Open Source · Actively Built

Atlan cockpit.

Chat, editor, terminal, live preview, agent fleet, and build pipeline — one surface that orchestrates every major coding agent on your own subscriptions, plus a free local model running on your own hardware. Built phone-first, bound to loopback, Apache-2.0.

I got tired of building across a dozen fragmented tools. Atlan is the answer: a single cockpit where the model, the code, the terminal, the preview, and the agent runs all live in one place — and where I can do all of it from a phone if that's what I have on me.

Every engine, one surface, your own subscriptions

The model picker holds Claude, Codex, Gemini, Grok, and Copilot as first-class coding agents, a dozen bring-your-own-key chat providers, and — at the bottom of the list — llama-server, running free on my own GPU. Six local models sit on disk, hot-swappable from inside the cockpit: pick one, and it repoints the symlink and restarts the service. Everything authenticates on your own subscription, within each vendor's terms. No API keys required to start, and no paywall between a person and their first build.

Probabilistic workers, deterministic walls

This is the thesis that runs through everything I build. Agents are excellent at proposing and unreliable at stopping. So the walls are code, not vibes:

  • Budgets that halt, not warn — a run stops dead at its token ceiling and waits for you to top it up
  • Deterministic checkers grade agent output — code grades it, never another model
  • An escalation ladder: free local model first, cheap cloud next, frontier only when the checkers demand it
  • Double-belted permissions — profile-level tool denial plus a runtime gate, because either one alone has been proven insufficient
Atlan Fleet tab showing agent runs tagged KILLED, DONE, and BUDGET HALT, each with token counts, cost estimates, and a top-up-and-resume button.
Fleet: real runs, hard budgets. A halted run waits for a deliberate top-up instead of quietly burning through your plan.

The preflight gate

Before anything is exposed, a Doctor-tab panel has to come back green: bound to loopback only, password-gated sessions with hashed tokens, keys encrypted at rest, no plaintext key files, secrets git-ignored, permission gate on, no active tunnels. The amber items sit right next to the green ones — the gaps are listed, not hidden, and the project ships a security document that names what's still open.

Atlan Doctor tab preflight checklist, all green: server bound to 127.0.0.1 only, access auth layer, keys encrypted at rest, no plaintext key files, secrets git-ignored, permission gate, no active tunnels, preview proxy local-only. Below it, a Scan panel offering a SAST audit powered by the PreFlight engine.
The preflight gate, plus the Scan panel — PreFlight's engine embedded directly in the cockpit.

The shipyard and the inspection bay

PreFlight — my free, in-browser security scanner — now runs inside Atlan as a Scan tab. The cockpit that writes the code also audits it before it flies, with the engine vendored in and version-pinned, and nothing leaving the machine. Two open-source projects that each dogfood the other.

Atlan chat tab: the free local llama-server model answering a question about why a hard token budget that halts an agent is safer than trusting the agent to stop on its own, with a 205 token counter.
The free tier answering for itself — a 35-billion-parameter model on my own hardware, no meter running.

Why there's no live demo

Because the honest answer is that you shouldn't host one. Atlan is a terminal, a filesystem API, and an agent with shell access — hosting that on the open internet is hosting remote code execution. It binds to 127.0.0.1 on purpose, and reaching it from elsewhere is a deliberate, authenticated act. A project that markets against its own security documentation isn't worth trusting. So: screenshots, source, and an invitation to run it yourself.