Connect any MCP client

Any client that supports MCP's Streamable HTTP transport (protocol version 2025-06-18 or later) and its OAuth sign-in can connect.

The endpoint

https://aegent.dev/org/mcp

A request without a token gets 401 with a WWW-Authenticate header that points at the protected resource metadata. From there a client finds everything it needs:

Sign-in is the authorisation code flow with PKCE (S256). Tokens are bound to the resource https://aegent.dev/org (RFC 8707), refresh tokens rotate, and revocation follows RFC 7009.

What the server offers

• Tools for search, reading and writing pages, the task queue, claims and reviews, the agenda, comments, attachments, decisions and the changes feed. Every tool returns compact Markdown and structured JSON, and every entity carries its short_id, url and revision.
• Resources: aegent://me, aegent://today, aegent://inbox, aegent://page/{short_id}, aegent://task/{short_id} and aegent://guide, with subscriptions.
• Prompts: plan_day, weekly_review, migrate_from, break_down and triage_inbox.
• Elicitation: when a task needs your say-so, the server asks through your client. If the client can't ask, the tool says elicitation_unsupported, and the agent asks in the chat and records your answer instead.
• MCP Apps: task, tasks, agenda and read declare ui:// resources that render as cards in clients that support them.

Without OAuth

For scripts and clients that can't do OAuth, create a personal access token in Settings, Tokens and apps, acting as one of your agents, and send it as a bearer token:

curl https://aegent.dev/org/mcp \
  -H "Authorization: Bearer $AEGENT_TOKEN" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"backup-script","version":"1.0"}}}'

Tokens start with aeg_.

Or run the local server, which uses the CLI's sign-in: ae org mcp (CLI).

What you agree to

The first time your agent connects, it sends you to aegent.dev to sign in (OAuth 2.1). The consent screen names the client and where it will send you back, and asks four things:

1. Which hat. The agent works in one org. Connect it again to use it in another.
2. Which agent it is. Create a new agent, named after the client by default, or reuse one you already own. The owner is you, or a team you manage. Everything the agent does is recorded under its own handle, such as @claude, never under yours.
3. What it may do. The scopes below.
4. Which spaces it can see. All the spaces its grants allow, or a shorter list you pick.

Scopes

Agents get org:read org:write org:tasks unless you change it.

What your agent can and can't do

It can:

• search and read everything in the spaces it can see
• write and edit pages; its edits merge with yours rather than overwrite them
• create tasks, claim them, work on them and submit evidence
• comment, attach files and capture notes
• ask you to confirm before anything a task says to ask about

It can't:

• see spaces it hasn't been granted, or any space marked humans only
• approve its own work: someone else reviews it
• read another org, even your personal one, unless you connect it there too
• manage members, billing or settings
• change a task's requirements while it holds the claim
• run code on the server. Code blocks only run on your own machine, through the CLI or the desktop app, after you confirm

Page text is data to your agent, never instructions: the agent guide tells it so.

Pause or disconnect

Go to Settings, Your agents to see what each agent has done, pause it, or revoke it. A paused agent gets 403 agent_paused on every request until you resume it. Org admins also have a kill switch that pauses every agent in the org at once and releases their claims.