Landscape

Tools people use to make coding agents cheaper or more reliable, and what keelyard took from them. "Measured" means we ran it; everything else is the project's own claim or a description, not our number. Stars and licenses from GitHub, October 2026.

project	stars	license	technique	effect	in keelyard
rtk	83k	Apache-2.0	per-command output filters	measured: 56,522 → 5,747 tokens on our 8 commands (keelyard: 5,901)	run, filter, run_compact (own code)
repomix	28.8k	MIT	pack repo, compress to structure	claimed large reductions; not measured	code outline, project_map
aider	49.5k	Apache-2.0	ranked repo map	described; not measured	project_map, swarm context packs
serena	30k	see repo	symbol-level find/edit over LSP	described; not measured	find_symbol, read_file ranges
context7	63k	MIT	fetch current library docs on demand	described; not measured	not adopted (network service, off-mission)
code2prompt	7.7k	MIT	repo to prompt with templates	described	not adopted (overlaps repomix idea)
LLMLingua	6.8k	MIT	model-based prompt compression	paper claims up to 20x	not adopted (needs a model at runtime)
caveman	111.8k	Apache-2.0	terse-reply rule	claimed output savings; not measured	one "Be terse" rule
claude-context	12.6k	MIT	vector code search via MCP	described	not adopted (needs embedding service)
ccusage	19k	see repo	usage from local logs	n/a	usage (own code)
superpowers	299k	MIT	skills: plan, debug, review workflows	n/a	workflow pack, lazy (240/181/182 tokens on demand)
spec-kit	141k	MIT	spec → plan → tasks	n/a	proof add requirements + plan workflow
claude-task-master	28k	see repo	task lists for agents	n/a	covered by proof requirements
claude-code-router	37.7k	MIT	route requests to cheaper models	n/a	route_task, swarm escalation
playwright-mcp	38k	Apache-2.0	browser control for agents	n/a	proof_screenshot (optional extra)