ClarkCant
GitHub

Open-source AI agent

Tell Clark
what you want.

ClarkCant works on your computer through a single conversation. Type it or say it. Clark handles the rest.

    A preview. Replies on this page are scripted; the real Clark runs on your own machine.

    Chapter one

    Clark couldn't.

    Clark Cant was our first AI agent, named as a joke on Clark Kent. In the beginning he could barely do anything. Every step ran into a guardrail: a refusal, a confirmation, a permission he didn't have.

    Honestly? He was almost useless.

    Chapter two

    Then Clark could.

    The guardrails didn't vanish. They moved into the system underneath: work confined to the folders you allow, a visible trail of what he did, and a Stop button that always works. The wall of dialogs came down.

    Clark became a teammate, and helped us build more than one product.

    Chapter three

    ClarkCant Can.

    We kept the name as a reminder of where he started. ClarkCant is that teammate, opened up for everyone: one conversation, one voice, and an agent that gets on with what you asked.

    Talk, don't operate

    One conversation.
    Everything else is plumbing.

    Most agent tools hand you the machinery and wish you luck. ClarkCant keeps it where it belongs: underneath. You never need to learn how an agent works to get something done.

    Clark

    Move last week's screenshots into a folder and name them properly

    Done. 37 screenshots are in Screenshots / Week 38, named by what's in them.

    Tell Clark what you want…

    What you see.

    • SessionsClark keeps one conversation. You never pick a session.
    • Context windowsWhat to remember and what to drop is his problem.
    • Models & routingChoose one if you care. Otherwise, don't.
    • Skills & toolsFound and used when the task needs them.
    • SubagentsIf work splits, it comes back into the same thread.
    • MCP serversConnected through chat, not config files.
    • Hooks & workflowsBehaviour you describe, not wiring you draw.
    • Runtimes & nodesYour laptop, your server. Same conversation.
    • PoliciesPick how bold Clark is, once. Change it by asking.
    • Audit trailEvery action written down, there when you look.

    What Clark handles.

    Rich replies

    Sometimes the best answer
    isn't text.

    When a list, a chart or a tiny app says it better, Clark draws one right inside the reply. No new window, no new app to learn. Go on, use these.

    Split a $186 dinner five ways, with tip

    Which of these three laptops should I get?

    Give me a 25-minute focus timer

    Find the contract I signed in March

    Examples built for this page to show the idea, with made-up data. In ClarkCant, widgets and mini-apps render inside the conversation, and every tool Clark uses appears as a widget you can open to see exactly what it did.

    Voice

    Voice is the
    same Clark.

    Not a separate assistant with a separate memory. Speak and it lands in the same conversation, with the same context and the same abilities. Start typing, finish talking, or the other way round.

    1. Draft a reply to Linh about moving Friday's meeting
    2. Drafted. Want me to suggest Monday at 10?
    3. “Make it Tuesday, and a bit warmer.”
    4. Updated. Tuesday, and friendlier.

    Uses your microphone only to move the Orb. Nothing is recorded or sent anywhere.

    What we hold ourselves to

    AI should make computers
    easier, not more complicated.

    1. Conversation is the app.

      No dashboard to learn, no sidebar of modes. Asking, seeing results, changing settings: it all happens in one place.

    2. Talk, don't operate.

      Say the outcome you want. Finding the right tab, tool or command is Clark's job.

    3. Voice is the same Clark.

      One agent, one memory, whether you type or speak.

    4. Sometimes the best answer isn't text.

      Tables, charts and small apps appear right in the reply, when they help.

    5. Power underneath. Not in your face.

      Details stay out of the way until you ask for them or something needs you.

    6. Don't make people learn the architecture.

      Nobody should need to understand sessions, runtimes or servers to finish a task.

    7. Honest, always.

      No fake progress bars, no “done” that isn't. If Clark doesn't know, he says so.

    Open source · Apache-2.0

    Yours to run, read
    and make better.

    ClarkCant runs on your own machine or server. The code is open, the roadmap is public, and it's meant to be approachable even if you've never touched an agent before.

    Early days v0.2 · bootstrap, not a release

    We'd rather tell you where things stand than impress you. The foundations are in and tested; parts of the experience are still being built.

    Working today

    • A conversation that streams replies and draws tool calls as widgets
    • A headless runtime you can run on your own machine or server
    • Read-only search across your files
    • Stop, and an audit trail of every action

    Not yet

    • The full desktop app
    • Live sign-in to services like Google Calendar
    • Installing capabilities end to end from chat
    • Native computer control on macOS and Linux
    Read the full, honest ledger →

    Try it from source

    Needs Node.js 22.19+ and pnpm. A friendlier installer is on the way.

    git clone https://github.com/digitopvn/clarkcant
cd clarkcant
corepack enable
pnpm install
node apps/runtime/src/main.ts --data-dir ./.data
    Star on GitHub Share an idea

    With gratitude

    Standing on Pi.

    ClarkCant is inspired heavily by the philosophy and work behind the Pi agent ecosystem, and its agent core is built on the Pi coding agent SDK. Small, understandable, yours. Thank you to Pi's authors and community.

    ClarkCant Can.

    You don't need to learn how the agent works.
    Just tell Clark what you want.

    Tell Clark