# OpenWorkProof [Health: Active]

**Category:** ⚖️ Legal  
**Repository:** https://github.com/dengyier/OpenWorkProof  
**GitHub Stars:** 262  
**Views:** 2  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/openworkproof

## Description
Verifiable execution for AI agents: Ed25519-signed work contracts and audit trails.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "openworkproof": {
    "command": "uvx",
    "args": ["openworkproof"]
  }
}
```

## Documentation

## What OpenWorkProof MCP server does

OpenWorkProof MCP server adds a verifiable execution layer to agent, CI, MCP, and multi-agent workflows. It records the intended task, the authority that approved it, decisions made before tool execution, the actions that occurred, the evidence attached to those actions, and the separate outcomes of verification and acceptance.

The resulting record can be exported as a signed evidence package. A third party with the package and public keys can replay the chain without connecting to the original platform. This makes it possible to inspect authorization, execution scope, evidence summaries, verification conclusions, and acceptance status as separate facts.

The protocol does not act as an operating system, model, orchestrator, payment service, legal arbitrator, or automatic acceptance authority. A `VERIFIED` result is not the same as `ACCEPTED`, paid, settled, legally audited, or adopted by an upstream system.

## How it works

A work chain proceeds through these linked objects:

- `WorkOrder` fixes the goal, source version, paths, tools, time limit, and acceptance conditions.
- `CapabilityGrant` gives a signed, limited authority that can be narrowed or consumed but not expanded.
- `PolicyDecision` records whether an action is allowed before the tool runs.
- `ActionReceipt` connects the request, authorization decision, execution result, and evidence digest.
- `VerificationDecision` records an independent assessment of the frozen scope and evidence.
- `AcceptanceDecision` records whether the designated acceptor accepted or rejected the delivery.

Ed25519 signatures, canonical JSON, an append-only ledger, and content digests protect the relationships between these objects. Changing an authorization, receipt, evidence item, public key, or causal parent set causes offline replay to fail.

Human Agency Profiles further restrict the usable capability set. The effective permission is the intersection of the WorkOrder, CapabilityGrant, and active profile. An action marked `reserved` stops before execution with `AGENCY_HUMAN_DECISION_REQUIRED`; an appeal records a signed review request but does not restore or enlarge authority.

## Setup and configuration

The published core can be installed from PyPI:

```bash
python -m pip install openworkproof
owp --help
```

For protocol or adapter development, the repository documents editable installation from source. The command-line interface supports offline surface verification, acceptance-bundle verification, and delivery-case inspection, verification, and export. The repository also lists MCP as an integration entry point and provides an MCP-specific guide.

External acceptance uses a `prepare → sign → commit` flow. The system creates a draft without the acceptor's private key, the designated acceptor signs independently, and the signed object is committed before an acceptance bundle is exported.

## Tools and capabilities

OpenWorkProof MCP server is suited to workflows that need signed authorization, pre-execution policy checks, action receipts, independent verification, separate human acceptance, and offline evidence review. The project also provides CLI, GitHub Action, AgentTeams, DeepSeek Harness, and Python API entry points, although the DeepSeek Harness integration is described as a local candidate rather than a published npm package.

Verification integrity checks can require coverage of the agreed object population and matching failure characteristics for negative controls. If those checks fail, the conclusion is closed as `UNKNOWN` rather than treated as verified.

## Limitations and notes

The protocol proves that recorded evidence follows the defined chain; it does not prove that a business outcome is universally correct. It does not transfer legal responsibility, certify compliance, hold funds, execute payments, or decide whether a customer should accept work. Acceptance remains the responsibility of the WorkOrder-bound human authority.

The README identifies core version 1.4.0 as published to PyPI, GitHub Releases, and the MCP Registry. DeepSeek Harness 0.1.0 is identified as an unpublished local candidate, so its local test results should not be treated as npm availability, customer adoption, or official vendor endorsement.

_Full upstream README: https://allmcps.com/mcp/openworkproof/readme_