Ctx filters tools upstream to cut token costs before context bloats
Ctx is an open-source tool that watches a repo and task, then recommends a small, scored bundle of skills, agents, MCP servers, and harnesses to load — preventing irrelevant tools from bloating context in the first place.
Score breakdown
Ctx shifts token-cost management to the pre-session stage, preventing context bloat from ever occurring rather than cleaning it up after the fact.
- 01Ctx selects a small, top-scored bundle of tools before a session starts, rather than compressing tokens inline.
- 02It watches the repo and current task, then walks a graph of available tooling to make recommendations.
- 03The author curated a backing dataset of 91k+ skills, 467 agents, 10.7k MCP servers, and 207 harnesses.
Ctx addresses a problem the author describes as developer "hoarding": over time, AI coding sessions accumulate skills, agents, MCP servers, harnesses, prompts, repo instructions, and local scripts, leaving the model with far too many options to choose from. Rather than compressing what's already in context, Ctx intervenes before the session starts — watching the repo and task, traversing a graph of available tooling, and surfacing a small, top-scored bundle of only the relevant skills, agents, MCP servers, and harnesses.
The goal, as stated, is to save tokens without requiring users to manually test and compare thousands of possible tooling combinations.
To ensure recommendations are grounded and repeatable rather than hallucinated, the author manually curated a dataset comprising 91k+ skills, 467 agents, 10.7k MCP servers, and 207 harnesses, using AI to generate it and then revising it for accuracy. The author explicitly positions Ctx as upstream and complementary to existing tools: `rtk` for compressing terminal output, terse-output tools for shorter responses, and Ctx for selecting the right tools to load in the first place. The goal, as stated, is to save tokens without requiring users to manually test and compare thousands of possible tooling combinations.
Key facts
- 01Ctx selects a small, top-scored bundle of tools before a session starts, rather than compressing tokens inline.
- 02It watches the repo and current task, then walks a graph of available tooling to make recommendations.
- 03The author curated a backing dataset of 91k+ skills, 467 agents, 10.7k MCP servers, and 207 harnesses.
- 04The dataset was AI-generated but manually curated and revised for accuracy.
- 05Ctx is designed to complement — not replace — tools like `rtk`, `caveman`, and `ponytail`.
- 06The author describes the core problem as developers accumulating too many tools, leaving models with too many options.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 17, 2026 · 10:39 UTC. How this works →