AI Agent

Describe it — the agent builds it

Tell the agent what you want in plain language ("build a landing page with a contact form", or "add Stripe checkout to my repo"). It writes the code, builds the Docker image, runs it on your VPS, can wire up a domain, and verifies it is actually live. Think of it as a developer teammate working on your server — you can read every step, correct it mid-run, or stop it anytime.

Your AI key, your model choice (BYOK)

The agent runs on your own AI provider key — added once under Provider Keys, encrypted at rest, and never shown again. You pick the model per project. This choice matters more than anything else: a stronger model writes better code, recovers from its own mistakes, and finishes in fewer steps.

Projects live inside your VPS

Each project is a workspace in your VPS's /projects folder. Start from scratch (greenfield), from an existing Git repository, or local-only. All the code, builds, and containers live on your own server — nothing runs in a hidden sandbox.

You stay in control

Set how much freedom each project has: Gated (approve every step), Auto to staging, or Auto to production. Steer a running agent with a quick correction without stopping it (type while it works), or hit Stop to halt it; set a spend cap so it halts at your budget. The context ring in the composer shows how full the model's window is — click it for details; long chats auto-compact near the limit. The agent can also SPAWN sub-agents (with a budget you approve first) to run parallel work on any of your VPSes — they appear in the "Spawned agents" panel right in the chat, where you watch each one live, answer its questions, steer it, or follow up after it finishes (it keeps its context). Pin your important chats to the top and rename any of them.

What it costs & how to budget

Each agent turn (iteration) carries a small platform compute fee; on the built-in Grapthway provider you also pay for the tokens it uses — fresh input, cached prefix (much cheaper), and output, each priced per million. Bring your own key and you pay your provider directly, with only the compute fee on top. In the Spend & Budget tab you set a Spend cap — the total ceiling across compute + tokens, where the run stops the instant it would be crossed — plus optional per-run limits on iterations, tokens, and time; these four controls are independent. A wallet ceiling always applies on top: a run can never spend more than your balance holds, so you never go negative. You are charged after each run settles, and the tab breaks usage down per day, model, and currency — with the iteration (compute) charge shown apart from the total so you can see exactly where your budget goes.

It uses the real platform, not a toy

The agent performs the exact same actions you would by hand: clone repos, build & run containers, create networks, expose ports, and add domains with SSL. Everything it does shows up in your Container Management and Domains menus — and changes you make there are visible to the agent. Use the Manage tab to review and tweak its work.

You can teach it — and it teaches itself

Attach reference files OR whole folders to any chat message and the agent reads them privately to do your task. Go further and teach it custom skills and tools — managed in the Agent Files tab, in two separate folders: skills (a SKILL.md playbook) under .agent-skills, and tools (runnable code with an entry script) under .agent-tools. A skill is the instructions; a tool is the executable, and a skill can list the tools it needs so the agent runs them automatically. Each folder keeps an auto catalog.json (the folder name is the default description until you set one; your edits are kept). On top of that, the agent shares a learned-skill catalog: when it solves something the built-in tools didn't cover, it saves the procedure so every future run — across all projects — can find it. The whole fleet gets smarter from real usage.

It keeps things running

Beyond the first build, the agent can maintain your app — scheduled dependency updates and automatic self-heal when a deploy or run fails — all within your guardrails and spend cap.

It runs routines on a schedule

Just ask in chat to run a task on a schedule — hourly, daily, a cron expression, or once at a time. For example: “every hour, scan my service logs for errors and report,” nightly dependency checks, or a periodic health probe. The agent sets up the schedule, runs it on cadence within your guardrails and spend cap, and you can list, edit, or cancel it anytime.