← Help

How do I use AccessLint from Claude, ChatGPT, or Codex?

AccessLint ships a connector: a hosted MCP server at https://mcp.accesslint.com/mcp. Add it once and you can scan a page, turn a sentence about a user journey into a draft flow, start a dry run, and read what the browser saw, without leaving the conversation.

The assistant does the talking. AccessLint does the auditing. Nothing about your account changes except what you ask for, and one thing can’t change at all: only a person can start monitoring, in the web app. See what the connector can’t do.

What you need

• An AccessLint account. Sign in the usual way: email or Google.
• For flows: a verified domain. Scanning a single page needs no domain at all, so you can start while DNS propagates.

Add the connector

Claude

1. Open Settings › Connectors and choose Add custom connector.
2. Enter https://mcp.accesslint.com/mcp and add it.
3. Claude sends you to AccessLint to sign in. Pick the workspace you want the connector to act in, and approve.

ChatGPT

Adding a custom MCP connector needs developer mode, on the ChatGPT web app, on a paid plan. Turn it on in ChatGPT’s advanced settings, then add https://mcp.accesslint.com/mcp and complete the same AccessLint sign-in and workspace choice as above.

OpenAI renames and moves this surface between releases: it has been called Connectors, Apps, and Plugins, and developer mode has moved with it. Look for developer mode in ChatGPT’s advanced settings, or read OpenAI’s own instructions. The URL you paste never changes.

On a workspace plan, an admin has to allow developer mode before you can see it.

Codex

Codex takes the connector as a streamable HTTP MCP server, from its CLI:

codex mcp add accesslint --url https://mcp.accesslint.com/mcp

Adding it starts the same browser sign-in, so pick your workspace there and approve. After that, ask Codex for a scan or a flow the way you would in a chat.

For a machine with nobody at the keyboard, skip the sign-in and hand it an API key instead:

export ACCESSLINT_API_KEY=alk_your_key
codex mcp add accesslint --url https://mcp.accesslint.com/mcp \
  --bearer-token-env-var ACCESSLINT_API_KEY

Checked against codex-cli 0.149.0. Run codex mcp add --help if your version disagrees.

Any other MCP client works the same way: it discovers where to sign in from the server itself, so the URL is the only thing you ever paste.

The workspace is fixed at sign-in

You choose a workspace on the approval screen, and every request from that connection uses it for good. To point the connector at a different workspace, remove the connection and add it again.

If you belong to two workspaces and want both, add the connector twice under different names.

What you can ask for

• Scan a page. “Check example.com/pricing for accessibility problems.” No domain setup needed.
• Add a site. You get back the exact DNS TXT record to add at your registrar. On a Google Workspace account, a domain your organization owns verifies immediately.
• Draft flows from journeys. “Monitor sign-up and checkout on example.com.” Each sentence becomes a draft flow with steps compiled from it.
• Dry-run a draft. The steps get walked in a real browser. If step 3 fails, ask for the screenshot and you’ll see the page the way the browser did.
• Fix a step, or rewrite the journey. Rewriting the description recompiles every step; editing one step costs nothing and leaves the rest alone.
• Read results. Open violations for a flow, newest run first, plus a link to the full report.

What the connector can’t do

• It can’t start monitoring. A draft monitors nothing until you approve its dry run in the AccessLint web app. No tool in the connector can approve one on your behalf.
• It can’t handle credentials. A step that would type a password, API key, or one-time code is refused. Sign-in is configured on the domain in the web app, not in a conversation.
• It can’t reach anything you can’t. The connector acts as you, in the one workspace you picked, under the same plan limits and quotas as the web app.

Findings are quotes, not instructions

A scan reports text and markup from the page it audited, and that text reaches the assistant as content. The connector labels it as data every time it hands it over, so an instruction hidden in someone’s page markup is reported rather than followed.

For scripts and agents

A client with nobody at the keyboard can skip the sign-in dance and send an account API key as the bearer token instead. Keys are minted on your account’s API keys page, and need the read, execute, and draft scopes for the full tool set.

Removing it

Remove the connector in your assistant’s settings, then open Connected apps in your AccessLint settings and revoke the connection. Both halves matter: removing it stops the conversation from reaching AccessLint, and revoking invalidates the token it was holding.

Revoking deletes nothing. Flows, domains, and scans stay exactly as they are, and the assistant can ask for access again whenever you want it back.

Questions: support@accesslint.com.