Skip to content

ctx — Skill & Agent Recommendation and Management for Claude Code

Watches what you develop, walks a knowledge graph of 1,789 skills and 464 agents, and recommends the right ones on the fly — you decide what to load and unload. Powered by a Karpathy LLM wiki with persistent memory that gets smarter every session.

Install

pip install claude-ctx

Optional extras: pip install "claude-ctx[embeddings]" for the semantic backend, pip install "claude-ctx[dev]" for the pytest/mypy/ruff toolchain. After install the ctx-scan-repo, ctx-skill-quality, ctx-skill-health, and ctx-toolbox console scripts are on PATH.

Why this exists

Claude Code skills and agents are powerful, but at scale they become unmanageable:

  • Discovery problem — with 1,700+ skills, how do you know which ones exist and which are relevant to your current project?
  • Context budget — loading all skills wastes tokens and degrades quality. You need exactly the right 10–15 skills and agents per session.
  • Hidden connections — a FastAPI skill is useful, but you also need the Pydantic skill, the async Python patterns skill, and the Docker skill. Nobody tells you that.
  • Skill rot — skills you installed three months ago and never used are cluttering your context. Stale skills should be flagged and archived.

ctx solves all of these by treating your skill library as a knowledge graph with persistent memory, not a flat directory.

What this is

ctx is not a collection of scripts. It is an agent with persistent memory and a knowledge graph.

The core idea comes from Andrej Karpathy's LLM-wiki pattern: instead of re-loading everything from scratch each session, an LLM maintains a wiki it can read, write, and query. The wiki becomes the agent's long-term memory.

ctx applies that pattern to skill management — and extends it with graph-based discovery:

  • A Karpathy 3-layer wiki at ~/.claude/skill-wiki/ is the single source of truth.
  • 2,253 entity pages (1,789 skills + 464 agents) with frontmatter tracking use count, last used date, tags, and status.
  • A knowledge graph (2,253 nodes, 454K edges, 93 communities) connects skills and agents by shared tags, enabling context-aware recommendations.
  • 74 auto-generated concept pages group related skills into named communities (e.g., Security + Testing, Python + API + Database).
  • PostToolUse and Stop hooks update the wiki automatically during each Claude Code session.
  • Skills over 180 lines are converted to a gated 5-stage micro-skill pipeline (956 converted) so the router can load them incrementally.
  • At session start, the skill-router scans your project and recommends the best-matching skills and agents.
  • Mid-session, the context monitor watches every tool call, detects new stack signals, walks the graph, and recommends relevant skills and agents in real time — nothing loads without your approval.

The result: you always know what skills and agents are available for your current task. The graph reveals hidden connections. The wiki learns from your usage. Stale ones are flagged. New ones self-ingest.

Explore the docs

  • Knowledge graph

    2,253 nodes (1,789 skills + 464 agents) connected by 454,719 weighted edges across 93 communities. Ships pre-built in graph/wiki-graph.tar.gz and powers the graph-aware skill recommendations.

    Knowledge graph

  • Dashboard

    ctx-monitor serve opens a local HTTP dashboard with live graph, skill grades + four-signal scores, session timelines, and one-click load/unload. Zero dependencies beyond stdlib.

    Dashboard reference

  • Toolbox

    Curated councils of skills and agents that fire at session-start, file-save, pre-commit, and session-end. Blocks git commit on HIGH/CRITICAL findings. Five starter toolboxes ship out of the box.

    Toolbox overview · Starter toolboxes · Verdicts & guardrails

  • Skill router

    Scans the active repo, detects the stack from file signatures, walks the skill-stack matrix, and loads exactly the skills that apply — no more, no fewer.

    Router overview · Stack signatures · Skill-stack matrix

  • Health & quality

    Structural health checks (missing frontmatter, orphan manifest entries, line-count drift) plus the four-signal quality score (telemetry · intake · graph · routing) that grades every skill A/B/C/D/F.

    Skill health · Memory anchoring · Lifecycle dashboard

  • Releases

    v0.6.1 — MIT, CI-matrixed (Ubuntu + Windows × Python 3.11/3.12), 1,363 tests passing. Ships 10 console scripts including ctx-init and ctx-monitor (local dashboard with graph + wiki + load/unload) plus the 9.6 MB pre-built wiki tarball with 2,253 nodes / 454,719 edges / 93 communities. Hardened across four Strix-audited security findings.

    CHANGELOG · Repository

Principles

  • Foundation first. Data model, CLI, and starter bundles ship before any hook integration. Each phase is independently usable.
  • User-configurable everything. Dedup policy, suggestion loudness, trigger set, council composition.
  • Evidence over opinion. Suggestions cite real usage data plus knowledge-graph edges. No black-box prompts.
  • Token discipline. Every council run honors max_tokens / max_seconds budgets.