Describe the outcome
Give Carole the goal, constraints, and the repository. It turns the request into a shared plan before work begins.
The local-first workspace for agent teams
A software team
you can talk to.
Plan, delegate, review, and ship with a coordinated team of AI agents—without losing sight of what they are doing.
GET STARTED
Install Carole.ai on your computer
1
pip install carole.ai2
caroleaiThen open 127.0.0.1:8000 to sign in or start building. Your workspace runs locally, not on this website.
Public source Local-first Model-flexible
carole / checkout-redesign
Team active
Team conversationShip the checkout redesign
AR
ArchitectMapping the dependency graph
BE
BackendImplementing the API route
FE
FrontendUpdating the workspace view
QA
ReviewerWaiting for implementation
64%
Redesign checkout, keep the existing API, and make failures easy to recover from.
I split this into the payment state machine, UI implementation, and a focused review pass.
Trace checkout flow Build recovery states Review file changes
Agent workflowDelegating and using connected tools
BE
Backend used PostgreSQL through MCP
doneinspect_schema · checkout_sessionsFE
Frontend loaded the UI review skillEditing recovery states and payment summary
workingAsk the team or give new direction…
Model Context ProtocolGive every agent
Give every agent
the right tools.
Connect the systems your team already works in. Carole routes each task to the relevant MCP tools while keeping access and approvals visible.
GitHubPostgreSQLLinearSlackSupabaseFigma
GitHub
PostgreSQL
Docker
Linear
Autonomous Swarm Active
Triaged GitHub issue #402, migrated PostgreSQL schema, containerized microservice & updated Linear sprint.
Live MCP routing
One request enters.
A coordinated team moves it forward.
How work movesFrom intent to reviewed code,
From intent to reviewed code,
in one continuous workspace.
Watch the team work
Specialist agents coordinate in parallel while every thought, tool call, handoff, and background task stays visible.
Review every decision
Approve sensitive actions, inspect exact file changes, and step in wherever judgment matters.
Work stays legibleSee the work.
See the work.
Shape it while it happens.
Carole turns agent activity into a readable timeline. Background jobs, tool calls, handoffs, and code edits stay connected to the conversation that caused them.
- Tool activityCommands, results, and failures with useful context.
- Agent presenceWho is working, what they own, and what comes next.
- Reviewable changesFiles and diffs grouped by the task that changed them.
Live activity 4 events
Architect delegated 3 tasksFrontend, Backend, and Reviewer
Backend completed a tool call
pnpm test checkoutYour approval is neededRun the checkout database migration
Reviewer started a diff review6 files · 157 additions · 24 deletions
Human control, built in
Nothing important hides behind a spinner.
When an agent needs more access, you see the exact action, its reason, and its scope. Approve once, reject, or redirect the team from the same conversation.
Explicit permission boundaries Inspect changes before shipping Stop or redirect work at any time
Public source on GitHubYour team. Your models.
Your team. Your models.
Your machine.
Inspect the system, adapt the workflow, connect the tools you trust, and choose the models that fit each role.
Explore the repositoryterminal
# Install and start your workspace
$ pip install carole.ai
$ caroleai
✓ Workspace ready at 127.0.0.1:8000Tools and infrastructureBuilt on the stack
Built on the stack
you already trust.
Open foundations across runtime, code intelligence, storage, observability, and agent connectivity.
PythonRuntime
Next.jsInterface
TypeScriptInterface
FastAPIRuntime
PostgreSQLData
LanceDBMemory
Tree-sitterCode intelligence
DockerInfrastructure
OpenLLMetryObservability
SentryObservability
MCPProtocol
WebSocketsRealtime
Works across model providers
Questions, answeredBuilt for teams that want
Built for teams that want
clarity and control.
Carole coordinates a team of role-based agents around one shared objective. You can follow each agent, inspect their work, and review the combined result from one workspace.
Yes. Carole is designed for local-first workflows and supports local model providers alongside cloud models, so you can choose where your code and inference run.
Actions that need your judgment appear as clear approval requests with the command, reason, and affected scope before the agent proceeds.
The source is public on GitHub, so you can inspect the project and run it locally. Check the repository for its current reuse terms.