Claude and ChatGPT (chat)
This is not available on mcp.align.tech yet. It runs on the preview environment
(mcp.preview.align.tech) while we test it. Nobody has completed a connect from claude.ai or
ChatGPT end to end yet. This page says which steps we have checked and which we have not.
The AI Assistants page covers coding agents. This page is for people who work in a chat window. You add Align as a connector, sign in with your normal Align login, approve what the assistant may do, and then ask questions in plain language.
What the assistant can do
A chat connection gets three read-only tools:
| Tool | Use |
|---|---|
search | Find decisions that match a question. Returns up to 10 results with a title and link. |
fetch | Read one decision in full, by the id search returned. |
check_a_plan | Check one plain sentence ("we will tell the customer the date is fixed") against past decisions. |
It cannot capture, edit, approve or delete anything. It sees only the decisions you can see in
Align. A decision whose title ends in brackets, such as (replaced), is not the current position.
Two things to know before you approve:
- Checks made from chat that match decisions are logged in Align's audit trail.
- The connection is tied to you and your workspace. Align takes the workspace from your sign-in, never from anything the assistant sends.
Connect from Claude
Checked against Anthropic's custom connector guide (support.claude.com, read 2026-10-20). Not yet done by us against Align.
- Pro and Max: open Customize, then Connectors, then + Add, then Add custom connector. On Team and Enterprise an owner adds the connector first under Organization settings, then Connectors, and members then click Connect.
- Enter a name, and the URL
https://mcp.preview.align.tech/tools. The path matters: the server answers MCP at/tools, and the root address returns 404. - When Claude asks about the OAuth client, pick the option where Claude registers automatically. Align supports dynamic client registration. It does not publish a client ID metadata document yet, so we do not expect Claude's "published identity" option to work. This is untested.
- Sign in to Align when the browser opens, read the consent screen, and choose Allow access.
- In a chat, click +, then Connectors, and switch the Align connector on.
Connect from ChatGPT
We could not read OpenAI's developer mode guide: the pages we tried returned 404. What follows comes from third-party write-ups and is not verified. Menu names and plan limits differ between them.
- Turn on developer mode in ChatGPT settings, under Apps and Connectors, then Advanced settings.
- Create a connector. Enter the URL
https://mcp.preview.align.tech/toolsand choose OAuth. - Sign in to Align, read the consent screen, and choose Allow access.
Align's tools are named search and fetch, the pair OpenAI documents for deep research and
company knowledge. We have not verified whether developer mode requires that pair. We built it in
anyway.
Revoke access
Open Settings, then Connected assistants in the Align web app. Each assistant has a Revoke button. Revoking ends that connection. The assistant can keep working for up to 30 seconds after you revoke, while the connector's check expires.
What can go wrong
- The connector says it cannot reach the server. Claude and ChatGPT connect from their own cloud, so a firewall in front of your Align must allow them. The preview host is public.
- You see a consent screen you did not start. Choose Deny. A link sent to you by someone else can open a screen for their assistant session. Check the app name and the host the page says you will return to.
- The assistant answers without calling Align. Name it in the question: "Ask Align why we chose X".
Limits today
- Preview only. Production is off.
- One permission level: read-only (
mcp:read). - Access tokens last one hour and refresh on their own. A connection that nobody uses for 30 days ends, and no connection lasts more than 90 days.