Coming soonA product workspace for AI-built software

Build with AI.Ship with proof.

Describe your idea or import a project. VibeYourFlow plans the missions, builds each one in a sandbox, and hands you tested code, a live preview and a passport of evidence. You decide what ships.

Release news on

--Days
--Hours
--Minutes
--Seconds
Ready
  1. Task inputqueued
  2. Planning agentqueued
  3. Coding agentqueued
  4. Run testsqueued
  5. Previewqueued
  6. Human approvalqueued
  7. Pull requestqueued
run ready
Mock provider · no keys, no network, no real repository

One change, start to finish

Every change in the app follows the same five stations. You always see which one it is at, and what it is waiting for.

  1. RequestDone

    “Change the headline on the home page to Appointments in two clicks.”

    You write it in Build, like a message.

  2. PlanDone

    A planning agent reads the code and writes down what it will change and how to check it.

    You can read the plan before anything is written.

  3. ChangeDone

    An agent edits the code on its own branch, inside an isolated sandbox.

    Your default branch is untouched.

  4. EvidenceDone

    Tests run and a preview starts. A step that did not run says “Not run”, never “Passed”.

    You see the result next to the changed files.

  5. DecisionWaiting

    The run pauses until a person approves, requests changes or rejects it.

    Nothing ships without your approval.

Watch a run happen

Scroll through one run. The app window follows the step you are reading, and every tab is live.

Change the headline on the home page

run 1284 · feature/headline-4f2a

Running
  1. Task inputdone
  2. Planning agentrunning…
  3. Coding agentqueued
  4. Run testsqueued
  5. Previewqueued
  6. Human approvalqueued
  7. Pull requestqueued
$ run start "Change the headline on the home page"
planner reading repository (14 files)

Hand a task to a team of agents

Every card on the board has a chat of its own. Pick the workflow that builds it, or let a main agent split it so sub-agents work side by side. Try it: the rules here are the app's own.

To do

1
  • #12

    Invoice export as CSV

In progress

0

    Review

    0

      Done

      0
        #12

        Invoice export as CSV

        Split into sub-tasks

        A main agent splits this task into up to 8 sub-tasks. After you accept the split, a sub-agent builds each one in its own Build chat, up to 4 at a time.

        The task and the split are examples. The workflows, the limits and every rule that starts, skips or moves a card are the app's own.

        Not only web apps

        Backends, command-line tools, libraries, games, desktop and mobile apps, kernels and firmware. Pick a kind to see its toolchain, how it shows itself and what “Build a feature” does with it.

        Operating system

        Kernels and bootloaders: built here and booted in QEMU, serial output in the run log.

        Toolchain
        NativeC and C++ (GCC, Clang, CMake), Go, Rust, Python with uv, Zig, QEMU, GRUB and the ARM embedded compiler
        Shows itself as
        Output in the run log

        Ready-made starters

        • x86 kernel in CNativeScreen (desktop app or VM)

          A Multiboot kernel on a GRUB ISO; the test boots it in QEMU and reads its serial port, the preview shows its screen.

        • Cosmos kernel in C#CosmosScreen (desktop app or VM)

          A Cosmos kernel in C# on a Limine ISO; the test boots it in QEMU and reads its serial port, the preview shows its screen.

        “Build a feature” for this kind

        1. Planner
        2. Install dependencies (with network)
        3. Implementation
        4. Install dependencies (with network)
        5. Tests
        6. Fix failing tests
        7. Install dependencies (with network)
        8. Tests (re-check)
        9. Review
        10. Fix review findings
        11. Install dependencies (with network)
        12. Tests (final gate)
        13. Approval
        14. Pull request

        Only the install steps have network. Agents, builds, tests and previews run without it.

        From the first upload to a version you can trust

        Ten parts of one product workspace, built and waiting for release. Every claim the app makes is backed by a run, a test or your own approval. Release news follows on 1 November 2026.

        1. Upload
        2. Brief
        3. Missions
        4. Build
        5. Passport
        6. Release check

        Coming soon

        Bring your code, keep every version

        Start from a .zip or .tar.gz right in the browser. Every approved run is saved as a new revision of the project, and any revision downloads again as source.

        Revisions

        1. Revision 3 · Approved run142 files · 1.2 MBCurrent
        2. Revision 2 · Approved run139 files · 1.2 MB
        3. Revision 1 · Upload131 files · 1.1 MB

        Illustrative example.

        Work from Claude, ChatGPT or your editor

        The app is an MCP server. Connect a client with an API token and it can create projects, start runs, split tasks and follow approvals, with your permissions and nothing more.

        claude mcp add --transport http vibeyourflow https://your-vibeyourflow.example/api/v1/mcp --header "Authorization: Bearer vat_…"

        The token comes from Settings → API tokens. Every call runs with that person's role in each project, through the same API, validation and audit log as the app.

        1. list_projectsReads
        2. list_runsReads
        3. get_runReads
        4. get_run_diffReads

        73 tools. Each one replays the API route of the same name, so a role that may not approve cannot approve here either.

        All tools by area
        Projects 8
        list_projects · get_project · create_project · list_github_installations · list_github_repositories · get_project_overview · project_activity · analyze_project
        Workflows 8
        list_workflow_templates · list_workflows · get_workflow · create_workflow · update_workflow · save_workflow_draft · validate_workflow · publish_workflow
        Runs 14
        start_run · list_runs · get_run · cancel_run · approve · reject · request_changes · get_run_report · get_run_diff · list_run_events · list_run_previews · rerun_step · fix_step · get_chat
        Pull requests 2
        get_run_pull_request · merge_run_pull_request
        Board 13
        get_board · create_task · get_task · update_task · move_task · comment_on_task · my_tasks · start_task_run · add_checklist_item · update_checklist_item · request_task_review · answer_task_review · my_reviews
        Sub-agents 13
        get_task_sub_agents · split_task · start_sub_agents · discard_sub_agents · cancel_sub_agents · list_sub_agents · create_sub_agent · update_sub_agent · set_task_sub_agent · route_task · get_task_routing · apply_task_routing · dismiss_task_routing
        Product 6
        plan_mvp · list_missions · update_mission · start_mission · list_memory · create_memory
        Team 4
        list_project_members · add_project_member · update_project_member · remove_project_member
        Environment 5
        list_sandboxes · get_project_preview · start_project_preview · get_security_overview · start_security_scan

        Built so that mistakes stay small

        Isolation is the first design rule, not an add-on.

        Sandboxed execution

        Repository code only ever runs in resource-limited containers on a separate runner.

        Branch isolation

        Every run gets its own branch and worktree. Your default branch is never force-pushed.

        Durable by design

        Temporal keeps runs alive across restarts, retries and long waits for approval.

        You merge

        Changes become pull requests. Nothing is merged without a person.

        Security scans

        Read-only AI reviews and dependency audits. Findings carry file and line, and a fix still waits for your approval.

        Evidence, not scores

        A check that did not run is never shown as passed, and a feature is only verified when a test backs it.

        A sandbox belongs to the project

        It is where the coding agents (Claude Code, Codex, Cursor) and their sign-ins are installed and where a run's commands execute. You decide who shares which one.

        Sandbox policy

        Who works in which sandbox

        Up to 10 sandboxes and 5 servers of your own per project. A run stays in the sandbox it was prepared in, even if a pick changes during the run.

        Default sandbox

        Platform runner · Claude Code, Codex signed in

        • Alex
        • Sam
        • Kim

        Build machine

        Project server · Claude Code, Cursor signed in

        • Nobody
        • Alex Owner
        • Sam Developer
        • Kim Developer

        Under one shared sandbox a pick is kept but ignored, so everyone shares the same sign-ins.

        Build the workflow your team actually uses

        This is the editor, in miniature. Add nodes, drag them around, open their settings, and watch the validator think along.

        workflows / demo-workflow · draft

        Node library

        Planning agent

        Reads the repository and writes a plan.

        Provider
        Anthropic
        Model
        ← Model router · model
        Output
        plan.md
        Valid. Tip: add a test step so changes are checked.

        The engine underneath: workflows you control

        Every part of the loop is a feature, not a plugin.

        1. Planning agentdone
        2. Coding agentdone
        3. Run testsrunning…
        4. Human approvalqueued

        Visual workflow editor

        Drag planners, agents, tests, previews and approvals onto a canvas. Every setting without a fixed value becomes an input port that an earlier step can feed, and the graph is validated before it can run.

        container
        1 CPU1 GBno egress

        no secrets, no platform access

        Isolated sandboxes

        Commands run in containers with CPU, memory and network limits. The platform never executes your code itself.

        agent edit src/headline.js
        tool read tests/headline.test.js
        agent 1 file changed
        $ npm test
        ✓ headline renders
        ✓ health route
        tests 3 passed
        preview after ready · 200
        agent edit src/headline.js
        tool read tests/headline.test.js
        agent 1 file changed
        $ npm test
        ✓ headline renders
        ✓ health route
        tests 3 passed
        preview after ready · 200

        Live run view

        Logs, file changes, tool calls and status stream into the browser as they happen.

        Fri 17:02 · pausedMon 09:14 · resumed
        deploy · restart · retry same run, same state

        Durable approvals

        Runs pause for as long as it takes and survive restarts, deploys and retries.

        Change the headline on the home page

        #42 · feature/headline-4f2a → main · +1 −1

        Merge pull request← that button is yours

        Pull requests, not force pushes

        Approved work lands on a candidate branch and becomes a pull request. Merging is always yours.

        PlannerAnthropicclaude
        AgentOpenAIgpt
        FallbackMockdeterministic

        Bring your own models

        Anthropic and OpenAI with your keys, plus a Mock provider for credential-free demos. Mix providers per node, with fallbacks.

        • Schedules & webhooks
        • CLI
        • Reports & artifacts
        • Teams & audit log
        • Templates
        • Build chat
        • Benchmarks
        • Upload, import or start from a starter
        • Pull request triggers
        • Plugins and Graphify for the agents
        • GitHub Actions builds

        Your keys, your models

        Pick a provider per node and let fallbacks take over when one is unavailable. Start with the Mock provider to try the whole loop without spending a cent.

        Try everything without a key

        A deterministic built-in provider plans, edits files and passes tests. Perfect for demos, CI and learning the workflow editor.

        provider: mock
        cost: 0
        network: none

        VibeYourFlow is coming soon

        Access is by invitation for now. News on the release follows on 1 November 2026. Already have an account? Sign in.

        Release news on

        --Days
        --Hours
        --Minutes
        --Seconds