mcp-server-akf
mcp-name: io.github.HMAKT99/akf
MCP (Model Context Protocol) server that exposes 10 AKF tools to AI agents.
Any MCP-compatible client (Claude Desktop, Claude Code, Cursor, Windsurf, etc.)
can check, stamp, validate, audit, and scan files using the Agent Knowledge Format.
A stamp costs ~15 tokens. Re-verifying costs 15,000. Agents stamp what they
verify; the next agent calls check_file and builds on it instead of redoing
the work.
Installation
pip install mcp-server-akf
Configuration
Claude Desktop
Add to ~/Library/Application Support/Claude/claude_desktop_config.json:
{
"mcpServers": {
"akf": {
"command": "python",
"args": ["-m", "mcp_server_akf"]
}
}
}
Cursor
Add to .cursor/mcp.json in your project root:
{
"mcpServers": {
"akf": {
"command": "python",
"args": ["-m", "mcp_server_akf"]
}
}
}
| Tool | Description |
|---|
check_file | One-line trust check — can an agent build on this file without re-verifying? (OK / LOW / STALE / UNSTAMPED) |
create_claim | Create an AKF claim with trust metadata |
validate_file | Validate an .akf file against the spec |
scan_file | Security scan any file for AKF metadata |
trust_score | Compute effective trust score for a claim |
stamp_file | Stamp trust metadata onto any file (20+ formats) |
audit_file | Compliance audit (EU AI Act, SOX, HIPAA, GDPR, NIST, ISO 42001) |
embed_file | Embed AKF metadata into DOCX, PDF, HTML, images, etc. |
extract_file | Extract AKF metadata from any supported format |
detect_threats | Run 10 AI-specific security detections |
Quick usage
Once configured, ask your AI agent:
"Stamp report.docx with AKF trust metadata, confidence 0.9, evidence 'quarterly review complete'"
The agent will call stamp_file through MCP and attach provenance automatically.
You can also run the server directly for testing:
On mcp 2.x, a tool result for a file (any tool called with path) carries that file's verification state in _meta under the vendor key dev.akf/trust:
"_meta": {
"dev.akf/trust": {
"v": "1",
"status": "OK",
"trust": 0.7447,
"evidence": ["human_review"],
"agent": "claude-code",
"age_days": 4
}
}
status is computed live (OK / LOW / STALE / UNSTAMPED), so a client sees whether the file changed since it was verified. The key is absent when the result isn't about a file. Clients that don't know the key ignore it.
To annotate your own server's results:
from mcp_server_akf.meta import TRUST_META_KEY, trust_meta
meta = trust_meta(path)
return CallToolResult(content=[...], _meta={TRUST_META_KEY: meta} if meta else None)
To read it in a client:
result = await client.call_tool("extract_file", {"path": "report.md"})
state = (result.meta or {}).get("dev.akf/trust")
if state and state["status"] == "OK":
...
Multi-Agent Support
The MCP server works seamlessly with multi-agent orchestration platforms:
- Agent Identity: Create and verify agent cards via
akf agent create / akf agent verify
- A2A Protocol: Export/import agent cards for cross-platform discovery via
akf agent export-a2a
- Team Certification: Use
akf certify --team for per-agent trust breakdowns in CI
- Delegation: Trust ceilings cap delegate output when agents hand off work
- Team Streaming: Multi-agent sessions with per-agent claim attribution
Supports: Claude Agent Teams, Copilot Cowork, Codex multi-agent, and any A2A-compatible platform.
Requirements
- Python >= 3.10
akf >= 1.0.0
mcp >= 1.0.0