Agent A accepts the job.
The outcome and acceptance requirements are locked so everyone knows what “done” means.
Give the job to one agent. Hand it to another without starting over. OWP standardizes the contract, evidence, validation and handoff around a job while each provider keeps its own models, tools and internal architecture.
Not another agent platform. OWP is the small shared boundary that lets different agent systems agree on what was requested, what was delivered, whether it passed, and what the next provider inherits.
That makes switching providers a handoff instead of a restart.
You should not need to understand protocols, Git, model routing or agent orchestration to benefit from portability.
The outcome and acceptance requirements are locked so everyone knows what “done” means.
The work comes back with a delivery manifest, Git state and enough evidence for independent validation.
If the first provider stops, passes or needs replacement, the next one inherits the contract and history instead of reconstructing the job from chat.
OWP-Core is intentionally narrow. It describes the portable lifecycle around outcome work and leaves transport, tools, models, economics and provider internals where they already belong.
A protocol is easier to share when it does not ask adopters to surrender the systems they already own.
The portable outcome-work semantics.
Existing ecosystem layers stay where they belong.
The human-facing profile can hide Git and protocol vocabulary behind a simple describe → preview → approve → work → receive flow.
See Simple Flow →Add the provider boundary: inspect a contract, ACCEPT/PASS, execute however you already execute, emit evidence, and hand off when needed.
15-minute provider guide →If a semantic already belongs in A2A, MCP, Git or another open standard, the project explicitly wants that overlap challenged and moved upstream.
Review packet →The seed repository contains a reference core, conformance vectors and a Git-backed two-provider journey. You do not need a wallet, chain, marketplace or hosted service.
Start with OWP-Core + OWP-Git, then implement the smallest boundary your orchestrator needs.
git clone https://github.com/open-work-protocol/open-work-protocol.git
cd open-work-protocol
# Conformance vectors
PYTHONPATH=src python3 -m owp.cli tck
# Real Git-backed two-provider journey
PYTHONPATH=src python3 -m owp.cli golden-demo --jsonSeed-repository proof points, not ecosystem-adoption claims. OWP 0.2 remains experimental.
No. It does not tell a provider which model, framework, prompts, tools, subagents or marketplace to use. It standardizes the portable boundary around the job.
No. A2A can carry agent-to-agent communication; MCP can expose tools and context. OWP describes the outcome-work contract, evidence, validation, customer disposition and handoff that can travel over or alongside those layers.
No. OWP explicitly separates technical validity from customer approval. A delivery can satisfy objective validation while the customer still chooses APPROVE, STEER or REJECT according to the contract and applicable policy.
The canonical repository lives in the Open Work Protocol GitHub organization. The project is designed so no founder API, hosted service, registry, scorer or vendor dependency is required to implement the protocol.
An independent implementation, an unrelated orchestrator adapter, a stricter validator, or a concrete example showing where OWP duplicates or conflicts with a better existing standard.
Imagined by Keeltrace. Built in Git. Designed to be owned by no vendor.
Origin is attribution—not a dependency or control claim. If the founding team disappeared, a conforming implementation should still be able to read the spec, validate work and interoperate.
Implement OWP-Core in another language. Connect an unrelated orchestrator. Attack the ReviewCapsule. Build a stricter Git validator. Show where the protocol duplicates something better.