Guidance
How agents should behave here. A few short guides an agent reads before it starts: tone, house rules, what to check first.
How it works
Structure. Maintain. Retrieve.
A context layer is only as reliable as its weakest layer. Atlas is built as three, and each does one job: your team structures what it agrees on, people keep it right, and agents read it over MCP.
The loop
Four steps, on repeat.
- StructureYour team writes down what it agrees on as skills, pages and metric definitions, each with an owner and a scope, in a live editor with comments and history.
- ConnectClaude Desktop, ChatGPT, your coding agent, a Slack agent or any agent that speaks MCP connects and signs in as a person.
- RetrieveAn agent searches and reads what its person can open, and answers from the same copy as every other agent.
- MaintainWhen an agent wants to fix something, it opens a change request, and a person approves the diff. Automations keep pages current between edits.
[01] Structure
Structure what your company knows.
Five kinds of context, each a typed object with an owner and a scope, each kept once.
Skills
How a job is done here. Written procedures that agents follow, like the month-end close or the deploy runbook. Each save is a new version with a note on what changed.
Knowledge
What is true here. Pages your team writes together, with comments, tasks and links between pages. Agents read the same pages people do.
Tables
Your warehouse, described. Each table carries notes on what it holds and how to use it, so every agent queries it the same way.
Metrics
What a number means here. A metric is SQL over your BigQuery or Snowflake warehouse, agreed once, so every agent gives the same number for churn.
[02] Maintain
Maintain it with people in the loop.
Nothing changes silently. Every edit is a version, every agent edit is a review, and every change is on the record.
Versions
Nothing is overwritten. Open any earlier version, compare it, and restore it as a new one. Deleted pages go to a trash you can restore from.
Change requests
An agent's edit arrives as a diff, like a pull request. A person approves or rejects it, block by block.
Automations
Scheduled and triggered work runs a model you connect, reads your sources and proposes updates, so context stays current without anyone chasing it.
Audit and insights
Every read and change is on the record, and insights show usage, stale pages and unanswered questions.
[03] Retrieve
Retrieve it from any agent.
One endpoint, the same access rules as the browser, and a warehouse that never gets written to.
One endpoint
Every agent, from Claude Desktop to a script on a schedule, connects to the same MCP endpoint and gets the same tools to search, read and propose changes.
Access travels with the person
Pages and skills sit in a teamspace or a private space. The browser, the API and MCP use the same rules, so an agent sees what its person sees.
Playground and evaluations
Ask as an agent would, see what it finds, and score retrieval against reference questions over time.
Your model, your key
Atlas calls a model only when a workflow you set up runs, with the provider and key you connect.
Operations
We run it for you.
Hosted in Paris
Ensemble Labs hosts Atlas on Fly.io. Each company is its own workspace, and access is invitation-only.
Set up with you
We bring one team's procedures, pages and definitions into Atlas together, then connect the agents you already use.
Start with one team.
Bring one team's procedures, pages and definitions. We run Atlas for you and connect the agents you already use.