Docs / MCP

Connect an agent over MCP

Safe write access for AI agents. Add one MCP server and your agent can record, correct and retract business facts. Every change is kept and attributed to the agent that made it, and any change can be undone.

Hosted on AvianSuite

Add this URL as a remote MCP server (Streamable HTTP):

https://mcp.aviansuite.com/mcp

Claude, ChatGPT and Cursor sign you in with OAuth and ask you to approve the connection. Approving adds the app as an agent connection on your store, so you can see, undo or revoke what it does. Other clients send an agent token from Agents as Authorization: Bearer.

Claude Code:

claude mcp add --transport http aviansuite https://mcp.aviansuite.com/mcp

Cursor, in .cursor/mcp.json:

{"mcpServers": {"aviansuite": {"url": "https://mcp.aviansuite.com/mcp"}}}

Agent frameworks and scripts:

{"mcpServers": {"aviansuite": {"url": "https://mcp.aviansuite.com/mcp", "headers": {"Authorization": "Bearer YOUR_AGENT_TOKEN"}}}}

No account yet

Add this URL as a connector instead. It needs no sign-in, so your agent can set everything up in the chat:

https://mcp.aviansuite.com/start

Agents can't add connectors themselves, so this step is yours: in Claude, open Settings → Connectors → Add custom connector and paste the URL, then start a new chat. In Claude Code, run claude mcp add --transport http aviansuite https://mcp.aviansuite.com/start. If the agent runs in a sandbox with an allowlist (a cloud environment, a container), also allow mcp.aviansuite.com and app.aviansuite.com.

Ask the agent to create a workspace. Its create_workspace tool asks for your email and what it will record, and the workspace works in that chat at once. The agent gives you a claim link to approve it, and a personal connector URL for later chats (treat it like a password).

Agents without MCP can do the same over HTTP:

curl -X POST https://app.aviansuite.com/api/agent-signup -H 'Content-Type: application/json' \
  -d '{"email": "person@example.com", "purpose": "Keep ticket status for the help desk"}'

The response has an MCP token the agent can use right away and a claim link for the person. Sandboxes allow 2 agent connections and 100 MB. They go read-only after 7 days unless the person claims them and adds a card, which starts the normal 14-day trial with nothing lost.

Receipts

Every write returns a receipt link. Agents include it when they report a change, and the person sees exactly what changed, which agent did it, and an Undo button. Only people signed in to the workspace see the change.

Self-hosted Stellar Jay

Run the same tools over stdio against your own store. Self-hosting is free.

go install github.com/kyle-visner/stellarjay/cmd/stellarjay-mcp@latest

{"mcpServers": {"aviansuite": {"command": "stellarjay-mcp", "env": {"STELLARJAY_URL": "https://store.example.com", "STELLARJAY_TOKEN": "writer-token"}}}}

Tools

record_fact: save a business fact with its evidence.
correct_fact: replace an earlier fact. The old one stays in history.
retract_fact: withdraw a fact with a reason.
get_entity: current facts and full history for one record.
list_changes: what changed in a time window, filtered by agent or record.
undo_changes: preview, then reverse everything an agent did in a time window. A dry run unless you confirm.
save_checkpoint: name the current state, such as before-import.
status: store health and current root.

Nothing is ever overwritten, so an undo is itself a change you can undo. Each write takes an operation_id, so a retried write is recorded once.