An open-source developer project called Caveman has expanded from a prompt skill that makes coding agents answer more concisely into a broader toolkit for reducing what agents read and write. The project supports Claude Code, Codex, Gemini, Cursor and dozens of other agent environments, according to its GitHub documentation.
The original skill instructs an agent to use compressed, so-called caveman-style language while preserving code, commands and error text. Users can select several levels of compression or turn the behavior off. The maintainers describe readability and faster interaction as the main benefit, with lower output-token use as a secondary result. They also warn that the skill adds prompt overhead and may increase total usage on workloads where responses are already short.
A newer component, Caveman Proxy, targets input rather than output. It runs locally and forwards requests to the provider selected by the user, compressing request payloads before transmission. The repository says recovery copies remain on the user's disk and that provider credentials pass through to the chosen service. Its engine identifies payload types, routes them through specialized compressors and can select context using relevance, recency and error signals while retaining chronological order.
The developers report that the proxy used 33.2% fewer provider-reported input tokens than direct Claude Code in a pinned 54-run benchmark, while passing 18 exact-answer checks. That result is a project benchmark, not an independent audit or a provider invoice. The documentation explicitly distinguishes controlled counterfactual measurements from inferred results produced by local use.
Caveman also includes an MCP server with compression, retrieval, statistics and encoding tools. A local analysis command can examine months of agent history, identify high-token flows and estimate how proposed changes might have affected prior sessions. The project says this analysis is read-only and does not require an account. Another feature converts installed skill instructions into image pages when the conversion is estimated to reduce context size, leaving frontmatter as text so discovery still works.
Licensing differs across the package. The skill, command-line interface and client-facing components are under MIT terms, while engine-linked runtime code uses the source-available BSL 1.1. Anonymous command and token-count telemetry is enabled by default, the documentation says, but users can disable it with a command or a standard do-not-track setting. Together, the additions position Caveman as an optimization layer around existing agents rather than a replacement for them.
The command-line package can also wrap several agent programs without rewriting their existing configuration files. For unsupported frameworks, the documentation offers base-URL integration recipes so applications can send provider requests through the local proxy.
