CLI · established

JuliusBrussee/caveman

Token-efficiency toolkit for coding agents: a concise-output skill plus a local proxy, recovery store, analyzer, browser, memory, and compression commands.

token efficiencycoding agentscontext management
Installation GuideInstruction Manual

What it contains

A portable output skill and a native/Node CLI that can wrap several coding agents, compress selected request content, retain recoverable originals locally, and report measured savings.

Typical uses

Reduce repetitive agent input and output, analyze token sinks, compress noisy commands, or wrap an agent behind a local provider proxy.

Why people choose it

Choose it when token flow itself is the problem and you want measured, reversible compression rather than simply shortening prompts by hand.

License
MIT for the skill and CLI; BSL-1.1 for the proxy runtime
Maintenance
active
Latest release
See GitHub Releases
Last meaningful update
2026-08-18
Maturity
established
Production readiness
development

Classification

Subject domains

token efficiencycoding agentscontext management

Task categories

input compressionoutput compressionusage analysis

Project phases

setupimplementationdiagnostics

Secondary types

skillMCP serverlocal proxy

Select when

  • Provider-reported input tokens are dominated by tool output, history, or repeated context
  • Want a concise-response skill across skills-compatible agents
  • Need local, recoverable command or context compression

Do not select when

  • Every byte must reach the provider unchanged
  • Cannot run a local proxy or inspect its request-routing behavior
  • Need guaranteed savings on sparse or short inputs
Implementation complexity
high
Setup effort
medium
Learning curve
medium
Confidence
high

Routing heuristic 88.0%

Routing specificity
5/5
Setup simplicity
4/5
Operational maturity
4/5
Composability
4/5
Local control
5/5
Security boundary clarity
4/5
First-party routing documentation
5/5

Strengths

  • Multiple independent modes
  • Local content-addressed recovery
  • Explicit measured-versus-inferred labeling

Poor-fit scenarios

  • Every byte must reach the provider unchanged
  • Cannot run a local proxy or inspect its request-routing behavior
  • Need guaranteed savings on sparse or short inputs

Limitations

  • Proxy mode changes request content and therefore requires careful evaluation
  • Savings depend on payload shape
  • Provider and agent compatibility can change
skills/caveman/SKILL.md

Instruct a compatible agent to answer concisely while preserving code, commands, and error text.

output compression
engine/

Route a supported coding agent through a local proxy that applies content-aware request compression.

proxyprovider routing
packages/cli/src/learn-tui.ts

Read local agent history without editing it, rank token sinks, and show a report with suggested fixes.

analysisreport
Shrink CLI command
packages/cli/

Run a noisy command through recoverable output compression with the caveman shrink command.

command output
Browse CLI command
browse/

Expose a focused, compressed representation of a locally controlled Chrome page.

browseraccessibility tree
Durable memory CLI command
engine/memory/

Remember and recall compact facts while retaining a recovery path for original content.

memoryrecovery
Explore agent
packages/agent/

Use a read-only subagent to locate relevant code with compact path-and-line evidence.

code exploration
engine/toon/

Re-encode suitable structured tool results into a smaller representation and decode them when needed.

structured data
Pixel mode workflow
engine/pixel/

Render exceptionally dense text slabs as images when the measured token gate predicts a benefit.

visiondense text
packages/cli/src/agent-mcp.ts

Expose compression, retrieval, statistics, and TOON operations to an MCP-compatible host.

MCPrecovery
Languages
Go, TypeScript, Shell, Markdown
Frameworks
Local HTTP proxy, MCP, Agent Skills
Runtimes
Node.js 18+ plus downloaded native binaries; Go and pnpm for source builds
Operating systems
macOS, Linux, Windows
Installation methods
npm CLI, signed installer, npx skills, or source build
Interfaces
CLI, local proxy, MCP server, Agent Skill, local report UI
Required credentials
Existing provider or coding-agent authentication is passed through in wrap mode
External services
Selected LLM provider; Chrome only for browse mode
Hardware
No special hardware
Major dependencies
Node.js; optional Chrome; Go/pnpm only when building from source

Compatibility notes

Apple silicon is supported through the packaged CLI path. Start with the output skill or analyzer, then evaluate proxy mode with a disposable session before routing important work through it.

One-sentence semantic summary

Token-efficiency toolkit for coding agents: a concise-output skill plus a local proxy, recovery store, analyzer, browser, memory, and compression commands.

Capability keywords

tokenscompressionproxycavemancontextshrinkrecoveryTOON

User intent phrases

  • Need to reduce coding-agent token consumption
  • Need recoverable compression of noisy context

Negative match phrases

  • Require byte-identical provider requests
  • Need only ordinary command-output filtering

Differentiators

Combines a deliberately terse response skill with a local proxy whose transformations keep recoverable original bytes.

Security considerations

Proxy mode handles provider requests and authentication paths. Review configuration, bind locally, protect the recovery store, and test with non-sensitive traffic first.

Privacy considerations

The analyzer is documented as local and read-only; proxy recovery copies stay on disk. Wrapped requests still go to the provider selected by the user.

Uncertainty

Unresolved items

None documented.

Inference notes

Routing guidance is an analyst synthesis based on the first-party repository and README.

Evidence

  • repository_metadata · Repository identity and current GitHub metadata checked on August 18, 2026.
  • readme · Primary first-party README inspected.

Repository usage guidance

installation summary
Install @caveman-ai/cli globally for the proxy toolkit and/or install the Caveman skill with npx skills
basic usage summary
Run caveman learn first, use the output skill independently, and only then evaluate caveman claude/codex wrap mode
documented entry points
caveman CLI and Caveman SKILL.md
key configuration files
Caveman user configuration and local recovery store
important directories
packages/, engine/, skills/
documentation paths
README.md, INSTALL.md, docs/technical/
example paths
docs/WRAP-BENCHMARK.md

Relationships

overlapping
high
Both reduce agent context cost. RTK concentrates on command output; Caveman spans output style, request compression, recovery, memory, browser data, and proxy routing.

Recommended combinations

Use RTK for deterministic shell-output filtering and Caveman only for distinct response or proxy layers; measure the stack to avoid double compression.

Routing rules

rule_050: Need measured, recoverable coding-agent context compression P950

Rationale

Token-efficiency toolkit for coding agents: a concise-output skill plus a local proxy, recovery store, analyzer, browser, memory, and compression commands.

Required conditions

Provider-reported input tokens are dominated by tool output, history, or repeated context

Preferred

JuliusBrussee/caveman

Fallback

Choose the closest narrowly scoped existing repository only after comparing permissions and workflow ownership.