Claude Code

You need an API key (Quickstart) and Claude Code. These steps were checked with Claude Code 2.1.

1. Set the address and the keyCopy Anchor Link

export API_URL=https://api.us.webagent.lookover.ai
export WEBAGENT_API_KEY=wa_us_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

Add this line to your shell profile so it is set every time you start claude.

2. Add the serverCopy Anchor Link

claude mcp add --transport http lookover-webagent \
  "$API_URL/mcp" \
  --header 'Authorization: Bearer ${WEBAGENT_API_KEY}'

The shell fills in $API_URL when you run the command, so Claude Code saves the full address.

Use single quotes around the header. The shell then does not expand ${WEBAGENT_API_KEY}. Claude Code saves the text ${WEBAGENT_API_KEY} and reads your environment when it starts the server, so the key itself is never written to a config file. If the variable is not set when claude starts, claude mcp list warns Missing environment variables: WEBAGENT_API_KEY and the connection fails with 401.

Things to know:

• Do not name the server webagent. Claude Code refuses it (this name is reserved). Any other name works. This page uses lookover-webagent.
• --transport http is the streamable HTTP transport. The webagent endpoint is POST /mcp and does not use SSE.
• The default scope is local: the server is available in the current project only and is not shared. Add --scope user to use it in every project. Use --scope project to write it to .mcp.json so your team shares the same line. The key is not in that file, only ${WEBAGENT_API_KEY}, so each person uses their own key.
• Webagent also accepts the key in an x-api-key header, if you cannot send Authorization.

3. Check the connectionCopy Anchor Link

claude mcp list

You want a line like this, ending in Connected:

lookover-webagent: https://api.us.webagent.lookover.ai/mcp (HTTP) - ✔ Connected

Inside Claude Code, type /mcp to see the server, its status and its tool count.

The server should offer the 15 fixed tools (create_session, run_task, get_task, smoke_test, get_report, list_sessions, get_session, stop_session, scrape, screenshot, pdf, crawl, list_skills, run_skill, get_skill_run), plus one skill_<name> tool for each skill your workspace can run, such as skill_smoke_uat. Skills appear only if they are installed in your environment. If you see only 15 tools and no skill_ tools, see Skills.

If it fails to connect:

4. Let Claude call the toolsCopy Anchor Link

By default, Claude Code asks before each tool call. To approve every webagent tool without a prompt, add an allow rule to .claude/settings.json:

{
  "permissions": {
    "allow": ["mcp__lookover-webagent__*"]
  }
}

Only do this for a project where you are happy for the agent to open any public URL you name in your prompt.

5. Run a first smoke testCopy Anchor Link

First install the agent setup snippet in your CLAUDE.md. Then try the Test Arena, a fake site meant for this:

Use the webagent tools. Smoke test https://webagent-test-web.pages.dev/app/login:
sign in with the email [email protected] and the password arena-pass. After signing in,
the text "Welcome" must be on the page. Give me the verdict, the failed steps, the
console errors and the report link.

Claude calls smoke_test once. It turns the sign-in into steps and "Welcome" into expectText. The tool opens a recorded browser, runs the steps and the checks, and waits up to 60 seconds. You can watch it in the dashboard under Sessions while it runs.

Say what must be on the page after the steps ("the text ... must be on the page"). If you only ask Claude to "check that the page shows" something, it can add that as a last step instead of an expectation. A step that runs while the page is still loading can then stop with low_confidence and fail the test, although the app is fine.

Claude Code 2.1.322 gives the model the whole report (structuredContent) as the tool result. Its summary field has the same sentences as the tool's text:

Smoke test of https://webagent-test-web.pages.dev/app/login PASSED in 1 s: 13 of 15 steps passed. No console errors or failed network calls.

The report has a verdict (pass, fail or error), the steps, console errors, failed network calls and links to the proof. If the test is not done after the wait, the verdict is running and Claude calls get_report with the id until it ends. That is normal, not an error. A test that ran and failed is a normal result: isError is set only when the test could not run. See Reading results.

For an open-ended goal, where the agent decides what to do, ask for a task instead. Claude then calls run_task:

Use the webagent run_task tool on https://webagent-test-web.pages.dev/: open the
Pricing page and tell me its heading.

A task has no assertions: ok means the agent says it is done. Claude can call get_report with the task id for a report with the console errors and failed network calls.

Recipes has more prompts. Reading results explains what each field means.

Remove the serverCopy Anchor Link

claude mcp remove lookover-webagent

If you added it with --scope user or --scope project, add the same --scope flag.