Use Sakana Namazu with a signed-in browser

Yes, you can use Sakana Namazu in a signed-in browser workflow, but Namazu does not open the browser itself. Call the hosted model through Sakana's API, and let a local agent host use Tabbit CLI to read authorized pages in your existing Tabbit Browser session. Keep the task read-only: list exact dashboard changes, link each source, and save a named draft for a person to verify.

Sakana's official model page identifies sakana-namazu-v1.0 as the model ID and sakana-namazu as its current alias. The API offers built-in web search and code execution, but those tools do not expose your signed-in browser pages. A Japanese API setup report on note.com describes a first PowerShell request after setting up credentials and prepaid credit. That is an API onboarding story, not evidence of browser access.

Choose the right route for the task

Sakana describes Namazu as a hosted, Japanese-specialized API, built on Kimi K2.6 and tuned for Japanese language and business workflows. Official docs list OpenAI Chat Completions and Responses APIs plus Anthropic Messages, with the same API base https://api.sakana.ai/v1. The current alias is convenient, while the versioned ID makes the selected checkpoint explicit.

The API is not a local model download. For a first request, use Sakana's documented OpenAI-compatible route and keep the key in an environment variable:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["SAKANA_API_KEY"],
    base_url="https://api.sakana.ai/v1",
)
response = client.chat.completions.create(
    model="sakana-namazu",
    messages=[{"role": "user", "content": "こんにちは！"}],
)
print(response.choices[0].message.content)

The official /v1/models endpoint lists supported IDs; check it before wiring a client. Namazu is pay-as-you-go, not included in a subscription. Sakana says built-in search and code execution incur separate tool charges, and each tool loop may add tool output to subsequent model input. Avoid enabling those tools when the task only needs the signed-in page.

Connect the API host to your signed-in browser

A local agent host has two separate jobs: call Namazu's remote API and invoke the browser tool. Tabbit Browser holds the site session; Tabbit CLI opens a task-owned tab in that environment. The client skill must be discovered, the CLI launcher must be reachable, and the page must be authorized. Verify each separately. See the Tabbit CLI guide, existing-browser connection notes, and first-task checklist.

1. Create an API key in Sakana's console and store it as SAKANA_API_KEY; do not paste it into a prompt or log.

2. Confirm the model ID with /v1/models, then send a harmless API request.

3. Sign in to the target dashboard yourself in Tabbit Browser.

4. In your local agent client, discover the Tabbit skill and separately test the local CLI launcher. Do not assume model selection installs either one.

5. Run a read-only smoke test before asking for real dashboard data. Use a separate task tab so your existing tabs remain available.

The following checks browser connectivity only; it does not call Namazu or prove an integration between a particular client and the model:

"$HOME/.local/bin/tabbit-cli" nodejs --task "Namazu browser smoke test" --request-id namazu-smoke-001 <<'JS'
await page.goto("https://example.com/", {waitUntil: "domcontentloaded"});
return await tabbit.observe({frames: "none", maxChars: 2000});
JS

For one task, such as reviewing a Japanese customer-support dashboard, ask for visible status changes and items that need verification. Keep the requested fields narrow. Do not ask the model to submit changes or infer intent from a label.

If an ID is duplicated, a login page appears, or history is not visible, report the limit and stop. For table extraction, see browser data to CSV; for source-backed research, use the browser research workflow.

Prepare a named draft, not a dashboard update

Ask the agent to create Namazu dashboard change review — CASE-EXAMPLE-104. The draft should list each exact record ID, the page URL, visible status text, timestamp, and any unknown value. Keep the source pages open in the task tab so a reviewer can compare the report with the dashboard before copying it to a ticket.

Use Sakana Namazu through the configured local agent and Sakana API. Use Tabbit CLI in my current Tabbit Browser environment. In a separate task tab, open [DASHBOARD_URL]. Read only records for [ACCOUNT_OR_CASE_ID]. Do not edit, submit, export, or change filters.

Return a Markdown draft named "Namazu dashboard change review — [ACCOUNT_OR_CASE_ID]". For each matching record, include the exact visible ID, status text, visible update time and timezone, page heading, and URL. Separate confirmed changes from fields that need verification. If the ID is missing, duplicated, or the page is not authorized, stop and report it.

Expected output: a concise, source-linked review list and “No write action taken.” This is an illustrative workflow, not a completed Namazu or production-site test.

The page content needed to answer is sent from the local host to Sakana's API. Do not use this workflow for secrets, unrelated personal data, or records your account is not authorized to view. For permissions and handoff boundaries, see browser MCP vs CLI, the Codex CLI setup, and browser permissions.

Sakana's get-started documentation and an independent WEEL API example both demonstrate Namazu as an API call; the WEEL example drafts a Japanese business email. A YouTube video titled “Sakana Namazu AI model for coding: Performance Review & Code Examples” is another public API demonstration, not a Tabbit browser test. Treat each as a setup or usage example, not a controlled quality result.

Common questions

The model's built-in tools and the browser connection have different access boundaries. In a local setup fixture, a protected page returned 401 before test sign-in; a second named Tabbit CLI task read Release notes review and GEO-104 after a short-lived test cookie was set. This supports session reuse in that fixture only. It does not show Namazu or any specific client operating the page end to end.

For a safer rollout, start with one non-sensitive record, compare its exact text and source, and review the named draft before any write. The MCP and CLI comparison and browser agent tutorials cover connection choices and other read-only tasks.

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.