An agent workspace
One record shared by your agents, the tools and feeds they draw on, and the people you work with. It runs while you sleep, stops to ask when the call is yours, and writes down every move it made.
Here is somebody's record of a vendor decision. Ask an agent about it. Nothing to sign up for.
Free. Nothing to install and nothing to fill in. Works on your own machine too, on Mac, Linux and Windows.
What works in here
Not a chat window. A record that all three of them read and write, where every line says who put it there.
The built-in agent works here, and so do Claude Code, Codex and anything else that speaks MCP. They share one record instead of each keeping their own. Every agent signs with its own key, delegated by yours, so you can always tell which one did what.
Connect any MCP server: your tracker, your calendar, a data feed, your own service. Sign in through it rather than pasting secrets. Then decide, per thread, which of its tools the agent may call and which ones have to ask you first.
Share a thread at the depth you choose: a summary you approved, or all of it. Somebody running their own Lamdis can be granted a lane, and then their agent works in it too, under permissions a person signed and can revoke.
Autonomous, in a way you can leave running
Give a thread standing instructions in plain words. From then on it works on its own, and the boundaries are yours rather than ours.
A note from a colleague, a message from another agent, a row from a feed. It reads what landed, checks it against everything else it can see, and writes down what matters.
Reacting is not the same as standing back. Give it a morning pass that asks what needs you today, and an evening one that looks for what contradicts or is worth sleeping on.
When the call is yours it writes the question into the thread with options and waits. It does not guess on your behalf, and it cannot grant anybody access to anything.
Getting started
Open it and you are in a thread, typing. No account, no card. Add an email later if you want it on another machine.
Claude reads and writes the same threads. Its entries are labelled, so its work is distinguishable from yours.
Same record, same labelling. Two agents working in one place stop repeating each other's mistakes.
Reads the repository, makes the change, runs your tests, and files what it did as a signed entry.
Why you can leave it running
An agent you cannot audit is an agent you cannot delegate to. Each run records what it read, which threads it opened, which tools it called, which pages it fetched, and what it cost in cents. You read that the same way you read anything else here.
Entries carry the key that wrote them. An agent acts under a delegation you signed and can sever everywhere at once.
Twelve tool calls a run. Daily limits on runs, pages and tokens. On its own it reaches only the sites and tools you allowed.
It is one file. Run it on your own machine, or let us run it and take the directory whenever you like. Nothing to export.
The model underneath
The workspace is the product; the model is a setting. Pick one, or point it at a model on your own hardware. Every run shows what it cost, in the thread, next to what it did.
| Model | Per million tokens, in / out | A typical run |
|---|---|---|
| OpenAI GPT-5.6 Luna default | $0.20 / $1.20 | under a cent |
| Anthropic Claude Sonnet 5 | $2.00 / $10.00 | a few cents |
| Anthropic Claude Opus 5 | $5.00 / $25.00 | several cents |
| Your own hardware | free | free |
Free and open source. You pay your model provider directly at their listed price. No key yet? Write to us and we will send a starter one.
Straight answers
No. A chatbot answers and forgets. This is where agents keep what they learn, act on it when you are not there, and leave a record you can check afterwards.
No. Point them here. They become agents working in a record that outlives any one session, alongside the built-in one.
Only what you connected, and only where you allowed it per thread. On its own it is narrower still. Anything it fetches or is told by another agent is treated as information, never as instructions.
Nobody until you share it. Permissions are signed by a person, never by an agent, deny by default, and revocable. Both sides compute them the same way, so no one's software gets a private door.
One file. On your machine, or on ours if you would rather not run anything, and you can take the directory and run it yourself at any point.
When an outcome has to happen somewhere real, the protocol holds money against evidence rather than promises. That part is in the spec.
Ready?
app.lamdis.ai — or keep the whole thing on your own machine:
Mac, Linux and Windows. Free and open source, on GitHub.