Shared Projects: How People and AI Agents Work Together in Orkas

Until now, everything you set up in Orkas was yours alone. You wrote project instructions, your agents built up memory, you collected reference files and planned work as to-dos — and the moment you needed a colleague, a contractor or a client to pick up part of the work, all of that context stayed on your machine. You could paste a prompt into a chat app, but a prompt is not a playbook.

Orkas 1.8 makes the project the place where people work together. You invite people into a project; they get the same instructions, memory, Library files, to-dos and shared agents; and each of them does their part with their own AI team.

How Orkas works for one person

Shared projects build on the parts of Orkas that already exist, so a short tour first.

Your computer — the Orkas desktop app (macOS · Windows)
├─ Commander — plans the work, picks agents, brings results together
│  ├─ Built-in specialists — research, writing, slides, images, video, office files, code, UI, SEO/GEO …
│  ├─ Your custom agents — instructions + skills + standards
│  └─ External CLI agents — Claude Code, Codex, OpenCode …
├─ Agent harness — run loop, tools, sessions, permissions
├─ Context — memory (agent · project · core), skills, connectors
└─ Projects — instructions, memory, Library, to-dos, automations

Models — Orkas official models, or your own provider keys
Orkas cloud — account and credits · optional personal sync · shared projects

Commander and specialists. You talk to Commander. It decides whether to answer itself, hand the work to one specialist, or send pieces to several agents at once and collect what they return. Specialists are agents with a narrow job and the tools for it. You can build your own agents in a chat: a role, a working procedure, standards it must meet, and skills it can use. If you already use Claude Code or Codex, you can connect them as agents too.

The agent harness. Every agent runs on the same runtime: a loop that calls the model, runs tools, checks permissions and keeps a session it can recover. We wrote about it in detail in The layer that turns a model into a product.

Memory and skills. Each agent keeps its own memory, a project keeps project memory, and Commander keeps a small core memory that carries across all your work. Skills are folders of instructions and scripts an agent can load when a job calls for them.

Projects. A project is the home for work that runs over days or weeks:

• Project instructions — goals and working rules, up to 4,000 characters, read at the start of every task in the project.
• Project memory — facts the project keeps across tasks.
• Library — the project's reference files. Agents search and read it, and can save deliverables back into it.
• To-dos — planned work, each with a status: Pending → In progress → Awaiting confirmation → Done. When an agent finishes a to-do, it moves it to Awaiting confirmation; a person checks the work and sets it to Done, or back to In progress.
• Auto — an optional switch that lets Commander advance one ready to-do at a time. It is off by default and advances at most 20 to-dos a day.

Models. You can use Orkas's official models with credits, or add your own provider key (OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter or any OpenAI-compatible endpoint).

All of this was built around one person directing an AI team. Shared projects keep that model and add more people.

Shared projects: a project can have members

When you share a project, Orkas uploads the parts a teammate needs and leaves the rest where it was. The dialog says it plainly before you confirm: "Project instructions, memory, files, to-dos and their attachments will be uploaded. Tasks and automations stay private."

Maya (owner)  ── her Commander, agents, models, credits, chats
Leo  (member) ── his Commander, agents, models, credits, chats
Ana  (member) ── her Commander, agents, models, credits, chats
        │
        │  instructions · memory · Library · to-dos · shared agents
        ▼
Shared project in Orkas cloud — storage counted to the owner

Each member's Orkas keeps a full local copy of the shared parts, and their agents work from that copy.

A walk-through

Maya runs a small content studio. She shares her Spring launch project with Leo, a copywriter, and Ana, a freelance designer.

1. Maya invites them. In the project's Members card she clicks Invite members. The first time, the dialog explains what will be shared and offers Start sharing and create link. She copies the link — it is valid for 7 days — and sends it to Leo and Ana.

2. Leo joins. He opens the link in his browser, signs in (a free account is enough), clicks Accept invitation and then Open Orkas. If Orkas isn't installed yet, the page offers a download. The project appears in his sidebar, marked as shared, and its files, to-dos and agents download in the background.

3. Maya shares her writing agent. Her custom Brand Writer agent knows the studio's tone. Custom agents aren't shared automatically, so she clicks Share on it. Orkas publishes it as a package: its instructions plus its skills, including skills it picked up while working. Its memory and chat history are not included. Leo's and Ana's Orkas install it automatically.

4. Maya plans the work. She adds a to-do — "Draft three launch emails from the brief" — sets Responsible member to Leo and attaches the brief.

5. Leo's AI does the work, on Leo's computer. Leo binds Brand Writer to the to-do and chooses Process now. Orkas first claims the run with the server: only the responsible member can start it, and only one run can be active at a time. A new task conversation opens on Leo's machine. The agent reads the project instructions, project memory and the attached brief, can search the Library, and uses Leo's model and credits. When it is done, it moves the to-do to Awaiting confirmation and saves the drafts to the Library, as the project instructions ask.

6. Everyone sees the result — not the conversation. Within a few minutes, or as soon as Maya opens the project or clicks Refresh, red dots appear on the To-do and Library tabs. She opens Leo's drafts from the Library and, once they're right, sets the to-do to Done. Leo's conversation with his agent stays on his side.

7. Two people edit the same file. Ana and Maya both tweak the brief. If their edits don't overlap, Orkas merges the text automatically. If they do, Project conflicts shows My copy next to Shared copy, and each person picks one.

One tip from step 5: an agent saves files to the Library only when it is asked to. Add a line such as "Save every deliverable to the Library" to your project instructions, and every member's agents will follow it.

Many people, many agents

In a shared project, each to-do has a person who is responsible for it, and an agent may be bound to do the work.

• Commander can plan for the team. In a project, Commander can read the whole to-do list — including teammates' items — create to-dos, assign them to members, and attach files. It can also split one unassigned piece of work across several specialists and combine their results.
• Only you choose the agent for your own to-dos. You can assign a to-do to Leo, but the agent that runs it is Leo's choice. Reassigning a to-do clears its agent.
• Shared agents keep everyone on one playbook. When Maya improves Brand Writer, Orkas republishes it and members get the new version automatically; a run that is already going keeps the version it started with. Only the person who shared an agent can change or unshare it.
• Built-in agents are available to everyone. Marketplace agents added to the project are installed for each member automatically.
• CLI agents stay personal. Leo can still bring Claude Code or Codex into his own tasks in the project — they get the project instructions and memory — but they can't be shared or bound to to-dos in a shared project. They rely on a CLI that is installed and signed in on one person's computer.
• Auto works per person. Each member can turn on Auto for themselves. In a shared project it only advances to-dos assigned to that member.

Why we built it this way

Share the project, not the conversation

A working conversation with an agent is full of false starts, half-formed requests and private notes. A teammate rarely needs it; they need the result and the rules that produced it. So shared projects sync the durable parts — instructions, memory, Library, to-dos and agents — and keep conversations private. Results reach the team the way finished work should: as files in the Library and a to-do waiting for confirmation.

Everyone brings their own AI

In a shared project, a member runs agents with their own models, keys, connector sign-ins, permission settings and credits — the same rules as their personal work. The owner doesn't pay for a member's AI use, and nobody's API keys or credits are shared. The server stores the shared files, checks who may do what, and hands out run leases. It does not run agents.

People own the work; agents do it

A to-do is assigned to a person, not to an agent. The agent decides how to do the work; it never replaces the person responsible. Only that person can start a run, and when an agent finishes it hands the to-do back for confirmation instead of marking it done. With several people and many agents in one project, someone has to be accountable for each piece — so it is always a person.

Asynchronous on purpose

Edits you make in Orkas save on your computer first and are published in the background. Others pull them when they open the project, return to the app, click Refresh, or at the next five-minute check. There is no live co-editing. Agents write a lot, and often all at once; a predictable checkpoint is easier to trust than files changing under you. When two edits collide, Orkas merges what it safely can and otherwise asks a person to choose.

Small teams first

The owner needs a paid plan to share; members join with a free account. That keeps the cost of saying yes to an invitation at zero for a freelancer, a client or a new hire, and the cost of sharing — the plan and the cloud storage — sits with the person who set the project up. Plans are sized for small teams — three to ten people in a project — not for departments.

Wrapping up

Orkas started as one person directing a team of AI agents. Shared projects keep that shape and widen it: a small team shares what makes those agents useful — the rules, the memory, the material and the plan — while each person keeps their own AI, keys and conversations, and stays responsible for their part of the work.

For four small-team setups you can copy, see Small-team shared projects.

If you already work with a colleague, a contractor or a client, try it on one project: write your shared playbook into its instructions and invite them in. Orkas is free to download for macOS and Windows.