Gemma 4 E2B IT can help turn a few signed-in team pages into a source-linked document draft, but it cannot read those pages by itself. Run the exact google/gemma-4-E2B-it checkpoint in a runtime that exposes structured tool calls to a local agent client, then use Tabbit CLI to open the approved pages in Tabbit Browser. Ask for a short evidence ledger before asking for prose, and save the result as a named draft such as “Weekly release update — Gemma E2B draft.” Stop before sharing or publishing. Google's function-calling guide makes the key boundary clear: the model requests a function, while the host application validates and executes it.
The useful part of this workflow is not autonomous writing. It is keeping every sentence tied to a page a teammate can reopen. A small model can help organize a narrow source set, but you still need to check that the runtime passes tool calls correctly and that each source belongs to the right workspace, date, and project. For the browser-session boundary, see connecting an agent to an existing browser. For a reusable task structure, see the browser-agent prompt template.
Key takeaways
The official E2B IT checkpoint is google/gemma-4-E2B-it; quantized builds may use different repository names.
The model emits structured calls. A runtime and client execute the page-reading tools and return results.
Read only pages you are authorized to use. Keep a source URL beside every material claim.
Separate collected evidence from the draft so you can spot missing, outdated, or conflicting support.
Save under a clear draft name and stop for a human review before sharing or publishing.
Build an evidence ledger before writing prose
Choose two to four pages you are allowed to read, such as a release note, a project decision, and an internal status page. Give the agent their URLs and the date range that matters. Ask it to record short observations rather than copy whole documents. The ledger should identify what the page directly says, not what the model infers.
| Evidence field | What to capture | Why it helps the reviewer |
|---|---|---|
| Source key | A short stable label such as SRC-1 | Lets each draft sentence point back to one entry |
| Page title and URL | Visible title plus the original page URL | Makes the page easy to reopen in the right workspace |
| Observation | A concise claim or status shown on the page | Separates evidence from later paraphrase |
| Date or section | Visible update date, heading, or record identifier | Helps catch stale or similarly named material |
| Confidence note | Directly stated, inferred, or unclear | Keeps interpretation from looking like source text |
Only after the ledger is complete should the agent organize a short document. Ask it to cite source keys inline and include a final “Needs review” section. If a source is inaccessible, it should report that fact and leave the claim out. If two pages disagree, the draft should show both dates and flag the conflict instead of choosing one silently.
Check E2B and the tool path first
Google's E2B model card identifies the exact instruction checkpoint and lists Apache 2.0. Google's Gemma 4 model card describes E2B as 2.3B effective parameters, 5.1B including embeddings, with a 128K context window. The official release announcement states that Gemma 4 is available under Apache 2.0. License and model-card availability do not prove that a hosted provider currently serves the exact ID.
Google's function-calling guide includes google/gemma-4-E2B-it as an example and demonstrates the application-side loop: receive a structured call, check the function and arguments, run it, then pass back the result. The vLLM Gemma 4 recipe documents the E2B serving ID and Gemma 4 parser/template configuration. Pin the serving version and confirm the exact endpoint before selecting it in an agent client. Do not assume that every endpoint with an OpenAI-compatible label implements the same tool-call behavior.
Pi can use custom compatible model endpoints configured in models.json; its documented Tabbit skill command is /skill:tabbit. Treat this as a possible local client path, not a verified Gemma E2B integration. Check the local server's model list, make one harmless tool call, and confirm the client returns the tool result to the model. The Pi custom-provider guide, model configuration reference, and Pi Skills guide explain the configuration and discovery sides.
Connect a local agent to the signed-in workspace
Install Tabbit Browser and its local Tabbit CLI skill. In Pi, enter /skill:tabbit; another client may expose the same skill through /skills or a picker. Sign in to the team workspace yourself in the intended Tabbit profile. Then check that the client can discover its skill, that CLI can reach the running Tabbit browser, and that the target pages still show the expected workspace. These are separate checks. Follow the Tabbit CLI quick start to install and verify the local connection.
The browser task should stay small. Ask the agent to open only the approved URLs, capture their visible titles and source links, and write a draft in the correct destination. A task tab is separate from your other Tabbit tabs, so you can continue researching and return to review the work. The first-task checklist helps define the boundary; the Tabbit CLI guide describes task tabs and the browser-agent research guide covers source handling.
Draft a source-linked team document
Use a prompt that names the allowed sources, the document, the output sections, and the stop point. Keep placeholders in the prompt rather than exposing team URLs in reusable public examples.
Task name: team-source-brief-draft
Use only the team pages listed below. Do not search unrelated pages.
Read-only source URLs:
- [approved release note URL]
- [approved project decision URL]
- [approved status page URL]
Before drafting, make a source ledger with:
source_key, page_title, source_url, visible_date_or_section,
short_observation, direct_or_inferred, unresolved_question.
Then create a document titled "Weekly release update — Gemma E2B draft" with:
1. Summary (maximum 5 bullets)
2. Changes confirmed by sources
3. Open questions and conflicting dates
4. Source list linking every source_key to its original page
Do not invent owners, dates, status, or approvals.
Keep claims without a source out of the summary.
Save as an unshared draft and stop before inviting, approving, or publishing.
Return the draft title and the source URLs you used for human review.The placeholders are examples; replace them only with pages the user is authorized to access. A “shared document” is still a write to another service. Use this workflow only when creating a draft is authorized, and stop before changing sharing settings. If the editor cannot save a private draft, produce the ledger and proposed text locally for review rather than bypassing its permissions.
Review the draft against the source pages
| Draft situation | Agent response | Human review |
|---|---|---|
| Source directly confirms a fact | Add the source key beside the sentence | Reopen the page and compare wording and date |
| Claim is an inference | Label it as an interpretation | Decide whether the inference belongs in the document |
| Pages disagree | List both sources and dates under open questions | Ask the document owner which source is current |
| Page is inaccessible or login expired | Mark it unavailable; do not guess | Restore access yourself or remove the claim |
| Claim has no source URL | Omit it from the summary | Add evidence manually before approval |
Do not treat a model-generated citation as proof. Open each link and confirm that it supports the sentence, belongs to the intended workspace, and is current. Check that the document title ends in “draft,” that no unrelated page was read, and that the sharing state has not changed. Leave the final decision to the document owner.
One Reddit user described using Gemma 4 E2B with an OpenAI-compatible Ollama endpoint to divide a machine-check task into a researcher and summarizer, run tools, and write a short Markdown report. The same author said the final prose was weaker than a larger model's and that the framework's tolerant JSON parsing helped. Another user testing E2B and E4B for calendar-like tasks said tool calls were sometimes missed even with safeguards. These are individual local setups, not evidence that E2B can safely draft in your team workspace. Read the E2B coordinator report and the small-model tool-calling discussion.
For a repeatable draft, keep the source set small, put URLs beside observations, and leave the document visibly unapproved. The source-linked research guide and CMS draft workflow offer adjacent patterns. Use organizing multi-model research only when another model adds a clear review step. If multiple agents share one browser, see one browser for multiple agents. For connection details, compare MCP and CLI, existing-browser setup, and local versus cloud browser agents.
When this workflow fits
E2B is worth trying for a short source bundle and a document with a fixed outline, provided your runtime and client pass a harmless tool test. It may be a poor fit when the workspace has many ambiguous documents, when the evidence is sensitive, or when the task depends on long chains of actions. Use a more capable or more reliable setup for high-impact records.
The useful handoff is a named draft plus its source ledger, not an automatic approval. Read each original page, resolve disagreements, correct the prose, and approve or share the document yourself.
FAQ
Can Gemma 4 E2B IT read my signed-in team documents by itself?
No. The model needs a local agent client that can run browser tools. Tabbit CLI connects a supported client to Tabbit Browser, where the selected profile may already be signed in to the team workspace.
What is the exact Gemma 4 E2B IT checkpoint ID?
Google's model card and Hugging Face repository identify it as google/gemma-4-E2B-it. A runtime may expose a quantized derivative under a different name, so record the exact file and serving ID too.
Does Gemma 4 E2B IT support tool calling?
Google documents structured function calling for Gemma 4 and includes E2B in its example model IDs. The application or serving stack still has to parse the request, validate it, execute the tool, and return its result.
How do I keep sources attached to a team-document draft?
Ask the agent to maintain a source ledger with a short observation, the original page URL, and a date or section label for each claim. Keep unresolved differences explicit, then reopen each source during human review.
Can I use Pi with Tabbit CLI?
Pi supports custom compatible model endpoints and documents the /skill:tabbit command. Verify that your local endpoint returns tool calls the selected client can execute; this is separate from Tabbit skill discovery and browser reachability.
Should the agent share or publish the document after drafting it?
No. Save it with a clear draft name and stop before sharing, approving, or publishing. A person should confirm the sources, account permissions, wording, and intended audience first.