North Mini Code can help draft a team brief from public references and pages you already have permission to read, but it cannot browse by itself. Cohere documents the model on Chat V2 and publishes open weights; an agent client must still connect the model to a browser tool. A practical local setup to evaluate is Tabbit Browser with Tabbit CLI from a client that can run local tools. Check the model route, CLI reachability, and website login separately, then ask for a named draft and review it before sharing.
Cohere calls North Mini Code a 30B total / 3B active open-weight coding model, trained for agentic coding. Its official docs list Chat V2, a model ID of north-mini-code-1.0, and tool use. The model card describes tool calling through Transformers chat templates; Cohere also points users to OpenCode for a hosted tryout. That proves the model has a tool-use path, not that every agent client knows how to drive Tabbit or that every hosting route uses the same ID. Cohere model docs · Cohere model card
Key takeaways
North Mini Code is a model, not a browser controller. The client supplies and runs tools.
Cohere lists it on Chat V2 and as open weights. Check the exact model identifier at the provider you choose.
Open weights do not automatically make the logged-in browser local. Decide whether the model runs locally or on a hosted API before sending page text.
Use a read-only task, collect links beside claims, and stop at a named draft.
A separate Tabbit task tab lets you keep working in your ordinary tabs.
Check the model route before connecting the browser
| Route | What Cohere documents | What you still need to check |
|---|---|---|
| Cohere Chat V2 | north-mini-code-1.0, tool use, evaluation access; Model Vault for production | Account access, current rate limits, data terms, and whether your client can pass tool results back |
| Open weights | Apache 2.0 release and model-card tool templates | Local runtime support, hardware, parser, and a client that can call the local tools you approve |
| OpenCode | Cohere links to an OpenCode tryout; OpenCode documents skills and shell tools | Select the available North Mini Code provider/model and confirm tabbit skill discovery plus local CLI access. No native Tabbit integration is implied |
The important boundary is the host. Cohere’s tool-use flag means the model can suggest a function call. A host must execute it, return the result, and decide what permissions the tool receives. OpenCode skills and OpenCode tools describe those client capabilities. For other clients, check their own skill loading and shell-tool rules. The Tabbit CLI complete guide covers the browser side; MCP and CLI compared explains the trade-off.
Set up a local browser path
Install Tabbit Browser and the supported Tabbit CLI skill using the official quick start. Then verify three independent checks: the client lists or loads the tabbit skill, its local command can reach the intended Tabbit instance, and the site is already signed in there. OpenCode discovers skills from supported ~/.agents/skills locations. If its list includes tabbit, you can request it with @tabbit or let the client load it through its skill tool. Do not expect /tabbit to be a universal slash command.
If your model is served remotely, browser execution can still be local only when the selected agent client has a supported local handoff. The model endpoint does not inherit browser cookies. The local vs cloud agent guide and existing-browser connection guide help check this boundary. A separate task tab isolates the work; it does not change the site’s permissions.
Turn team notes into a reviewable draft
Use a task you are authorized to perform, such as comparing a public release note with a matching entry in your team’s signed-in project wiki. Read only the named pages. Extract stable fields and attach a source URL to each one. Do not let the agent edit the wiki page or publish a summary.
| Step | Ask the local agent to do | Human checkpoint |
|---|---|---|
| 1. Open | Open the two URLs in the intended Tabbit profile | Confirm domain, account, page title, and record ID |
| 2. Compare | Return only the requested fields; mark missing values unknown | Follow each source link and resolve conflicts |
| 3. Draft | Prepare North Mini Code — release note brief in a local draft area | Check every claim against its cited page |
| 4. Stop | Wait at the review checkpoint | You decide whether to copy or save the text |
Example prompt (fixture name is illustrative):
Use the tabbit skill available in this local client. Read only these two pages: [public release note URL] and [signed-in team wiki URL]. Compare record ID GEO-104 by status and date. Return a table with source page, record ID, status, date, and a URL for each value. Mark missing or conflicting fields as unknown. Prepare a draft named “North Mini Code — release note brief”. Do not edit either page or click Save, Publish, Send, or Submit. Stop for my review.This is a proposed workflow, not a test of North Mini Code inside Tabbit. The project separately verified Tabbit CLI session reuse with a local fixture; that does not verify a North Mini Code integration. The first-task checklist and agent research with sources show how to keep claims traceable. Community discussion on browser automation also warns that end-to-end AI runs can be unreliable when quality matters, so leave a person in charge of conflicts and final edits. Reddit discussion
Protect the session and review the output
A signed-in page is still private even when the browser stays on your machine. If you use Cohere Chat V2 or another hosted route, page text may leave the device as model input. Check the provider’s current terms and keep sensitive pages out of the first trial. Do not export or paste session cookies into the prompt. Limit the task to read access, keep the draft local, and inspect every cited page before sharing. See how to reuse a browser login safely and browser-agent permissions.
Related reading
FAQ
Does North Mini Code browse the web by itself?
No. It generates tool calls, but an agent client must provide and execute the browser tool. Tabbit CLI is one local route to the signed-in Tabbit Browser session.
Can I run North Mini Code locally?
Cohere publishes open weights under Apache 2.0 and documents local inference paths. Check the current model card and runtime requirements for your hardware before downloading.
Can Cohere’s API use tools with this model?
Cohere lists North Mini Code on Chat V2 and labels tool use as a capability. The host still has to execute each returned tool call and pass its result back.
Will my site login be shared with the model?
The local browser keeps the login in Tabbit. Page text you ask the agent to summarize may still be sent to a hosted model provider, so check its data terms first.
Can the agent save the shared document?
Use a draft checkpoint. Ask it to prepare a named draft, show source links, and stop before clicking save, publish, send, or submit.