
A run knows which workflow, trigger, and story it served.
Knowledge
Agents, workflows, chats, memory, dashboards, and atlases are plain files in your LatchAI home. Chat transcripts mirror to markdown with YAML frontmatter, so an Obsidian vault indexes them as-is.
A file knows which run wrote it.
Every capability below is a surface on the same engine — the same tool registry, the same permission gate, the same event stream.
Graph JSON on disk is the single source of truth — the canvas renders it, the executor runs it, the SDK compiles to it. Arm a workflow on cron, a watched folder, or a local webhook and it runs without you.
Read the docs →Named personas in agents/*.md: system prompt, model, tool allowlist, turn budget. Turn on memory and the agent keeps durable notes it consults before it works.
Read the docs →The assistant panel shares the engine's provider, tool registry, and filesystem — so it reads, edits, and runs shell in the same workspace the editor is showing you.
Read the docs →Point an agent at a repo you've never seen and it publishes layered, interactive diagrams of how the system fits together — and a node on the map opens the real file.
Read the docs →A dashboard is a natural-language brief, not a saved query. An agent gathers the data with its tools and emits a validated document the UI renders as widgets.
Read the docs →A shadow git repo snapshots the workspace around every run, so any run has a diff. Drop a review gate into a workflow and it pauses for your approve/reject.
Read the docs →Declare stdio or HTTP MCP servers in mcp.json and their tools register beside the built-ins. Agents see one tool surface; the permission gate stays intact.
Read the docs →Download a model in-app and it runs entirely on your Mac — private, offline, free. Or connect Ollama, LM Studio, vLLM, OpenAI, or Anthropic. One Models page, one default, picked per chat or per agent node.
Read the docs →Not "privacy-friendly cloud". The engine is a process on your laptop, bound to loopback, reading and writing directories you can open in Finder.
The daemon binds 127.0.0.1 and rejects non-local Host/Origin on HTTP and WebSocket. Nothing phones home. Agent shell runs with a secret-scrubbed environment, and API keys live in the macOS Keychain rather than in any LatchAI file.
Download a model from the in-app catalog and run it on your own hardware — or connect a local vLLM, Ollama, or LM Studio server, OpenAI, or Anthropic. One OpenAI-compatible API serves them all to other apps, and a clearly-labeled demo mode keeps LatchAI explorable with no model at all.
Workflows are JSON. Agents, skills, dashboards, and atlases are markdown with frontmatter. Run logs are append-only JSONL. You can grep it, diff it, commit it, and open the whole thing in Obsidian.
LatchAI is an early project and there is no packaged download yet. You clone the repo, install, and start the engine and the UI — first boot walks you through picking a home and a model.
npm install
npm run engine # daemon on :7777
npm run ui # workbench on :5173