Lattice
no model
Agent networks · no code

Build the team.
Then reheat the work
until it holds.

Put agents on a board and tell each one its job in plain words. Wire in tools and what you already know. Then wrap a loop around the part that has to be good — it runs again and again until a reviewer signs off.

Frontier APIs Ollama & LM Studio WebGPU, in this tab
LOOP · UNTIL APPROVED PASS 3/5 BRIEF AGENT Writer REVIEWER Scores 1–10 KNOWLEDGE Style guide TOOL DELIVER
How to build one

Four moves, in order

Each one depends on the one before it. Do them in sequence and you will have something running in about three minutes.

FIRST

Pick your agents

Drag an agent onto the board and say who it is in plain words — “You review copy for tone and cut anything vague.” That description is its entire job.

THEN

Connect them

Drag from one agent's right dot to another's left dot. Whatever the first produces becomes the second's input. That is a connection — there is nothing else to it.

THEN

Give them tools and facts

Wire in a web request, a calculator, or a slice of your knowledge graph. Agents can only reach what you wire in — nothing else on your machine, nothing else online.

LAST

Wrap a loop around it

Select the blocks that have to be good and wrap them in a loop. Set what makes it stop: a reviewer approving, a score clearing a bar, or a fixed number of passes.

Legend

The parts you can place

Start from something real

Networks you can open now

Every one of these runs. Open it, see how it is wired, then change the parts you do not like.

Where the thinking happens

Three kinds of model, one board

Set a default for the whole network, then override it per agent. A cheap local model can draft while a frontier model reviews — or the other way around.

Frontier APIs
Anthropic, OpenAI, Google, Groq, OpenRouter, or anything OpenAI-compatible. Your key is encrypted with a passphrase and stays in this browser.
Your own machine
Ollama or LM Studio, here or elsewhere on your network. Lattice reads the model list off the server, so what you see is what is installed.
In the browser
WebGPU models load into this tab and run on your graphics card. Slow to start, then offline and free. Good for loops that need many cheap passes.

Everything — networks, keys, knowledge, run history — is kept in this browser and nowhere else. The only thing that leaves is the text you send to the model provider you chose. Clearing site data erases all of it, so export anything you want to keep.

Untitled network

Empty board

Place an agent to begin, or open a working example and take it apart.

Knowledge graph

Facts your agents can look up

A knowledge graph stores things and how they relate — "Acme supplies Northwind", "Northwind owns the Redmond site". Agents ask a question, Lattice walks the connections outward from the closest matches and hands back only what's relevant. Paste text and it will pull the graph out for you.

Add knowledge
Uses your active model. It reads the text and proposes things and relationships — you approve them before they're saved.

Add one by hand
Try a lookup

          
0 things · 0 relationships
Drag to pan, pinch or scroll to zoom, drag a circle to move it. Colour groups things of the same kind.
Models

Choose where the thinking happens

Set up one provider to get started. The one marked Active is used by any agent that hasn't been given its own model. Keys are encrypted with a passphrase you set and stored only in this browser.

Default for this network
Nothing set up yet
Cost guard

Loops can run away. Lattice stops a run when it hits either of these, whichever comes first.

Runs

What happened, step by step

Result
No run yet. Open the Studio and press Run.
Trace