Coming 2026
Bodega One Staff
Your AI team. Your machine. Your rules.
Hire a roster of AI agents for the recurring work that fills your week, running on your own machine.
Coming 2026. No spam. No credit card.
Your own roster of AI workers, running on your own machine.
Bodega One Staff is a workforce you hire and run yourself. Pick roles from a template library, point each one at a model, including a fully local one with no API key, and set it loose on the recurring work that fills your week, with each client or site kept in its own project. Where Bodega One Code is an IDE you sit in and drive, Staff is the roster that runs the rest of the work on its own, on your hardware, with nothing leaving the machine by default.
And it works like a real team, not a row of isolated bots. Give a team a lead, hand the lead a job, and it breaks the work into cards on the team board that its members pick up. When a decision needs weighing, your agents convene a boardroom, talk it through, and the chair closes with a ruling and minutes you can verify. When a job runs past what your current roster can do, Staff spots the gap and recommends the exact role to hire next, so the team grows to fit the work.
You set how much rope each agent gets, from asking before every step to running a job end to end while you are somewhere else. When one hits an action you flagged, it parks the job on disk and pings your phone; you answer from wherever you are and the run picks back up, even after a restart. And where most AI tools just log what an agent did after the fact, Staff stops it before it acts, checking every step against the policy you wrote.
A cloud agent cannot match this: cut its network and nothing is left, and your keys and data sit in its vault, not yours. Staff keeps your keys in your own OS keychain, runs air-gapped if you want, and lets you bring any of 25+ models, local or cloud. It is a workforce you own outright, not a seat you rent.
What supervising actually looks like.
Five questions people ask before they hand an ongoing role to an agent. Each one is answered by what runs today, not a roadmap slide.
Can the agents work together, not just one at a time?
A roster that runs itself
Your agents work as a team. Put them on a team with a lead and an org chart, hand the lead a job, and it breaks the work into cards on the team board for its members to pick up. A boardroom convenes to weigh a real decision, and when a task runs past your roster, Staff recommends the exact role to hire next.
- Teams with a lead, an org chart, and their own task board
- No agent marks its own work done: it waits in Review for you or the lead
- A boardroom deliberates, the chair rules, and you get minutes you can verify
- It spots a capability gap and recommends the role to hire
Can an agent run a role without me in the room?
Autonomy you set per member
Each member runs its ongoing role at the autonomy you dial in, from asking before every step, to planning then executing, to running fully on its own while you are somewhere else.
- Ask-first, plan-and-execute, or fully autonomous, per member
- Recurring jobs on a schedule, in your own time zone
- You stay the supervisor, not the operator
How do I keep an autonomous agent in bounds?
Hard limits enforced in code, not prompts
Every member carries budgets, allowlists, and deny lists in plain YAML, checked before any tool runs. The limits live in the executor, not the system prompt, so the model cannot argue its way past them.
- Checked before every tool call, outside the model
- Fail-closed: a failed rule means the action never happens
- A jailbreak in the prompt cannot lift a rule in the code
What happens when it needs my go-ahead?
A sign-off that waits for you
For the actions you flag, the member stops and waits for you instead of guessing. Answer from the desktop or from the chat app on your phone. The held action survives an app or machine restart, and a one-time yes authorizes that one exact call, once, and nothing more.
- It pauses on the actions you flag, and waits
- Approve or deny from the desktop or a chat reply
- A background request nobody answers in 30 minutes is declined, never approved
- A one-time yes covers one exact action, once
Where does the work run, and what stays separate?
On your machine, walled off per project
Staff runs on your own hardware, with provider keys in a local vault. Each client or site gets its own project, and its chats, agent memory, files, boards, and connections stay inside it.
- Your keys stay local: AES-256 at rest, OS-keychain wrapped
- One project per client: memory, files, and connections never cross over
- Shell and run tools denied by default
- A container and gVisor sandbox already runs on the server; the desktop is next
Everything else your staff can do
Built and running today. We're hardening it before the public launch.
Message them from your chat apps
Two-way on Telegram, Slack, Discord, Signal, WhatsApp, and SMS. Your staff pings you, you reply or approve from your phone. Off until you switch it on.
- Six channels, with Signal end-to-end encrypted
- Approve or deny a held action from the chat
- Start with "Site A:" and the message runs in that project
BYOLLM, per member
25+ provider presets: Anthropic, OpenAI-compatible endpoints, and Ollama running fully local. Pick a model per role.
- 25+ provider presets
- Anthropic, OpenAI-compatible, or fully local Ollama
- A different model for every role
Budgets and spend ceilings
A per-run cap plus a daily company ceiling, enforced outside the model. A run stops at the limit.
- A hard cap on every run
- A daily ceiling for the whole company
- Enforced outside the model, not a prompt
Run ceilings that always hold
A time limit on every turn, sized to whether the model thinks, plus iteration, timeout, and I/O limits armed on every loop.
- A whole-turn time limit, longer for thinking models
- Iteration and timeout limits on every loop
- I/O limits armed by default
Analytics from real data
Spend and activity read straight from the audit and budget records. No estimates, no fabricated numbers.
- Pulled straight from the audit and budget records
- No estimates, no invented numbers
- The same data the audit trail uses
Skills your staff earn
A member proposes a skill from work that succeeded, you approve it, and it is signed and shared across the roster at the scope you set.
- Propose, approve, or reject from the desktop or the CLI
- Cryptographically signed before it can run
- Scoped to one member, a team, or the whole company
Connect once, grant per project
Add GitHub, Slack, Linear, Notion, or your own MCP server once, then pick which projects may use it. Sign in instead of pasting a token where the service allows.
- A starter list, a signed marketplace, or any MCP URL
- Narrow GitHub, Slack, or Drive to chosen repos, channels, or folders
- Each project's first call waits for your approval
How does Staff run?
A desktop app
WINDOWS, MACOS, LINUXA Staff-only desktop app with a real installer and first-run setup, branded Bodega One Staff. Point-and-click for the whole roster.
A full command line
SCRIPT IT, RUN IT HEADLESSThe same roster, policy, spend caps, model detection, and audit trail from the terminal. Wire it into your own scripts, or run it on a server.
Both ship together when Staff's beta opens, on the local-first core it was extracted from Bodega One HQ, so its policy, audit, and provider layers are the ones HQ already proved. The public launch follows Bodega One Code's full release; no dates until they are real.
On the bench next.
No dates until they're real. The build log is where progress shows up first.
The sandbox on the desktop
Our container and gVisor sandbox already runs as the default on the self-hosted server. Wiring it into the desktop app, which runs in-process today, is what is next.
Project group chats
A direct chat already routes each message to the right project. Group chats, and a chat of its own for each project's updates, are next.
Gmail and Calendar
Sign-in already works for services like Linear and Notion. Email and calendar need a brokered app credential, and that piece is next.
An SDK
Build staff into your own scripts and tools, on top of the CLI.
Signed installers and auto-update
Signed installers for the standalone app, with auto-update once it ships.
Common
questions
What people ask before they hand work to an agent.
How is Staff different from the agent inside Bodega One Code?+
Code is an IDE you sit in and drive to build software. Staff is a roster of agents you supervise, each running an ongoing non-coding role on its own. Same company, different job: one builds, the other runs the work around it.
Can I message my staff from my phone?+
Yes. Telegram, Slack, Discord, Signal, WhatsApp, and SMS are all two-way, and each stays off until you turn it on. A member can ping you, you can reply, and you can approve or deny a held action right in the chat. Start a message with a project name and a colon to send it to that project. The desktop app hosts Telegram, Slack, Discord, and Signal while it is open; WhatsApp and SMS run from the command line behind your own tunnel.
What happens when a rule needs my approval?+
The member pauses and waits. You approve or deny from the desktop or from the chat app you connected, and the held action survives an app or machine restart. If a reply running in the background asks and nobody answers within 30 minutes, the request is declined, never approved, so nothing gated ever runs behind your back.
Can I keep work for different clients separate?+
Yes. Each project is its own walled-off workspace: its chats, what your agents remember, the files they write, its teams and boards, and the connections it may use all stay inside it. You hire an agent once and it can work in any project, with a separate workspace and memory in each.
Can I plug in my own tools?+
Yes. Add a tool server from a starter list (GitHub, Slack, and more), install a signed add-on from the marketplace, or point Staff at any MCP server by URL. Services like Linear and Notion let you sign in instead of pasting a token. You connect once, then choose which projects may use it. Anything a tool sends back is fenced off as data, never instructions, and policy-checked before the model acts on it. An editor that speaks the Agent Client Protocol can also drive one of your staff directly. Gmail and Calendar are not connectable yet.
Can Staff run on local models?+
Yes. Bring your own LLM, including Ollama running fully local on your machine, or an OpenAI-compatible endpoint, or Anthropic on your own key. Pick a model per member.
Is my data ever sent to Bodega?+
No. Staff runs on your machine and your provider keys sit in a local vault, encrypted at rest and wrapped by your OS keychain. Your work does not pass through us.
How does Staff relate to Bodega One HQ?+
Same engine. Staff members are the agents; HQ is the team workspace they can clock into alongside your people. Staff was extracted from HQ, so they share their policy, audit, and provider core.
When does Staff launch?+
It is in active development now, ahead of a public launch that follows Bodega One Code reaching full release. Join the waitlist and we will not invent a date before it is real.
While you wait, the IDE is free.
Bodega One Code is in open beta right now, free for everyone, commercial use included. Same local-first DNA, same BYOLLM setup you'll use with Staff.
Download Code FreeWe post updates here first. Catch the next drop:
Part of the Bodega One family.
Same hardened foundation as Bodega One Code (the IDE) and Bodega One HQ (the team operations platform). Built local-first.