CTX Exposes Claude Code's Blind Spot: No Tool Discovery
CTX is a real-time skill and agent recommendation engine for Claude Code. It walks a knowledge graph of 1,768 skills and 443 agents, powered by a Karpathy LLM wiki with persistent memory, and suggests the right tools on the fly.
A developer named Steve Solun just shipped CTX, a real-time recommendation engine that watches Claude Code users and suggests which of 1,768 skills and 443 agents to invoke. This is the first serious attempt to solve the tool-discovery problem for AI coding assistants — and it's a third-party project, not an Anthropic feature.
- Steve Solun released CTX, a real-time recommendation engine that watches Claude Code usage and suggests skills/agents from a knowledge graph of 1,768 skills and 443 agents.
- CTX is powered by a Karpathy LLM wiki with persistent memory, enabling context-aware suggestions that improve over time.
- The key tension: Anthropic hasn't built this into Claude Code, leaving a third-party project to define the standard for tool discovery in AI-assisted development.
Why Is a Third-Party Project Solving Claude Code's Biggest Usability Gap?
Anthropic shipped Claude Code in early 2025 with impressive code generation capabilities, but one glaring omission: there's no built-in system to help developers discover or manage the growing ecosystem of skills and agents. As of April 2026, the community has created over 1,768 skills and 443 agents — a sprawling toolkit that no human can memorize. CTX fills this void by monitoring what a developer is actually doing and suggesting relevant tools in real time. It's powered by a Karpathy-style LLM wiki with persistent memory, meaning it learns from each interaction. The fact that Steve Solun had to build this externally tells me Anthropic either didn't prioritize tool discovery or didn't anticipate how quickly the skill ecosystem would explode.
This is a classic platform failure: the platform owner (Anthropic) ships a powerful engine but neglects the discovery layer, leaving room for third-party middleware. CTX is the first serious attempt to build that middleware for Claude Code.
How Does CTX Actually Work — and Is the Knowledge Graph Enough?
CTX watches your development activity in real time — what files you open, what code you write, what commands you run — and walks a knowledge graph of 1,768 skills and 443 agents. It then recommends the most relevant tools for your current task. The knowledge graph is not static; it's updated via a Karpathy LLM wiki, which means the system can incorporate new skills and agents as the community creates them. Persistent memory means CTX remembers your past choices and adjusts recommendations over time.
The question is whether a knowledge graph is the right representation for this problem. Skills and agents don't exist in isolation — they have dependencies, version conflicts, and usage patterns that a simple graph might not capture. CTX's real test will be whether it can handle the combinatorial explosion of tool combinations as the ecosystem grows to 10,000+ skills.
Who Wins and Who Loses From CTX's Emergence?
| Stakeholder | Win/Lose | Why |
|---|---|---|
| Claude Code users | Win | Reduced cognitive load from tool discovery; faster onboarding to new skills |
| Skill/agent creators | Win | CTX provides a discovery channel for their tools, increasing adoption |
| Anthropic | Lose | Third-party middleware exposes a product gap; users may blame Claude Code for lack of built-in discovery |
| OpenAI (Codex/ChatGPT) | Neutral | CTX is Claude-specific, but the pattern could be replicated for other assistants |
| Cursor/Sourcegraph | Lose | CTX shows that context-aware tool recommendation is viable, raising expectations for all coding assistants |
| Verdict | CTX wins the early lead, but Anthropic must acquire or replicate this within 6 months |
Can Persistent Memory and a Wiki Really Keep Up With a Fast-Moving Ecosystem?
The Karpathy LLM wiki approach is elegant: treat the knowledge base as a living document that the LLM can read and write, rather than a fixed database. But there's a tension between freshness and reliability. If CTX automatically ingests every new skill posted to GitHub, it risks recommending low-quality or malicious tools. If it uses manual curation, it can't keep up with the community's pace. CTX's persistent memory mitigates this by learning which recommendations users actually accept, creating a feedback loop that should filter out noise. But I'm skeptical that a single developer can maintain the quality bar as the ecosystem scales.
What Does This Mean for the Future of AI Coding Assistants?
CTX points toward a future where coding assistants are not just code generators but full development environments with intelligent tool orchestration. The assistant should know not just how to write code, but which linter to run, which testing framework to use, and which deployment agent to invoke — all without the developer explicitly asking. CTX is a step in that direction, but it's still a recommendation engine, not an autonomous orchestrator. The next logical step is a system that not only recommends tools but also chains them together to complete complex workflows. I expect Anthropic to either build a native version of CTX or acquire the project by Q4 2026.
My thesis is clear: CTX is a necessary but insufficient step toward solving the tool-discovery crisis in AI coding assistants, and its real value lies in exposing Claude Code's lack of built-in context awareness. In the short term, CTX will make Claude Code users more productive by reducing the friction of tool discovery. Developers who adopt CTX will find themselves reaching for the right agent or skill faster, leading to fewer context switches and more flow-state coding. The long-term impact is more significant: CTX sets a precedent that third-party middleware can define the developer experience for AI assistants, not just the platform vendors. This is a power shift from platform owners to ecosystem builders. The biggest loser here is Anthropic, because CTX highlights a product gap that should have been obvious from day one. The biggest winner is the community of skill and agent creators, who now have a distribution channel for their work. I predict that Anthropic will acquire CTX or build a competing feature by Q4 2026, because leaving this gap open risks losing developer mindshare to platforms that offer integrated tool discovery. The concrete prediction: Anthropic will announce a native skill recommendation system for Claude Code by October 2026, citing community feedback.
- Anthropic will announce a native skill recommendation system for Claude Code by October 2026, citing community feedback.
- CTX will reach 10,000 stars on GitHub within 60 days of this article, as developers flock to solve the tool-discovery problem.
- At least one major coding assistant (Cursor, GitHub Copilot) will announce a similar context-aware recommendation feature within 90 days, validating CTX's approach.
- CTX reveals that Anthropic shipped Claude Code without a critical discovery layer, leaving a third-party to define the standard.
- The knowledge graph approach is a good start, but may not scale to handle dependency conflicts and versioning as the ecosystem grows.
- Persistent memory is CTX's secret weapon — it learns from user behavior, creating a feedback loop that improves recommendations over time.
- The real battleground is not just tool recommendation but autonomous tool orchestration — chaining skills together to complete complex workflows.
- CTX is a canary in the coal mine: if Anthropic doesn't respond, it signals that third-party middleware will define the developer experience for AI assistants.
Discussion
Add a comment