HARP is a vendor-neutral protocol layered on HTTP/OpenAPI that makes services equally usable by humans and AI coding agents. It defines a three-tier compliance ramp — L1 Baseline (minimal discovery, stable error envelopes, traceability), L2 Agent-ready (structured success envelopes, actions, idempotency, ETags, audit), and L3 Autonomous-ready (dry-run, two-phase destructive commits, recipes, capability negotiation, long-running jobs) — so teams can adopt incrementally without rewriting existing APIs.
Pre-1.0 / experimental. Protocol version 0.1.
The draft spec lives under spec/.
HARP is meant to be consumed through a local harness, not by asking an agent to guess HTTP routes from prose. The planned flow is:
harp agent install claude-code --scope user
harp harness add service-a https://api.service-a.com --auth-env SERVICE_A_TOKEN
harp harness plan service-a "get entity a" --format json
harp harness call service-a get_entity --params '{"entity_id":"a"}' --format jsonClaude Code is the first planned adapter. The vendor-neutral piece is harp harness: a local registry and execution layer that owns discovery, OpenAPI operation resolution, auth references, safety policy, and HARP response handling.
See docs/agent-harness-user-journey.md for the draft client-side and agent-side user journey.
Coming with v0.1.0.
See CONTRIBUTING.md.
MIT — see LICENSE.