TabbitBlog

Use Claude Opus 5 with a signed-in browser

Verify the exact Opus route, inspect a support ticket with Tabbit CLI, and leave a draft for human review.

In this article
  1. Key takeaways
  2. Choose the inference route
  3. Turn a support ticket into a checklist
  4. Troubleshoot by symptom

Claude Opus 5 can work with a signed-in site inside a local agent that has browser tools. The API model ID is claude-opus-5. Confirm your provider still offers it, then let the agent call Tabbit CLI. API configuration alone does not grant browser access. Anthropic announced Opus 5 on July 24, 2026. Its API docs use claude-opus-5; Opus 5.5 is a later release and a separate model choice. Use the exact version for reproducible work. Check the launch post, model lifecycle page, and Opus 5.5 announcement.

A Reddit user argued that agent actions involving spend need human checkpoints. It is one user’s view, but useful for ticket work. Post.

Key takeaways

  • Model hosting and browser access are separate layers; a local agent calls Tabbit CLI for page actions.

  • Check the exact model and runtime, then verify the account in the selected Tabbit profile.

  • Leave a reviewable draft; do not let the agent submit or change the site.

Choose the inference route

RouteUse whenVerify
Anthropic APIDirect model access in a local agentModel ID, account access, API terms
Claude CodeLocal agent workflowSelected model and Tabbit skill availability
Another local agentYou already use a tool stackProvider support and local CLI permissions

Discover the Tabbit skill through your client’s supported method. Pi uses /skill:tabbit; some clients list skills through /skills. Discovery does not prove the CLI can run: check the skill, local tabbit-cli diagnose, and model host separately.

Turn a support ticket into a checklist

Use a synthetic or approved low-risk ticket. Confirm the signed-in workspace, then read only its visible number, status, and linked policy title. Return each item with a source URL and mark missing evidence. Save a named local draft; do not send a reply or change status. A Reddit user discussing Opus 5 argued for deterministic human checkpoints before spending actions. Treat that as one person’s experience, but the boundary is useful here too.

An X user described repeated prompts for a browser-based Three.js scene with Opus 5, not a signed-in business workflow. Post.

Task name: Support ticket review — fixture Q-104
Use the installed Tabbit skill from this local agent and Claude Opus 5 (model ID claude-opus-5). In a new task tab, open the ticket URL I provide. Read only the visible number, status, and linked policy title. Return a checklist with the exact source URL for each item; mark missing values “not visible”. Save a local draft named “Q-104 support review”. Do not reply, change status, or close the ticket. Wait for my review.

A task tab keeps this job separate from your existing browser work. If the site is signed out, sign in yourself and verify the account. Do not put passwords or private records in the prompt.

Troubleshoot by symptom

CheckpointEvidenceStop if
ModelProvider and exact IDIt routes to another version
BrowserVisible account and final URLLogin or workspace is wrong
TicketNumber, status, sourceA value is inferred or absent
ActionNamed draft onlyThe agent reaches reply or close

For setup, read the Tabbit CLI guide, connect an agent to an existing browser, and the first-task checklist. Compare browser MCP and CLI, review profiles and workspaces, follow the CMS draft workflow, see what an agentic browser is, and use the Codex setup example.

Sources checked: Opus 5 announcement, model lifecycle docs, Opus 5.5 announcement, and Claude Code CLI docs.

FAQ

Is Claude Opus 5 available by API?

Anthropic lists claude-opus-5 in its API documentation. Verify your account and the current lifecycle page before relying on access.

Is Opus 5.5 the same model?

No. It is a later release with a separate model ID. Do not substitute it in a reproducible task without approval.

Does Claude for Chrome connect Tabbit CLI?

They are separate integrations. Check that your local agent can discover the Tabbit skill and run the CLI.

Can the task reuse my browser login?

It uses the selected Tabbit environment, which may have a valid login. Verify the account and page yourself.

Should Opus send the support reply?

No. Ask it for a named draft; a person should send replies or change ticket status.

Take the next step

Let Tabbit work alongside you.

Research across tabs, automate repetitive browser work, and keep every piece of context within reach.