How to use the original Fugu Ultra with a signed-in browser

To use the original Fugu Ultra with a signed-in website, pin Sakana’s June 2026 version in a client that can also run local tools, then use Tabbit CLI to inspect only the authorized workspace page. Keep public web research and signed-in evidence in separate columns, cite both, and stop at a named draft for human review. This is a setup guide, not a claim that Fugu Ultra has a native Tabbit integration.

A Reddit reviewer who tested the Ultra and standard Fugu said they had “mixed results with Ultra” and preferred standard Fugu for their coding use because Ultra felt slower and more expensive. That is one person’s account, not a universal result. It is a useful reminder to reserve a higher-cost orchestration route for a task where depth matters. Read the original discussion.

Key takeaways

• The original Ultra is a versioned route: Sakana lists fugu-ultra-v1.0, also called fugu-ultra-20260615.

• The unversioned fugu-ultra alias currently resolves to v2.0. A provider catalog entry such as sakana/fugu-ultra may also be an alias, so verify the exact route before using it.

• Fugu’s model API and browser access are separate. The client must support the chosen Sakana model and local CLI tools.

• Tabbit CLI can work in the selected Tabbit browser profile. It does not discover a client skill or make a remote agent local by itself.

• Keep the brief read-only, source-linked, and open for a person to review.

Check the version before choosing a browser workflow

Sakana’s model catalog and getting-started guide list the versioned Ultra IDs and the current alias behavior. The model is offered through Sakana’s API; a third-party provider may have a different catalog or may not expose the pinned version. Check that before building a workflow. Do not silently substitute v2.0 when the task is specifically about the original release.

This distinction also applies to the client. Sakana documents a codex-fugu launcher and API setup, but that only addresses model access. The local agent still needs the Tabbit skill and permission to run the browser CLI. In Pi, the skill entry is /skill:tabbit; clients with a /skills picker should use that picker and confirm the Tabbit instructions are loaded. Never assume a /tabbit command exists. See the Tabbit CLI setup guide and how to connect an agent to an existing browser.

Separate public research from signed-in workspace evidence

Use the model’s documented API route to reason over the public material your client can retrieve. Use Tabbit CLI only for the private workspace page that you are allowed to access. They are separate inputs: a model name does not grant browser access, and a browser session does not prove that the model endpoint is available.

For example, an authorized internal launch note may describe an older release while Sakana’s public catalog names a newer alias. The brief should preserve both facts and explain which claim each source supports. It should not infer that the private note is obsolete or that the public alias is pinned. For a broader primer on these boundaries, see web search versus browser control and what an agentic browser can do.

Set up a local Tabbit task without moving the login

Before opening a site, run the installed launcher’s tabbit-cli diagnose check and confirm the local controller is running. This verifies CLI reachability; it does not discover the model or skill and does not sign in to the website.

Install Tabbit Browser and the Tabbit skill in a local agent client that can call the selected Sakana endpoint. Follow the official Sakana setup for the API or its supported launcher, and the Tabbit CLI quick start for browser access. A remote runner cannot see a local profile unless its client has a separately supported local handoff.

Before asking the agent to read a page:

1. Check that the Sakana model appears in the client’s model list. For the original Ultra, verify the pinned ID, not only the alias.

2. Open the target site yourself in the Tabbit profile you intend to use and confirm the login is active.

3. Load the Tabbit skill from the client’s documented skill picker. Its instructions should tell the agent to use the installed CLI.

4. Ask the agent to open a separate task tab and read one harmless, authorized page. Keep the profile’s other tabs untouched.

5. Confirm the task tab shows the expected account and page before requesting any extraction.

The local fixture recorded for this workflow returned 401 before a test session was established and 200 after a second named task reused that synthetic session. That verifies session reuse in the local test environment only. It does not test Fugu Ultra, an external agent host, or a real third-party account. For first-run checks, use the browser-agent checklist; choose between CLI and MCP using the MCP-versus-CLI guide.

Ask for a cited brief and leave it open

Use a read-only request with a clear output name and an explicit stop condition. For example:

Use Fugu Ultra fugu-ultra-v1.0 if this client and provider expose that exact model ID. Compare the official public release documentation with the authorized signed-in workspace note I specify. In Tabbit, open a separate task tab and read only that page. Create a draft named “Sakana release evidence brief” with columns for claim, public source URL, workspace page and section, agreement or conflict, and confidence. Do not fill, edit, submit, or publish any website form. Leave unknown claims marked for review and show me the draft before making changes.

The expected result is a short table, not an untraceable summary. Each public claim should link to its original page; each workspace claim should identify the page and visible section. If the two sources disagree, show both versions. A second Reddit commenter described their own Fugu Ultra experience as slower and more expensive than standard Fugu for a coding task. That is a subjective trade-off, not evidence about browser performance or your account. See the original thread.

Before sharing the brief, verify the model route, open every source link, and confirm that the private source is from the intended account. Keep identifiers and personal details out unless they are necessary. For safer boundaries, use the browser-agent security checklist as a final review.

When this route fits

Use pinned original Ultra when the task specifically requires the June 2026 release and your provider still offers it. If you only need Sakana’s current Ultra, the alias may be simpler, but it currently points to v2.0. Use standard Fugu for a lighter task when its route is available and sufficient. Neither choice replaces a human check of private evidence.

Tabbit is useful when the second source is behind a login already held in a local browser profile. It is not needed for public-only research, and it cannot make an unavailable model endpoint appear in a client. Start with a one-page, read-only brief, inspect the named draft, then decide whether to expand the task.

Tabbit CLI guide · existing-browser connection · first-task checklist · CLI or MCP · web search or browser control · agentic browser basics · Codex setup · Playwright MCP setup

Secholar.L

Growth & Content at Tabbit

Secholar.L writes hands-on guides about AI browsers and agent workflows, comparing tools with real research and citing verifiable sources.