Clockwork is the state machine.
You bring the intelligence.
Clockwork doesn't run your AI — it coordinates it. When a task is ready, Clockwork emits an event. Your agent executes and posts results back. Clockwork advances the graph. The same model works whether the executor is Claude, GPT, a human, or a vendor.
How it works
Clockwork owns workflow state. Your agents own execution. The boundary is clean.
Clockwork owns
- Workflow state (pending / running / blocked / done)
- Dependency graph — what activates what
- Deadlines and slip propagation
- Human approval gates
- Handoffs between any executor type
You own
- The AI model — Claude, GPT, Gemini, anything
- What actually happens inside each task
- How inputs are processed and results are stored
- Which model handles which task type
- How good the intelligence is
If OpenAI ships a better model tomorrow, Clockwork becomes more valuable — not less.
Because there are now more capable executors available to route to. Clockwork's value is in coordination, not intelligence. AI companies want to focus on intelligence. Clockwork owns the layer they don't want to build: scheduling, dependency tracking, vendor coordination, human approvals, resource allocation.
OpenAI
Wants Clockwork coordinating GPT tasks — not competing with GPT.
Anthropic
Wants Claude to execute. Clockwork tracks state and tells Claude when to run.
Enterprise IT
Wants one coordination layer across agents, humans, vendors, and tools.
The executor abstraction
Every task has an executor. Clockwork doesn't care what kind. A human, a Claude agent, a GPT agent, a vendor portal — they're all just executors. Ownership can transfer mid-workflow as many times as needed.
When a task becomes ready, Clockwork emits an event to the executor. When the executor completes, Clockwork records the output and activates what's next. State lives in Clockwork. Intelligence lives with the executor.
// Task assignment
{
"task": "Summarize findings",
"executor": {
"type": "agent",
"provider": "anthropic",
"agent_id": "research-agent"
},
"status": "waiting"
}
// When Task A completes, Clockwork fires:
{
"event": "task_ready",
"task_id": "B",
"inputs": { "prior_output": "..." }
}MCP server primitives
The Clockwork MCP server exposes an execution coordination API. Every primitive maps to something that needs to happen in a long-running multi-agent workflow.
Any AI that can call an MCP server can use Clockwork as its coordination backend — generating workflows, assigning tasks to agents or humans, querying state, and triggering dependent work.
create_workflow(goal, participants)
Start a new execution graph with any mix of human, AI, or vendor participants.
create_task(workflow_id, executor, description)
Add a task and assign it to any executor type.
create_dependency(from_task, to_task)
Define ordering — Clockwork activates the next task automatically when the prior one completes.
assign_task(task_id, executor)
Transfer ownership mid-flight — human to AI, AI to vendor, or back.
complete_task(task_id)
Mark complete and trigger downstream activation.
delay_task(task_id, reason)
Propagate a slip through the graph — all dependents reschedule automatically.
query_schedule(workflow_id)
Get the current state: what's done, what's active, what's blocked, what's next.
instantiate_template(template_id, inputs)
Launch a parameterized workflow from any published template.
A real mixed-executor workflow
Clockwork manages state and dependencies. Each executor is independent — different providers, different systems, same graph.
“Research 50 HVAC franchises I could buy.”
webhook → Claude runs → posts output back
activates when step 1 completes
different model, same graph
Clockwork parks here until approved
activates on human approval
vendor via vendor portal
Templates become agent skills
Every template in the marketplace is also an agent skill. AI calls instantiate_template() with typed inputs — and gets back a fully-scheduled, dependency-aware workflow.
Marathon Training
Inputs
- race_date
- current_weekly_miles
- goal_time
Output
48-task training workflow
Startup Research
Inputs
- industry
- region
- budget
Output
Multi-executor research workflow
Home Renovation
Inputs
- rooms
- budget
- timeline
Output
Contractor coordination workflow
Works with every AI
Because it's MCP, any AI that supports tool use can call Clockwork — no custom integration needed.
Add Clockwork to your own repo
Connect any MCP-capable agent — Claude Code, Cursor, Windsurf — in three steps. No SDK, no custom integration.
Get an API key
Create a ck_live_… key in the app under Settings → API Keys (shown once). Sign up free if you don't have an account yet.
Add the MCP server
Drop this into your client's MCP config (e.g. claude_desktop_config.json):
{
"mcpServers": {
"clockwork": {
"command": "npx",
"args": ["-y", "@clockwork/mcp"],
"env": {
"CLOCKWORK_API_KEY": "ck_live_…",
"CLOCKWORK_API_URL": "https://app.clockwork-co.com"
}
}
}
}Download the agent skill
The skill teaches your agent when and how to drive the Clockwork tools. Save it to ~/.claude/skills/clockwork/SKILL.md:
mkdir -p ~/.claude/skills/clockwork curl -o ~/.claude/skills/clockwork/SKILL.md \ https://app.clockwork-co.com/clockwork-skill.md
Building a typed client instead? The OpenAPI spec lives at /api/v1/openapi.json.
Connect your AI to Clockwork
We're onboarding AI companies and teams building agent workflows. Request early MCP access below.