Command A+ can help prepare a document draft from a signed-in team site when an agent client can call the exact Cohere model and execute local tools. Select command-a-plus-05-2026, install Tabbit Browser, sign in to the intended workspace, and connect the local agent through Tabbit CLI. Then ask for a source-linked draft and stop before sharing or editing the source document. Cohere documents tool use and the model’s API ID, but it does not document a built-in Tabbit integration.
Cohere’s model documentation gives the exact API ID and lists tool use; its tool-use guide explains that the calling application executes returned tools. Cohere’s release note confirms Apache 2.0 weights and standard API availability.
That boundary matters. In a Reddit report, one Pi user said the model “stopped after tool calls” through an OpenAI-compatible API. That is one setup, not a benchmark. It is a good reason to verify the entire tool loop before you point an agent at a real team document.
Key takeaways
Cohere’s exact API model ID is
command-a-plus-05-2026; the hosted model page lists tool use as a capability.A model, an agent client, the browser connector, and a human reviewer are separate parts of this workflow.
Cohere API is the clearest documented hosted route. OpenRouter also lists
cohere/command-a-plus; confirm that route in your client before relying on it.Use Tabbit CLI from a local agent to read the existing signed-in browser session. Do not assume a remote model runner can reach your computer.
Save the result as a named draft, check every source link, and keep the agent away from publish controls.
The four parts of the workflow
| Part | What it does | What to verify |
|---|---|---|
| Model | Generates a response or requests a tool | Exact ID: command-a-plus-05-2026; tool call appears in the API response |
| Agent client | Sends requests to Cohere and runs approved local tools | It supports this provider and can invoke local tools in the same task |
| Tabbit CLI | Lets the local agent work with Tabbit Browser tabs | Skill is discovered and tabbit-cli diagnose sees the browser |
| Reviewer | Checks source links and the draft | The source document is unchanged; edits are saved as a separate draft |
A hosted model call and a local browser connection are different checks. Cohere’s model docs list text and image input plus tool use, and its API docs explain the tool round trip. The client still has to turn a tool request into a local action and send the result back. If the client only returns plain text, it cannot operate the browser through that response alone.
Confirm the route before opening team material
Cohere’s current model page names the API ID command-a-plus-05-2026. Cohere also offers Model Vault for production use. Its deployment reference lists one B200 or two H100 GPUs for the W4A4 configuration; do not read “25B active” as a desktop hardware estimate. For a hosted workflow, start with Cohere’s documented API. OpenRouter lists a cohere/command-a-plus route as well, but provider availability and tool behavior can change.
In your local client, select the exact provider and model, then run a non-browser tool test. Ask it to return one fixed value through a harmless function. Inspect whether the client received a structured tool call, executed the function, and passed the result back to Command A+. If the loop ends after the call, stop there. The community report above describes that exact failure shape.
For the browser side, follow the Tabbit CLI quick start. Supported clients may expose the skill as /tabbit; Pi uses /skill:tabbit, and other clients can use /skills or a skill menu. Read the installed skill instructions instead of guessing. Run tabbit-cli diagnose, open the target site in Tabbit Browser yourself, and confirm the account and workspace. A client’s skill list, the CLI launcher, and the website login are separate gates.
| Preflight check | Pass condition | Stop condition |
|---|---|---|
| Model route | The client shows Cohere command-a-plus-05-2026 or a deliberately selected current provider route | A similarly named model is selected silently |
| Tool loop | A harmless call returns structured arguments and a result | The call appears as prose, hangs, or loses its result |
| Skill and launcher | The client finds Tabbit’s skill and diagnose sees the browser | The client is remote-only or cannot invoke local tools |
| Website scope | You verify the intended signed-in account and page | Wrong account, unclear access, or unexpected write controls |
Turn a team page into a reviewable draft
Choose one document or one small set of approved pages. Ask the agent to extract only fields relevant to the brief, attach a source URL to each claim, and label anything it cannot verify. It should create a new draft rather than edit the source page. Keep the draft in a separate task tab so the user can continue work in the main window. Tabbit’s browser connection guide and first-task checklist cover that separation.
Use this prompt as a starting point:
Use Tabbit CLI in this local task. Read only the signed-in team document at [approved URL]. Do not edit, share, publish, or submit anything. Extract the decisions, owner, due date, and open questions. For every item, include the source URL and a short exact excerpt. Mark missing or conflicting fields as needs_review. Create a separate draft named "Q4 launch notes, 2026-11-10" with a source column and a reviewer column, then stop. Show me the draft location and any fields you could not verify.The title and date are examples. Replace the URL and document name with a page the user is allowed to read. Review each citation in the browser, compare the draft with its source, and remove sensitive material that does not belong in the note. Do not let the model infer missing owners or deadlines.
A useful draft has one row per decision, not a polished summary that hides where claims came from. If two pages disagree, preserve both links and flag the conflict. The browser MCP versus CLI guide explains when a command-line connection fits better than a protocol server. For setup details, see the Tabbit CLI guide, Codex setup example, Tabbit Browser overview, browser best practices, and agentic browser explainer.
When this workflow fits
Command A+ is a reasonable candidate when your client can use Cohere’s current endpoint and run local tools, and you want a model to organize source material into a review copy. It is a poor fit if your client cannot parse tool responses, if your deployment cannot meet the chosen model’s hardware needs, or if the work requires unreviewed edits. A hosted model does not receive a browser session automatically.
Cohere publishes Command A+ weights under Apache 2.0, and its hosted service has separate usage terms. Choose the route that matches your data and deployment requirements. For ordinary team drafting, the safe first move is a read-only pass, a named draft, and a human review. Download Tabbit Browser if you need a local browser for that workflow.
FAQ
What exact Command A+ model ID should I select?
Cohere documents `command-a-plus-05-2026` for its API. OpenRouter currently lists the route as `cohere/command-a-plus`. Check the provider and full ID in your agent client before using a task.
Can Command A+ open a signed-in website by itself?
No. The model can request tools, but a local agent must execute those tools. Tabbit CLI can connect that local agent to Tabbit Browser when the client discovers the installed skill and can launch the CLI.
Is Command A+ open source?
Cohere publishes Command A+ weights under Apache 2.0. Hosted Cohere API use follows the terms for that service, which are separate from the weight license.
Can I run Command A+ on a laptop?
Do not infer hardware needs from its 25B active parameters alone. Cohere’s published W4A4 deployment reference is one B200 or two H100 GPUs. Hosted inference is a more practical starting point for many local-browser workflows.
What should I do if a tool call hangs?
Stop the task and inspect the selected model ID, provider, thinking settings, tool schema, client adapter, and returned tool-call result. Repeat with a harmless read-only page before granting write access.
Can a hosted Command A+ endpoint use my local Tabbit browser?
Only when the agent running locally can call the hosted endpoint and execute local tools in the same task. A remote runner needs a separately verified local handoff; model availability alone does not connect it to your browser.