okf-mcp: Skill for Claude Code

Publish OKF Knowledge

Follow this workflow exactly:

download -> enrich -> server dry-run -> explicit confirmation -> submit -> MCP verify

Boundaries

• Read and query published knowledge through the configured remote MCP server.
• Download, dry-run, and submit through the okf knowledge CLI.
• Do not call private HTTP endpoints directly.
• Do not use MCP write tools for this workflow.
• Do not install or upgrade the CLI automatically.
• Never put credentials in this Skill, the OKF bundle, command arguments, or logs.
• A successful dry-run is not approval to submit.

1. Check The Runtime

Before changing knowledge, run:

command -v okf
okf --version
okf knowledge --help

If okf is missing or lacks the knowledge command, stop and report that the agent runtime must be updated. Do not use curl, npm install, or direct HTTP requests as a workaround.

2. Confirm The Target

Confirm the server, bundle, local workspace, and intended knowledge scope. Use MCP list_bundles when the bundle is not known. Never guess a bundle ID.

3. Download The Current Snapshot

okf knowledge download \
  --url https://knowledge.internal.example \
  --out ./okf-work

When the hosted service contains more than one writable bundle, also pass the confirmed --bundle <id> value.

Treat the returned revision as mandatory baseline state. Work only inside the downloaded workspace. Do not wrap its files in an extra bundle directory.

4. Enrich Locally

Apply OKF v0.2 rules:

• every non-reserved Markdown concept has YAML frontmatter and non-empty type;
• preserve unknown frontmatter and extension fields;
• prefer native v0.2 generated, sources, lifecycle, and relation fields;
• use safe bundle-relative Markdown paths;
• do not invent facts, schemas, URLs, sources, or identifiers;
• keep concept paths stable during updates;
• removing a file expresses an intended deletion and must be called out clearly.

Local validation is useful but does not replace the server preview.