Think of SpecWerk as an OS for AI coworkers: they read specs, call tools, and leave an audit trail.
Spec-first AI workflows for real work.
Replace manual back-office work with spec-driven automations your engineers trust and your execs can approve.
No five-LLM trench coat. Just specs, tools, and a thin agent layer.
Built for teams who want AI employees they can audit, not AI 'agents' they have to believe.
Jobs, Specs, Tools, Worker
SpecWerk is an OS for AI workers. You don't wire up agents; you hire jobs by writing specs.
The promise: "Reconcile yesterday's invoices every morning at 9am and email finance a summary."
The contract for that job: inputs, steps, tools, outputs, checks. A YAML file your team can read and review.
Deterministic services (internal or MCP) the job is allowed to use. They do the actual work.
SpecWerk runtime + LLM that executes the spec and narrates what happened. One worker, no agent telephone game.
AI agents that behave like software, not improv actors.
SpecWerk is an open-source runtime for AI workflows that behave like well-structured software, not ad-hoc chat logs.
- Hire jobs by writing specs, not wiring up agents.
- Describe workflows in simple, human-readable specs.
- Run work through deterministic tools and MCP servers.
- Use a small agent step only where judgment or language is needed.
- Inspect every run in SpecWerk Studio's Kraftwerk-style control panel.
Result: AI that can replace parts of a team without turning into a black-box transcript.
Why SpecWerk?
Where it sits between agent frameworks and classic automation.
- Conversation-first, multi-agent role-play
- Logic buried in prompts and chat logs
- Hard to replay or audit
- Deterministic but rigid scripts
- All glue code written by hand
- Slow to adapt to new workflows
- Job-spec OS: hire jobs by writing specs
- Single worker, no agent-to-agent chat
- Spec-first, human-readable workflows
- Thin agents for judgment and language only
- Deterministic tools via MCP or internal code
- Studio UI for runs, logs, and debugging
How it works
Hire a job by writing a spec: describe the promise, then the workflow as ordered steps in a YAML file your team can read.
Implement tools as deterministic services, or let Tool Forge scaffold MCP servers and tests from a tool spec.
SpecWerk executes the spec deterministically, calling tools and agents exactly as defined.
Review runs, logs, and summaries in SpecWerk Studio's clean, Kraftwerk-inspired UI.
Who it's for
Everyone else is racing to build more agents, more canvases, more connectors. SpecWerk is for the teams who already know what job they want done, and just want it done the same way every day, with a clear spec and a log.
- Hire AI workers for recurring jobs: reconciliation, approvals, reporting.
- Keep a clear, auditable trail for every run.
- Offer a safe job-spec OS as a platform primitive.
- Standardize on specs, tools, and logs instead of ad-hoc "AI experiments."
- Upgrade scripts and bots into job-driven automations.
- Version specs, tools, and run logs alongside your code.
Get started
- Read the overview
- Run the sample workflow
- Swap in your own tools
pip install specwerk
specwerk run specs/invoice-recon.yaml -p '{"date":"2025-01-01"}'