The company context layer

Reliable context, for every agent.

Structure what your company knows. Maintain it with people in the loop. Retrieve it from Claude, ChatGPT, your coding agent, Slack and any agent that speaks MCP, all reading the same copy.

Read by the agents your team already uses

  • Claude
  • ChatGPT
  • Claude Code
  • Codex
  • Cursor
  • Slack agents

[01] One copy

Kept once. Read by every agent.

Skills are how work gets done. Knowledge is what is true. Metrics are what words mean. Each is kept once, in Atlas, and every agent retrieves the same copy under the same access rules.

skillsknowledgemetricsand 64 moreand 412 moreand 87 moreskills/month-end-close.mdskills/deploy-runbook.mdskills/incident-review.mdknowledge/ideal-customer.mdknowledge/fy27-goals.mdknowledge/pricing-policy.mdknowledge/payout-schedule.mdknowledge/refund-approvals.mdmetrics/churn.sqlmetrics/net_revenue_retention.sqlmetrics/arr.sqlClaude Desktop43 peopleChatGPT18 peopleCoding agent12 peopleSlack agent1 agentCommand line9 peopleSupport agent2 agentsAny MCP agent4 agents
atlas
scope · version · source

[02] The product

Three layers. Each one on point.

A context layer is only as reliable as its weakest layer, so Atlas is built as three. Pick a layer, then a part, to see it inside the app.

guide · before you start · read by every agent
skill · month-end-close · v0.1.4 · published
page · finance / payout schedule · live
table · analytics.orders · 14 columns · read-only
metric · active customers · bigquery · run preview
scope · finance › payout schedule › maya chen (edit)
change request · payout schedule · from slack agent · pending review
audit · payout schedule · who, what, when
automation · weekly churn notes · mondays 9:00
insights · 12 stale pages · 6 unanswered questions
app.atlascontext.ai/mcp · sign in as yourself
search · "refund policy" · only what you can open
playground · "when do payouts settle?" · 3 sources
evaluation · finance questions · 48 / 52 passing

[03] Under the hood

Reliable by default.

The unglamorous parts are what make context trustworthy. They are on in every workspace, for people and agents alike.

01Access
Agents sign in as a person over OAuth. Search checks access before it ranks, so an agent never sees more than its person.
02Changes
Agent edits arrive as change requests. A person reads the diff and approves it, block by block.
03History
Every save is a version. Compare any two, restore any one. Deleted pages go to a trash you can restore from.
04Audit
Every read and every write, by people and by agents, in one workspace-wide trail.
05Warehouse
BigQuery and Snowflake, read-only. A metric is SQL agreed once, so every agent gives the same number.
06Models
Atlas calls a model only when a workflow you set up runs, with the provider and key you connect.
07Hosting
Run for you by Ensemble Labs on Fly.io in Paris. One workspace per company, invitation-only.

[04] Connect

One endpoint. Any agent.

Point your agent at Atlas and sign in as yourself. It reads what you can read, and its edits come back to your team as change requests.

  1. Your agent opens Atlas in a browser. You sign in and pick a workspace.
  2. You approve access for this agent. It gets its own grant, never your session.
  3. Every call checks your current access. Lose access to a page and so does your agent.

terminal

claude mcp add --transport http atlas https://app.atlascontext.ai/mcp/

Run /mcp in Claude Code and sign in.

[05] Solutions

One layer, two ways in.

Company brainOne place your company agrees on what is true.Procedures, facts and numbers written down once with an owner, kept current by people, and read by everyone the same way.Explore Agent memoryGive every agent the same memory.Each agent signs in as a person, reads what that person may read, and remembers through Atlas instead of on its own.Explore

Start with one team.

Bring one team's procedures, pages and definitions. We structure them with you, run Atlas for you, and connect the agents you already use.

Book a demo