@axsept/mcp-server

The reference Node.js MCP server for the SkillChain integration protocol. Captures human–AI interaction events from Claude Code (or any MCP-aware client) and ships them to the SkillChain BFF.

This package is the canonical rich-signal reference implementation per §5 of the integration protocol. It exposes three MCP tools that the host LLM can invoke during a conversation to record the skills the user demonstrated.

Install

When used as part of the skillchain Claude Code plugin, this server is launched automatically via npx. To install standalone for development or non-plugin use:

npm install -g @axsept/mcp-server

Configure

Set the SkillChain API key as an environment variable (or via Claude Code plugin config):

export SKILLCHAIN_API_KEY="sk_user_..."
# optional — defaults to https://skillchain.tech
export SKILLCHAIN_API_BASE_URL="https://staging.skillchain.tech"
# optional — when set, validates events but does not POST to BFF (for local testing)
export SKILLCHAIN_DRY_RUN="1"

Run

skillchain-mcp-server   # starts the MCP server on stdio

Or directly via npx:

SKILLCHAIN_API_KEY=sk_user_... npx @axsept/mcp-server

Tools

log_activity

Record a discrete activity the user demonstrated a skill on. Called once per meaningful work unit — not once per message.

Input fields (Zod-validated, see LogActivityInputSchema):

• description (required) — short prose describing what the user did
• skills (required, 1–50) — one or more skills with name, proficiency (0–1), optional taxonomy_node_id, evidence_snippet
• human_contribution (optional) — prompt_quality, directing_pattern, iteration_count, accepted_unchanged, modifications_made, rejections — see protocol §2 for definitions
• tools_used (optional) — up to 20 tool/framework/language names

record_session_summary

Record a session-wide summary at natural session boundaries (long pauses, topic changes, completion of significant work). Captures the user's directing/orchestration skill across the whole session.

Input fields:

• summary (required) — session-level prose
• skills_demonstrated (required, 1–50) — skill names
• notable_decisions (optional) — short descriptions of decisions made
• session_duration_minutes (optional)

get_profile_snapshot

Read-only. Returns the user's current SkillChain signature (VDS, CSS, top skills, freshness). Used by Claude to avoid re-logging the same skills repeatedly and to personalize coaching when the user asks.

What the server does

For each log_activity or record_session_summary call:

1. Anonymize all text fields locally via @axsept/anonymizer — PII patterns (emails, phone numbers, SSNs) and secret patterns (AWS keys, GitHub tokens, JWTs, etc.) are stripped before any network call. Skill vocabulary is preserved.
2. Compose a HumanAIEvent that conforms to the published @axsept/event-schema. Includes a deterministic event_id derived from the session ID, monotonic sequence number, and content hash — so retries dedupe at the BFF.
3. POST to the SkillChain BFF with Authorization: Bearer <key> and a X-Request-Id header. Handles 429 with exponential backoff (1s → 60s cap, 5 attempts), halts on 401, retries 5xx.

The server itself does not call any LLM. All extraction work is done by the host model (Claude) as part of the conversation, per the protocol's rich-signal contract.

License

Apache-2.0