Lobstack documentation
Lobstack is a desktop app where bots do real work and ask before they change anything, and the API it runs on: 17 models from 5 providers behind one OpenAI-compatible endpoint.
The Lobstack app is free for Windows, macOS and Linux. Bots work in your files, repositories and connected tools. A step that only reads runs straight through; a step that changes something waits for your approval. The API, the Lobstack Gateway, is what the app runs on, and you can call it from your own code too.
The app in three steps
- 1
Download it
From lobstack.ai/download. It is a public beta and the builds are not code-signed yet, so Windows and macOS warn the first time you open it.
- 2
Sign in and pick a bot
One sign-in through your browser, and the first one from the app adds $5 of credit for 30 days. A new workspace starts with three bots; make your own with New bot.
- 3
Give it a job, and approve what changes something
Searching and reading run. Writing a file, running a command or posting a document stops on a card that shows exactly what will happen, until you approve it. The whole walk-through is the app quickstart.
The app, the API and your account
Lobstack app
A desktop application: bots that do real work on your machine, behind an approval gate. Public beta.
Gateway
One OpenAI-compatible endpoint over 5 providers, with complexity-based routing and per-request cost metering.
Console
Watch what the Gateway did and manage what can call it: traces, spend, request logs, keys, the model catalogue, webhooks.
Where to start
With the app
App quickstart
Download, sign in, give a bot its first job, and approve the step that changes something.
The Lobstack app
Engines, the approval gate, connectors, the built-in browser, and what leaves your machine.
Skills & connectors
What a skill teaches a bot, what a connector lets it reach, and how to write a skill.
Webhook triggers
Start a bot from a Stripe, GitHub or Console webhook.
With the API
API quickstart
An account, a key, one request, and the receipt it comes back with.
Concepts
Keys, tiers, routing and metering — the vocabulary the API pages assume.
Chat Completions
The endpoint in full: body, response, streaming, tools.
Metering & cost
What every response tells you about what it cost, and what the saving was measured against.
Pricing & plans
Four plans, an allowance in dollars of model spend, three meters, and what the ceiling does.
Errors & retries
Six classes, their status codes, and which of them are worth retrying.
CLI
The Gateway from a terminal, with the cost of each call as it happens, and a local OpenAI-compatible proxy.
MCP server
Four tools inside Claude Desktop, Claude Code, Cursor or Zed — including what a call would cost before you make it.
The API in one request
If you have used the OpenAI SDK, you already know the Gateway. Change the base URL and the key; everything else — the request shape, the response shape, streaming — stays as it is. What changes is what comes back with the answer: the model that served it, the tier it was scored into, and the price, on the response itself rather than in a dashboard you have to go and open. "auto" hands the choice to Nex, the router.
The key comes from the Console; the API quickstart goes from a new account to this request and the headers it returns.
What we do not claim
Routing a simple request to a cheaper model is not a differentiator — OpenRouter, LiteLLM and Requesty all do it, and LiteLLM does it free and open source. We would rather you heard that here than found it out later.
What is unusual is the receipt, and specifically that it says what its own numbers mean. A dollar saving comes back with the model it was measured against and baseline_reason saying why that model: named when you asked for it, plan_ceiling when you sent "auto" and the comparison is against the best model your plan allows. Those are different claims and the response never presents them as the same one. Where there is nothing honest to compare against, the field is null rather than a number. The rules are on Metering & cost, and you can run the router on your own prompt without spending a token on the routing playground.