An agent that receives too many logs may lose room for its next step. A research agent may need repeatable domain instructions, while a Claude Code team may want extensions without building each one itself. Three GitHub projects address those separate needs: Context Mode filters tool output, K-Dense provides scientific skills, and Anthropic maintains a Claude Code plugin directory. These are source-based picks, not a ranking or a hands-on test.
Filter tool output before it fills the context
Context Mode places a sandbox between a coding agent and tools such as a browser or command line. It can return selected results or let the agent search saved material instead of pulling an entire log or page into what the model sees for that task. The project's README reports reductions of 94% to 100% across its examples, including a Playwright snapshot reduced from 56.2 KB to 299 bytes. Those are maintainer benchmarks, not independent tests, and results will depend on the output a team routes through the tool. Setup also varies by agent host: some support automatic hooks, while others need a routing file. The repository's Elastic License 2.0 prohibits providing the software to third parties as a hosted or managed service when users receive access to a substantial set of its features, so teams considering that model should read the full terms.
Start research agents with domain-specific instructions
Scientific Agent Skills packages documented procedures for research tools, data sources and workflows across fields such as biology, chemistry and data analysis. Its maintainers list 177 skills and installation paths for compatible agent hosts. Version 2.72.0, released October 1, removed four document-handling skills that had been bundled under Anthropic's license; users who need the DOCX, PDF, PPTX or XLSX workflows must obtain them separately. The repository is MIT-licensed, but K-Dense says individual skills can carry different terms in their own SKILL.md files. Dependencies and access rules for scientific data sources also vary, so a team should check the specific workflow before relying on it. A skill is guidance for an agent, not a validated scientific result.
Add extensions inside Claude Code
Anthropic's Claude Code Plugins Directory separates plugins built by Anthropic from external plugins supplied by partners and the community. Users can browse in Claude Code's Discover screen or install by plugin name. That makes the directory relevant to teams already using Claude Code; it is not a general plugin manager for other agent hosts. The repository is Apache-licensed, but its README directs users to each plugin's own license. Anthropic also warns that it does not control the software included in plugins or verify that it will work as intended or remain unchanged. A directory listing is a discovery route, not a security review.
The useful distinction is the layer each project addresses: Context Mode filters tool output, Scientific Agent Skills supplies research procedures, and the plugin directory helps Claude Code users add extensions. Before adopting one, check the installed version, the agent host and any service or data-access terms that apply.
Sources
- Context Mode README at inspected commit and Elastic License 2.0
- Scientific Agent Skills README at inspected commit, v2.72.0 release notes and repository license
- Claude Code Plugins Directory README at inspected commit and repository license
