Pareto can synthesize source material, but it does not control your signed-in browser. Use a local agent with the Unbiased pareto API route, connect that agent to Tabbit Browser through Tabbit CLI, then save a named document draft with a source URL beside each claim. Unbiased documents OpenAI and Anthropic API formats and an OpenCode setup; that establishes model access routes, not a native Tabbit integration. Check Unbiased's API guide.
Key takeaways
Pareto is Unbiased's blended API model. Its documented API model string is
pareto; provider catalogs may use a qualified ID such asunbiased/pareto.Let the browser agent gather permitted source passages and URLs. Ask Pareto to organize them, not invent missing citations.
Create "Weekly research brief, draft 01" in the target document app and stop before sharing or publishing it.
Keep source URLs, evidence snippets, and unresolved claims visible so a reviewer can reopen each page.
Pick where each task runs
| Task | Run it in | Required input | Do not assume |
|---|---|---|---|
| Read a public announcement | Local agent or model client | Public URL | Public access proves a private workspace is available |
| Read a signed-in team page | Tabbit Browser task via local CLI | Site login in that browser profile | The model API can see browser cookies |
| Summarize and group findings | Pareto through a supported local client | Source excerpts and URLs | A fluent answer has verified citations |
| Save a collaboration draft | Authorized document app | Named draft and review owner | Drafting grants permission to share or publish |
Unbiased's developer docs name pareto as the API model string and document OpenAI and Anthropic-compatible formats. Their OpenCode setup shows the custom provider route. New Unbiased platform accounts require review before an API key is issued, according to the same docs. Verify account access and your client's tool support before starting. This article does not claim to have run Pareto inside OpenCode or tested a Pareto-to-Tabbit integration.
Keep the evidence attached to the draft
For team notes, vendor comparisons, or a weekly policy roundup, ask the browser task to read only pages you are allowed to access. Pair each extracted statement with its exact source URL. Send Pareto the small evidence set, with instructions to label inference and gaps. This makes the output reviewable even if the model blends multiple answers behind a single response.
The result should be a draft, not an update to the underlying source pages. If a team page requires a login, use the Tabbit environment where you signed in yourself. If access is missing, preserve the page URL and report the permission message rather than finding another account.
Prompt for a source-linked draft
This is an illustrative workflow, not a recorded integration test. The prompt asks for a concrete review document and leaves the source material untouched.
Use the installed Tabbit skill from this local agent to read only the team pages I am authorized to access. Keep each page URL with the exact excerpt that supports a claim. Do not edit any source page.
Send the evidence bundle to Pareto. Create a document named "Weekly research brief, draft 01" with columns: claim, source URL, supporting excerpt, confidence (direct / partial / unresolved), reviewer note.
Group related claims under short headings. Do not add facts without a source. Mark conflicting or missing evidence as unresolved. Save the document as a draft, show me the named draft, and stop before sharing, submitting, or publishing.Expected output: a short document with a URL and excerpt for each claim, plus an unresolved section. A reviewer can open each source and compare the quoted text. Do not pass passwords, cookies, private keys, or unrelated personal data to the model.
Set up Tabbit Browser for the local reading step
Install Tabbit Browser and sign in to the team site yourself. Use a local agent that can discover ~/.agents/skills/tabbit/SKILL.md and run ~/.local/bin/tabbit-cli. OpenCode documents @tabbit for a discovered skill. Other clients may use a skill panel or a different syntax. Do not invent a /tabbit command for a client that does not list it.
Check the model route, skill discovery, launcher reachability, site login, and document write permission separately. A local Tabbit CLI test in this environment confirmed a second named task could reuse the test browser's short-lived signed-in state in its own task tab. It did not test Pareto or a document integration. Keep a separate task tab for source reading while you continue in the rest of Tabbit. Read the Tabbit CLI guide, agent connection steps, first-task checklist, and browser MCP vs CLI guide before granting more access.
Review before sharing
| Draft check | Pass condition | Reviewer action |
|---|---|---|
| Source traceability | Every factual sentence has a direct URL and supporting excerpt | Open the source and compare it |
| Conflicts | Different pages or dates are shown as separate claims | Decide which source governs; do not let the model merge them silently |
| Missing evidence | Unverified details are labeled unresolved | Find a source or delete the claim |
| Document state | Draft is visible and not shared or published | Approve the exact document and audience before sharing |
If you need a repeatable local handoff, see how to connect an agent to an existing browser, reuse a signed-in session, organize browser tabs and workspaces, and the browser-agent security guide. These guides cover browser access and review controls; they do not establish native Pareto support.
FAQ
Can Pareto edit a shared document in my signed-in browser?
Pareto is an API model, not the browser controller. A local agent must read authorized pages through a browser tool, send the relevant evidence to Pareto, and create a draft in a document tool. Keep source pages read-only until a person approves edits.
What is Pareto's API model string?
Unbiased's developer docs use `pareto` with its API. The OpenRouter catalog route is `unbiased/pareto`. Confirm which provider and model string your selected agent client supports.
Does Unbiased require account approval?
Its API quick start says new platform accounts are reviewed before an API key is issued. Check current access before writing a client configuration.
Can Pareto cite the pages it reads?
Only if the local agent keeps each source URL with the extracted claim and passes that evidence into the draft. Ask for one URL per claim and mark unsupported statements as unresolved.
Can a remote agent use my local Tabbit login?
Not automatically. The remote agent needs a documented local handoff. Otherwise, run the browser reading task locally and give the model a minimized, approved source bundle.
Should the agent publish its document?
No. Have it create a named draft and stop at a visible review checkpoint. A person should inspect the sources and approve any shared edit or publication.