# inkcap

> A small, self-hosted AI workbench for chat, scheduled loops, and push notifications.

Low complexity by design: one boring stack—PostgreSQL with raw SQL, a Bun/TypeScript backend, and server-rendered Eta/HTML/CSS views. No ORM. No separate frontend build. Most pages need no JavaScript; live chat uses one small vanilla JS file.

## What it does

- Connect local llama-server and hosted OpenAI-compatible models.
- Chat with reasoning controls, message branches, and server-rendered Markdown.
- Connect MCP tools; inspect calls and require human approval where needed.
- Save a prompt, model, tools, and schedule as a manual or recurring loop.
- Get push notifications when loops finish, fail, or wait for approval.

## How it works

Runs are durable server jobs: streaming, tool loops, and persistence continue without a browser. PostgreSQL is truth; browsers render state, subscribe to replayable SSE, and submit ordinary forms. Credentials remain server-side. Messages form a tree, so edit, regenerate, and fork preserve branches.

Browser footprint: 5.9 KB Brotli-compressed chat JavaScript, 9.7 KB shared CSS, no frontend framework, and no webfonts. HTML and model traffic excluded; measured 2026-07-12.

## Run it

Requires Bun and PostgreSQL. Install dependencies, configure `DATABASE_URL` and `SESSION_SECRET`, run migrations, then start the server:

```sh
bun install
bun run db:migrate
bun run dev
```

Published image: `ghcr.io/jakswa/inkcap:latest`.

## Links

- Site: https://inkcap.click/
- Setup: https://inkcap.click/getting-started.html
- Source: https://github.com/jakswa/inkcap
- Current status: https://github.com/jakswa/inkcap/blob/main/docs/STATUS.md