I built Claude sub-agents that keep my context clean (github.com)

House Agents (v0.2.0‑beta) is a set of Claude Code “sub‑agents” — house-research, house-git and house-bash — that run in their own Claude instances and context windows to quarantine heavy operations (large file searches, long docs, build logs, diffs) and return condensed summaries to your main conversation. Measured in production, the agents processed ~133.6k tokens and returned ~4.5k (examples: house-research 70.1k→3.2k, house-git 42.9k→~500), trimming 95–99% of token bloat. Installation is via cloning the repo and copying .claude agent files either project‑level (.claude/agents) or user‑wide (~/.claude/agents), and each agent is an .md file with YAML frontmatter (tools, model, metadata) plus a system prompt you can customize. Technically, each sub‑agent is spun up with its own system prompt and authorized tools (Read, Grep, Bash, etc.), executes workloads off‑context, and returns actionable summaries (file:line refs, impact ratings, test summaries) rather than raw outputs. Tradeoffs: ~2–5s extra latency per invocation, no shared memory between calls, and more API calls (but often lower total token cost). A known limitation: sub‑agents cannot access MCP tools due to bug #7296, blocking planned agents (mcp, data, vision). The project is extensible (edit tool access, token budgets, output formats) and aimed at large codebases, long docs, CI/CD and workflows where keeping the main Claude session focused is critical.