Step 5 Preview can request tool calls, but it cannot operate a browser on its own. To use it with a signed-in browser, run it in an agent client that exposes local tools, then use Tabbit CLI as the browser executor. The client must actually offer Step 5 Preview and allow local tool use; model support alone does not connect a remote runner to your computer.
A useful first task is reconciling one record across two logged-in web pages. Ask the agent to use a stable ID, keep both source URLs, and return a field-by-field difference table. Tabbit CLI uses the current Tabbit Browser environment, so you can inspect the pages and continue reviewing in a separate task tab.
What Step 5 Preview provides
StepFun’s official cookbook documents the step-5-preview model ID and a tool-calling loop: the model chooses a function and arguments, the host validates and executes the function, then returns the result with its matching call ID. The application owns execution and stopping limits. This is general function calling; the documentation does not say the model has native browser control. StepFun’s tool-calling guide
StepFun’s repository also documents API serving and deployment requirements for local weights. A local model deployment is a separate infrastructure project; the existence of weights does not mean a standard laptop agent can run them. Official repository and deployment notes
| Question | What the source establishes | What you still need to check |
|---|---|---|
| Model identifier | StepFun cookbook uses step-5-preview | Whether your selected client/provider exposes it today |
| Tool calls | Model can request a declared function with arguments | Whether the host validates and runs the local tool |
| Browser access | Not established by model docs | Whether the agent runs locally and can launch Tabbit CLI |
| Login state | Not a model capability | Whether the intended account is already signed in inside Tabbit Browser |
Prepare the local browser path
Install Tabbit Browser, install Tabbit CLI from the official quick-start guide, and confirm the client you intend to use can select Step 5 Preview. Then confirm that client can load a local skill and run local commands. The CLI skill entry is /tabbit in clients that support named skills; Pi uses /skill:tabbit. Some clients expose a /skills list instead of a slash command. Skill discovery does not prove that the launcher is reachable, and launcher reachability does not prove the site is signed in.
Use the browser skill to discover the installed launcher syntax and its task lifecycle. The CLI is invoked as tabbit-cli nodejs --task ...; use the installed launcher path from the setup guide, and follow its current request-ID requirements. Do not paste passwords or session cookies into prompts.
Before asking the agent to touch a record, open both sites yourself in Tabbit Browser and confirm that the expected account and record are visible. An existing login may be reused by a CLI task operating in that Tabbit environment. A separate task tab gives you room to keep reading or reviewing while the agent works. That session behavior was checked in a local test fixture; it does not establish that a particular external model/client integration is available.
Reconcile one stable record across two pages
Start with a test or read-only record. Give the agent the identifier and the fields to compare, not a vague instruction such as “check what changed.” Keep a source URL beside each extracted value.
| Checkpoint | Pass condition | Stop and review when |
|---|---|---|
| Identity | Both pages show the same stable record ID | IDs differ, are hidden, or are ambiguous |
| Fields | Each requested field has a value or unknown | The page omits a field or formats it differently |
| Evidence | Every value points to the page it came from | The source URL or visible label is missing |
| Action boundary | Task only reads and compares | Any save, send, delete, or approval is proposed |
Give the agent this prompt, replacing the bracketed values. It should return a concise comparison and a named review checkpoint.
Task: compare record [RECORD_ID] on [PAGE_A_URL] and [PAGE_B_URL] using the signed-in pages already open in Tabbit Browser.
Use Tabbit CLI from this local agent to inspect the pages. Do not sign in, submit, save, send, delete, or change settings.
For each field [FIELD_LIST], return the value, source page URL, and whether the values match. Use "unknown" when a value is not visible; do not infer it.
First confirm the record ID on both pages. If either page shows a different ID, a login prompt, CAPTCHA, or an unexpected dialog, stop and report it.
Save no changes. Finish with a draft named "[RECORD_ID] — reconciliation review" and a visible checkpoint listing: matched fields, mismatches, unknowns, and the two URLs for human review.This is a suggested workflow prompt, not a report of a completed Step 5 Preview browser test. Review the returned record ID and each cited value in the actual pages before using the summary. If the page layout changed, ask the agent to re-observe rather than repeating a stale click. If access is blocked, stop and resolve the browser state yourself.
What Tabbit CLI adds
Tabbit Browser supplies the browser environment; Tabbit CLI lets compatible local agents work with it. The model proposes tool calls, the agent client decides whether and how to use them, and the CLI performs browser steps. The local session check for this assignment used two named tasks against a localhost fixture: the first received 401 before a test sign-in; a second task in the same Tabbit environment received 200 and read the fixture record. This only verifies session reuse in that test setup. It does not verify Step 5 Preview inside Codex, Claude Code, or any other client.
If your chosen client does not list Step 5 Preview, use a supported provider or do not select it for this workflow. If the agent is remote and cannot hand off to a local process, use a local agent for the browser step. For setup and troubleshooting, see the first-task checklist, existing-browser connection guide, CLI versus MCP guide, Codex setup, complete CLI guide, login reuse guide, browser prompt template, click troubleshooting, API/MCP/browser decision guide, and source-based research workflow.
What early community feedback does and does not show
A Reddit user’s Terminal-Bench comparison says Step 5 Preview was “very slow” in that benchmark context, while the author reports a lower token count than one comparator. That is a single user’s measurement on a coding benchmark, not a browser-use result. A YouTube reviewer titled a video “Step 5 Preview First Test” and included a “Browser OS Test” chapter; the accessible page exposed the video metadata but not its comments in this review. Neither source proves that Step 5 Preview can control a real signed-in website. Reddit discussion · YouTube test video
Verdict
Use Step 5 Preview for signed-in browser work only when your chosen client exposes the model and can run local tools. Pair it with Tabbit Browser and Tabbit CLI, keep the task read-only at first, and verify every field in the source pages. The model’s documented function calling is a useful building block; the browser connection and login context come from the client and browser setup, not from the model itself.
FAQ
Can Step 5 Preview control a browser by itself?
No. The model can return tool calls, but the host application must expose tools and execute them. A local agent with Tabbit CLI can provide a browser execution path when the model is available in that client.
Does the model reuse my website login?
The model does not hold a browser login. Tabbit CLI runs against the Tabbit Browser environment, where an existing site session may be available; confirm the account and page in the browser before acting.
How do I verify a two-site comparison?
Use a stable record ID, preserve each source URL, and compare exact fields. Ask the agent to return missing fields as unknown and verify both pages before it summarizes.
Can I use a cloud agent with Tabbit CLI?
Only if that agent has a supported local handoff. A remote runner cannot reach a local browser merely because its model supports tool calls.
What should I do if the page changes or blocks access?
Stop before submitting anything, re-read the page, and ask the local agent to report the new state. Sign-in prompts, CAPTCHA, and permission dialogs need your direct review.