Every session starts cold.
The goal and constraints get reconstructed from conversation instead of carried with the work.
v0.0.0 · Open protocol · Markdown-first · MCP-ready
Shape the work before the model touches it.
HAWP gives AI agent work a stable shape for intent, context, mission, constraints, and output, so agents can execute with less drift and hand off without starting from zero.
Apache-2.0 · Go CLI · Static Markdown artifacts · No lock-in
WORK
Drag a field or focus it and use arrow keys to reshape the work packet
The failure mode
AI work often loses shape between the idea and execution.
The expensive part is not always generation. It is re-explaining intent, rediscovering constraints, recovering decisions, and correcting drift across sessions and tools.
Execution drifts.
Agents optimize locally while the original mission, boundaries, or expected output fade from context.
Handoffs lose decisions.
The next model, tool, or person re-derives what was already known and repeats avoidable work.
The protocol
Five fields. One portable shape for the work.
HAWP keeps the protocol intentionally small. The artifact remains readable by humans, useful to agents, and portable across tools.
Input
The request as received, before interpretation changes it.
Context
The smallest useful background the agent needs to reason correctly.
Mission
The concrete objective in one clear, execution-ready statement.
Constraints
Hard boundaries, quality bars, dependencies, and things that must not change.
Output
The observable end state that defines what “done” actually means.
$ hawp work new
✓ work item scaffolded
input → captured
context → bounded
mission → explicit
constraints → preserved
output → verifiable
$ hawp mcp
✓ hawp_work_intake · compound shaping
✓ hawp_work_doc · canonical work paths
Tool-independent by design
The work survives the agent.
HAWP uses Markdown artifacts and exposes search and workflows through MCP, so the same project shape can move across coding agents without getting trapped in one vendor's memory or chat history.
In v0.0.0 , compound intake can retrieve relevant context, reshape the request through a local LLM, and return structured mission, constraints, output specification, and done signal states without creating a work item automatically.
Claude CodeGitHub CopilotCodexCursorContinue
v0.0.0 repository evidence
Less context. Faster retrieval. Explicit uncertainty.
Current repository evidence covers search retrieval, context shaping, and structured intake shaping. These are HAWP-local measurements intended to make workflow claims inspectable.
Benchmarks are project-local measurements, not universal model performance claims.
How it fits
Shape, execute, and hand off.
Lock intent before execution.
Capture the request, relevant context, mission, constraints, and expected output as a work artifact.
Give the agent bounded context.
Search the kit and work index, then pack only the useful context instead of repeatedly loading everything.
Carry the shape forward.
The next agent or session reads the same portable state instead of reconstructing the project from conversation history.
Latest release
Get HAWP for your platform.
Download the HAWP CLI from the latest GitHub release. The version and asset links update from the release feed when a new build is published.
Checking GitHub for the latest HAWP release.
FAQ
Small protocol, deliberate boundaries.
What is HAWP?
HAWP is an open-source Human-AI Workflow Protocol for shaping AI agent work before execution. It carries intent, context, mission, constraints, output, evidence, and handoffs in portable work artifacts.
Is HAWP an AI agent framework?
No. HAWP is a workflow protocol and CLI for shaping and carrying work. It does not replace your agent runtime, orchestrator, or model.
Does HAWP require a database or cloud service?
No. The work model is Markdown-first and repository-local. Search indexing can be generated locally, so the protocol remains a set of portable files rather than a hosted dependency.
Why does HAWP use MCP?
MCP gives connected agents a standard way to search the HAWP kit and work index without forcing the protocol into a specific editor or vendor.
Can I use HAWP with more than one coding agent?
Yes. Provider guides in the repository cover Claude Code, GitHub Copilot, Codex, Cursor, and Continue.
Give AI agent work a shape that survives the session.
HAWP is open source, Apache-2.0 licensed, and designed to stay readable even when the agent changes.