All projectsProject 07 of 22 · Build Agents That Remember

Celsius

What we built: Automation Flows. Screenpipe automations backed by n8n, the popular open-source workflow engine with hundreds of integrations.

The project

Problem: Screenpipe already records what you do on your computer and automates tasks with “pipes”: AI skills that carry out a whole task on a schedule or a keyword. They are powerful, but an AI that runs everything end to end is hard to predict, audit or fix. On Screenpipe’s Discord and GitHub, users keep asking for deterministic, reviewable automations, and companies won’t hand real work to an automation they can’t read.

What we built: Automation Flows. Screenpipe automations backed by n8n, the popular open-source workflow engine with hundreds of integrations. Each automation is an explicit n8n graph of steps, and AI only sits in the steps that need judgment. Before anything runs, Screenpipe shows a one-sentence AI summary of the Flow, then redraws the real n8n graph with one plain sentence per step. Nothing runs until the user clicks Turn on, and any step can be inspected and edited in n8n.

How it works. Screenpipe spots repetitive work in your recorded activity and shows it on a task page, with real screen captures and a few clarifying questions. An AI builds the n8n workflow using only nodes your n8n really has. You review it in Screenpipe’s Automations tab, where you also follow each run, get failure notifications and retry. We also built three Screenpipe nodes for n8n:

  • a deterministic trigger: a keyword said, a window opened, a pipe or Flow run;
  • an AI trigger that watches a plain-language condition, such as “when I talk about a quote in a meeting”;
  • a node that searches your Screenpipe memory.

Screenpipe and n8n run on your machine. Fewer AI calls mean lower cost, faster runs, and automations you can trust.

How it’s built

We built on Screenpipe’s own stack:

a Rust Flow service inside screenpipe-engine (axum routes under /automation-flows);
the review screens in TypeScript/React in the Tauri + Next.js desktop app, with React Flow redrawing each n8n graph;

n8n, self-hosted in Docker and driven through its REST API (create, publish, pause, executions, retries);

three custom n8n nodes written in TypeScript;

Bun, Vitest and cargo for tests.

Screenpipe is the core of the project, not a plug-in:

Inside the app: Flows live in Screenpipe’s Automations tab, and the task page builds on its detected-workflow page and notifications.
Inside the engine: the Flow service reuses Screenpipe’s local API auth and its encrypted secret store for the n8n key.

Screenpipe memory: our n8n nodes read it through the local API (/search, /activity-summary, pipe runs). They start workflows on what you say or which window opens, or when an AI judges a plain-language condition against your recent screen and audio.

AI steps call models through Screenpipe’s local AI proxy, and failed runs come back as Screenpipe notifications with Retry.

🇫🇷 made with love <3 by @nass