Atrium
Context gateway · operational

Your team's context, loaded into every agent session.

Atrium is the context gateway for AI agents. One call gives Hermes, Claude Code, Codex, or your own agent the company, the active sprint, blockers, and the decisions your team already made — so the first message is useful instead of a re-explanation.

Sign in to dashboard → Health check
1 callto bootstrap context
7-daytrial, no card
Any agentHTTP or MCP
Capabilities

Everything an agent needs before it starts work

These are the capabilities shipping today in the Atrium gateway and admin dashboard.

Agent context bundles

Pre-load company, sprint, blockers, and team roster for Hermes, Claude Code, Codex, or any HTTP client — one request instead of five.

Sprint generation & scheduling

LLM-powered sprint plans from your backlog, then auto-scheduled across phases with capacity estimates per assignee.

Meeting intelligence

Drop in a transcript and Atrium extracts action items and decisions, then persists them as long-term agent memory.

Agent delegation

Assign a sprint to an agent and track it from assignment to completion, scoped per company so credentials never over-reach.

Project intelligence

Kanban board, Gantt timeline, sprint burndown, and an EVM panel (CPI/SPI) computed from live task data.

Scoped keys & device auth

Multi-company API keys with route scopes, plus browser-approved device authorization for the CLI so secrets stay in the OS keychain.

Public venture profiles

GET /token/:chain/:address renders a token venture anonymously — every claim carries its own evidence class (L0 self-reported through L4) and a source, with an honest empty state when nothing is verified.

Wallet-signed operator claims

An operator signs an EIP-191 challenge with their own wallet. The gateway recovers the secp256k1 signer and revalidates it against a live chain RPC before granting anything — an unreachable endpoint is refused, and nothing is published until a claim actually succeeds.

Append-only goal history

Goal and milestone changes are appended, never overwritten: earlier versions stay on the page, struck through, and supersession is derived from version order rather than a stored flag. The authenticated writer route is live; the public history remains empty until a real goal is recorded.

Authenticated goal recording

POST /api/v1/ventures/:id/goals validates and appends one goal or milestone version through the database writer, returns explicit refusal states, and is deliberately non-billable.

Venture profiles

A venture page anyone can read, with every claim labelled by its evidence

GET /token/:chain/:address is server-rendered HTML at the public root: no account, no key, no JavaScript. Each claim is shown with the evidence class it earned and the source behind it, so a reader can tell the operator's own word from something independently checkable.

L0_self Self-reported The operator's own word, and it is shown as such: dashed edge, no checkmark, never rendered as verified.
L1_signed Wallet-signed Signed by an address. It proves who spoke — not that what they said is true.
L2_source Source-verified A resolvable artifact URL: a commit, a pull request, or a deployment.
L3_onchain Onchain-verified Read from a public chain. It proves value moved — never that it was revenue.
L4_legal Legally verified Only with a named, jurisdiction-identified legal review. An anonymous document never reaches this class.
Open LUDI on Bankr →
Agent workflow

The loop that is live today

What an agent actually does against this gateway, end to end — and where the evidence gate sits in that loop.

1

Read the task board

The agent pulls the workspace's open work — the active sprint's task board and its action items — as the starting state for the session.

2

Per-task delegation

The board is worked per task: a single task is delegated to an agent and tracked from assignment to a reported status, scoped to one workspace.

3

Classify the evidence

Each completion claim is checked against the artifact behind it by the evidence classifier. An over-claim is downgraded rather than accepted: a transaction asserted against a pull request artifact is downgraded to the source class the artifact actually supports.

4

Write back to the sprint

The agent's status and result are a write-back into the sprint, so the next session's context bundle already reflects the work that was done.

How it works

From empty repo to context-aware agent in three steps

1

Create a workspace

Describe the company and project. Atrium starts a 7-day trial and seeds a starter project with tasks.

2

Connect your agent

Generate a scoped key or run the CLI device-approval flow. No secret is ever pasted into a prompt.

3

Start a session

One POST to /api/v1/session/start returns the full context bundle your agent plans from.

API

Core endpoints

All write endpoints require a scoped API key. Health and the API index are public.

GET /health Liveness check
POST /api/v1/session/start Pre-load context for any agent
POST /api/v1/meeting/:id/transcript Transcript → action items + decisions
POST /api/v1/projects/:id/sprints/generate LLM-powered sprint generation
POST /api/v1/projects/:id/sprints/:sid/assign Assign an agent to a sprint
GET /api/v1/projects/:id/sprints/current Active sprint with burndown
GET /api/v1/action-items Kanban-ready, grouped by status
GET /api/v1/projects/:id/detail Full project: tasks, sprints, metrics, teams
GET /token/:chain/:address Public venture profile — HTML, anonymous, no JavaScript
POST /api/v1/ventures/:id/claims Submit a wallet-signed operator claim
POST /api/v1/tasks/:id/evidence Attach evidence to a task; the class is derived, never asserted
GET /api/v1/tasks/:id/completion The completion claim as independently classified
PATCH /api/v1/projects/:id/sprints/:sid/assignments/:aid Agent reports status back into the sprint
Payments

Agents pay per call — and the meter says what it cannot yet do

Metering is per call and recorded, so a caller is never charged for something the gateway cannot serve.

x402, per call

A caller with no allowance gets a 402 carrying the price of that call and an x402 payment challenge. The gateway verifies and settles the payment and returns a receipt — one call, one price, no subscription needed.

A fixed allowance order

A call is metered in one order and stops at the first step that covers it: a trial or included allowance (free tier, subscription, or grant), then a prepaid credit balance, then invoice overage, then x402. The billing tables are append-only.

No top-up route yet

There is no self-serve top-up: you cannot buy credits through the gateway today, so a 402 says that in words instead of linking a page that does not exist. Built, but carrying no data yet.

Quickstart

Start a session with context

Replace the company id with the workspace you created in the dashboard.

# Start a session with context
curl -X POST https://atrium.cool/api/v1/session/start \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <atrium_key>" \
  -d '{"company_id":"<uuid>","agent_type":"hermes"}'
# Or hand one company to your agent through the CLI
hermes-atrium <company> "brief me"

The CLI takes a single company argument — your workspace name — and hermes-atrium <company> "brief me" returns that workspace's context. Device approval keeps the key in the OS keychain, so no secret is pasted into a prompt.

Payment is per call: an agent without an allowance pays for a single call with the x402 challenge the gateway returns, and there is no subscription to set up first. Prepaid credits are not purchasable in the gateway yet — see the honest status below.

Honest status

This page states only what is true today

  • Live: the context gateway and its session bundles; the public venture page with its L0–L4 evidence classes; wallet-signed operator claims checked against a live chain RPC; the agent workflow including the evidence classifier; and x402 per-call payment.
  • Built, but carrying no data yet: the public venture page renders an honest empty state — Atrium's registry has not imported LUDI evidence yet; and the append-only goal and milestone history remains empty until a real goal is recorded through the live writer route.
  • Not built: no self-serve top-up. Nobody can buy credits through the gateway today, and nothing here pretends otherwise.