Kinetic Gain MCP Server

mcp-kinetic-gain

One MCP server, all eleven Kinetic Gain Protocol Suite specs + the v0.1.0 implementation tooling. Drop into Claude Desktop, Cursor, or any MCP-compatible client with a single config entry. The agent gains 63 tools (47 spec + 16 implementation-preview, v0.7.0): AEO Protocol, Prompt Provenance, Agent Cards, AI Evidence Format, MCP Tool Cards, AI Tutor Cards, Student AI Disclosure, Classroom AI AUP, Clinical AI Disclosure, AI Incident Card, AI Procurement Decision Card — plus hash attestation (ed25519), audit-stream event composition + chain verification (offline AND live against a running audit-stream-py via AUDIT_STREAM_URL), cross-spec drift detection, Decision Intelligence preview (Decision Card → PolicyBundle, rubric status inference, incident remediation planning). New in v0.7.0: audit_event_emit, audit_events_query, audit_chain_verify_live — Claude can now read and write to the audit-stream spine from inside a chat.

This is the unified read-side companion to kinetic-gain-visualizer: the visualizer renders any of the 11 specs for humans, this server exposes them as callable tools for agents.

Install

npm install -g mcp-kinetic-gain

Or run without installing via npx:

npx mcp-kinetic-gain

Claude Desktop config

Add to your claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\):

{
  "mcpServers": {
    "kinetic-gain": {
      "command": "npx",
      "args": ["-y", "mcp-kinetic-gain"]
    }
  }
}

Restart Claude. All 63 tools appear in the tools panel. Try:

"Use aeo_inspect on https://mizcausevic-dev.github.io to summarize the entity declaration, then use ai_evidence_verify_hash to check the content_hash of an evidence object against my candidate text."

CLI mode (v0.5.1+)

The same binary doubles as a Suite JSON validator outside any MCP host. Useful in CI, pre-commit hooks, or local sanity-checks.

# Validate a single document
npx mcp-kinetic-gain validate path/to/ai-entity.json

# Validate a tree of well-known files
npx mcp-kinetic-gain validate ".well-known/**/*.json"

# Multiple paths or globs
npx mcp-kinetic-gain validate cards/clinical-*.json cards/incident-*.json

# Other commands
npx mcp-kinetic-gain --version
npx mcp-kinetic-gain --help

The CLI auto-detects which Suite spec each file belongs to via its top-level version field (aeo_version, clinical_ai_card_version, aup_version, etc.) and validates it against the same zod schemas the MCP tools use. Output is GitHub-Actions-aware: when GITHUB_ACTIONS=true, failures emit ::error:: workflow commands so they surface as PR annotations.

Exit codes:

Running mcp-kinetic-gain with no arguments still launches the stdio MCP server — existing Claude Desktop / Cursor configs are unaffected.

Tools

Specs with a well-known URL convention (AEO, Agent Cards, Tool Cards) get fetch tools. Specs without one (Prompt Provenance, AI Evidence — these usually travel inline with answers or in repos, not at fixed paths) get parse tools that take a document_json string.

Why one server instead of five?

• One Claude Desktop config entry instead of five
• Cross-spec workflows are atomic — an agent can agent_card_tool_disclosure to find a Tool Card URI, then call tool_card_inspect on that URI in the same conversation, all through one server
• Shared schemas + utilities keep the implementation cohesive
• Deprecation path — if mcp-aeo-server (the AEO-only predecessor) gets retired, the AEO tools live on here with the same names and contracts

Architecture

src/
├── server.ts              # MCP entrypoint, handler dispatch
├── tools.ts               # 18 tool descriptors (JSON Schema inputs)
├── schemas.ts             # zod schemas for every spec
├── common.ts              # fetchJson, canonicalSha256, pretty
└── handlers/
    ├── aeo.ts
    ├── prompt-provenance.ts
    ├── agent-card.ts
    ├── ai-evidence.ts
    └── tool-card.ts

Each handler module is independent and could be split into a separate package if needed.

Hash canonicalization

ai_evidence_verify_hash follows the AI Evidence Format spec's canonical SHA-256 rules:

1. Read content as UTF-8
2. Normalize line endings to \n
3. Strip a single trailing newline
4. SHA-256, lowercase hex, prefixed sha256:

If your candidate_text produces an unexpected mismatch, check CRLF vs LF and trailing newlines first.

Tests

74 unit tests against an in-process Node HTTP server (no external network). Every tool's happy path + at least one error path, plus a live local-HTTP synthetic-index test for incident_index_fetch:

npm install
npm run typecheck
npm test
npm run build

License

This server: AGPL-3.0. Reference implementation. Commercial SaaS hosts must share modifications back.

The specs themselves: MIT. Maximally permissive. Anyone may implement, validate against, or extend any Kinetic Gain Protocol Suite specification. The dual-license split is deliberate: the protocol stays open, the reference server is copyleft.

Kinetic Gain Protocol Suite

Suite hub: suite.kineticgain.com Companion visualizer: kinetic-gain-visualizer Red-team bench: prompt-injection-bench

---

Connect: LinkedIn · Kinetic Gain · Medium · Skills