What's New in Jobbit: Agent Browser, Model Choice, Projects and More (September 2026)

September's release is the biggest set of changes to Jobbit since launch, and most of it is about one idea: letting you see, direct and share what the AI agent is doing. The agent's browser is now on screen and you can take the wheel. Paid plans can choose the model behind the agent, including Claude, GPT, Gemini, DeepSeek and Kimi, and the models that make images and video. Projects give a group of chats shared files, instructions and secrets, and a team to work in them. A message queue, chat forking and a Goal mode change how you steer long runs. Here is everything that shipped, what it is for, and where to read more.

Agent browser (beta): watch, then take control

When a task needs the web, the agent's cloud browser now appears in the workspace panel beside your chat, live, with the page title and address. Press Take control to log in, click or type yourself, then Give back and the agent carries on. The agent never asks for passwords or codes; you handle those moments in a browser you are driving. The browser lives inside the run, so nothing touches your own machine, and the last page a finished run saw stays on screen.

Read Jobbit's agent browser: watch your AI agent browse the web, then take control.

Choose your model: Jobbit tiers and named third-party models

The model button in the composer now lists Jobbit's own 1.13 Lite, Pro and Ultra tiers alongside named models from Anthropic, OpenAI, Google, DeepSeek, Moonshot, xAI, Alibaba, Meta and Z.ai, each with its provider mark, context window, maximum output and price per million tokens. Two named models are open to every account; the rest, and the Pro and Ultra tiers, are part of paid plans. Switch models in the middle of a chat and the context carries over.

Read choose your AI model in Jobbit.

Choose your generation models

In Settings under Models, paid plans can now pick which model produces images, video, music and speech, from providers including Black Forest Labs, Recraft, Krea, Kling, Runway, MiniMax, ByteDance, Google and OpenAI. Each role shows the default, your choice and what will actually run, and every model lists its price in credits per image or per few seconds of video, so quality and cost are a deliberate trade. Free accounts keep running on the defaults.

Projects: shared files, instructions, secrets, notes and members

A project gives a group of chats the same files and knowledge. Drop in documents and the agent reads them so any chat can search inside; each file shows whether it is findable by name, being read through, searchable inside or too long to read in full, and you can describe files so the agent picks the right one. Write instructions once and every chat inherits them. Store secrets such as API keys by name so builds and automations can use them without the values ever appearing in a conversation. Leave notes with bylines for the team. Invite members with join links that can expire or be limited to a number of uses; shared projects appear under "Shared with you". The project page shows the room left as a ring, in tokens explained as pages, and says how much of the project a file will take before you add it.

Read Jobbit Projects: give your AI agent shared files, instructions, secrets and a team.

Message queue: talk to a working agent

The composer stays open while the agent runs. Messages you type join a server-side queue of up to ten, go out in order when the agent can take them, and survive a refresh. Use it for corrections, additions and follow-ups without stopping the run; remove a queued message before it goes if you change your mind.

Fork a chat from any message

Any chat can be forked into a new one, from a message you choose or as a whole, from the header or the sidebar. Both chats keep a "continued from" link to the branch point, and forks inside a project keep the project's files, instructions and secrets. It is how you explore two directions from one starting point, or hand a colleague a copy to continue.

Execution modes: Auto, Plan and Goal

Next to the model button, choose how independently the agent works. Auto picks the best approach. Plan writes and shows a plan before acting. Goal keeps working until your request is done, reporting progress along the way. Plan mode is the right start for anything big; Goal mode is for clear outcomes you want finished.

Read steer your AI agent mid-run: message queue, chat forking and Goal mode.

Workspace file explorer

Everything the agent creates in a chat now appears in a proper file explorer in the workspace panel, on desktop and on phones: browse folders, search inside a folder, upload files by drag and drop, rename, duplicate and move, and download a selection as one archive. Files the agent makes show up as you watch, and long lists stay put while you scroll.

Context meter, compaction and notifications

The composer shows how much of the model's context a conversation is using. When a chat grows long, compact it: the agent summarises what matters and continues with room to think. Turn on browser notifications and your browser tells you when a run finishes or the agent needs a decision, so you can leave the tab. Voice input is in the composer too: record, and Jobbit transcribes.

Plans and hosting

Plans are unchanged in shape and clearer on the page: Free with 100 credits a day and permanent hosting that resets every 30 days; Plus at £15 a month with 6,000 credits, ten concurrent tasks and an 8 CPU, 8 GB hosting pool with a custom domain; Pro at £79 with 35,000 credits and a 32 CPU, 32 GB pool; Max at £159 with 80,000 credits, twenty concurrent tasks and a 64 CPU, 64 GB pool. Extra tasks on every plan wait in a queue and start automatically.

Read Jobbit pricing explained.

Smaller improvements

• The workspace explorer, projects and chat list are usable on a phone, with the selection bar and menus fitted to a narrow screen.

• Chats show who started them in shared projects, and long chat lists load as you scroll.

• Renaming and deleting a chat each have a clear dialog.

• Tool calls in the chat are named the way the agent names them, and a task that carries a shell command shows it.

• The light theme has a background of its own, and the dark theme's ink animation is slower and available in both.

• Recording costs are quoted correctly, and the project counter reflects notes as well as files.

Why this release matters

Taken together, these changes move Jobbit from "an agent you talk to" toward "an agent you work with": you can see what it is doing, correct it while it does it, choose the engine behind it, give it your context once, and share all of that with a team. That is the direction the whole category is heading, as we argued in the future of AI agents, and it is available today on every plan, with the model catalogue and generation choices on Plus, Pro and Max.

Frequently asked questions

What are the biggest new features in Jobbit's September 2026 update?

A live agent browser with take-control handoff, the choice of Jobbit tiers and named third-party models on paid plans, generation model selection for images, video, music and speech, Projects with shared files, instructions, secrets, notes and members, a message queue for steering runs, chat forking, Auto, Plan and Goal execution modes, and a workspace file explorer.

Do I need a paid plan for the new features?

The agent browser, Projects, message queue, forking, execution modes, workspace explorer, compaction, notifications and voice input are available on every plan within its limits. Choosing models beyond the free tier and the two open named models, and choosing generation models, is part of Plus, Pro and Max.

Which third-party models can I use in Jobbit now?

Named models from Anthropic, OpenAI, Google, DeepSeek, Moonshot, xAI, Alibaba, Meta and Z.ai appear in the model picker with their context size and token prices. The catalogue is served live, so it reflects what is currently available; two named models are open to free accounts.

Is the agent browser secure?

It runs in Jobbit's cloud inside a single run, treats page content as data rather than instructions, pauses before irreversible actions, and never asks for your passwords: you type them yourself while you have control, and control ends when the run ends.

Where can I see the full pricing?

On the pricing page at jobbit.uk/pricing, and in our plain-English guide to Jobbit pricing, which explains credits, concurrent tasks, hosting allowances and which models each plan includes.

Everything above is live now. Start free on Jobbit and try the new release on a real job.