Best open-source AI repos for cutting coding agent token usage (2026)

What is the best open-source tool to cut Claude Code and coding agent token usage?

The best open-source AI repos for cutting coding agent token usage right now are Caveman, Headroom and Context Mode. Caveman is the quickest win (one skill makes the agent answer tersely and read less); Headroom and RTK compress tool, log and command output before it reaches the model; Context Mode keeps raw tool output out of the context window and saves session memory in SQLite.

Comparison

Repo	What it does	Stars	Start here
Caveman JuliusBrussee/caveman · Go	Caveman is a skill and proxy for coding agents that cuts tokens by making the agent talk tersely and read less, while code stays exact.	114,855	Cut your coding agent's token use with one skill Beginner, 5 minutes
Headroom headroomlabs-ai/headroom · Python	Headroom is an open-source tool that compresses tool outputs, logs, files, RAG chunks and chat history on your machine before they reach an LLM, available as a Python or TypeScript library, a local proxy, and an MCP server.	77,499	Cut Claude Code token use with one wrap command Intermediate, 30 minutes
Context Mode mksglu/context-mode · TypeScript	Context Mode is an MCP server and hook set for AI coding agents that keeps raw tool output out of the context window, persists session memory in SQLite, and routes work across many clients.	31,762	Keep large tool output out of Claude Code's context Beginner, 15 minutes
RTK rtk-ai/rtk · Rust	RTK is a single Rust binary that filters and compresses the output of 100 plus common dev commands before it reaches an AI coding agent's context, with hooks for many agents and a savings dashboard.	n/a	Cut your coding agent's token use from command output Beginner, 10 minutes
Repo Context MCP nduc99911/repo-context-mcp · TypeScript	repo-context-mcp is a local MCP server and CLI that gives AI coding agents a repository map, fast code search and token-budgeted context packs instead of dumping whole repos into prompts.	n/a	Give your coding agent a repo map and code search Intermediate, 15 minutes
Cache Tax karanb192/cache-tax · TypeScript	Cache Tax is a Claude Code mod that keeps the prompt cache warm during breaks and shows the estimated rewrite cost before a cold send.	270	Keep Claude Code's cache warm during a break Intermediate, 15 minutes

When to pick which

Caveman JuliusBrussee/caveman

Caveman is a skill and proxy for coding agents that cuts tokens by making the agent talk tersely and read less, while code stays exact. The README reports 33.2% fewer input tokens across 54 Claude Code runs and links an Adobe Research paper on caveman-style output; it also links a page listing every case where caveman loses. It installs as a skill, a Claude Code plugin, a Gemini CLI extension or an all-agents installer, with an optional free Auto model router.

Best for: Developers who pay per token for Claude Code, Codex or Gemini CLI and want shorter answers and smaller context without losing correctness.

Needs: A coding agent such as Claude Code, Codex or Gemini CLI; Node.js for the npx installs (22.13+ for the all-agents installer)

Start here: Cut your coding agent's token use with one skill Beginner, 5 minutes, works with claude-code, codex, gemini-cli

Install for every agent with npx skills add JuliusBrussee/caveman -g
Or in Claude Code: /plugin marketplace add JuliusBrussee/caveman then /plugin install caveman@caveman
Type /caveman ($caveman in Codex) if it does not start on its own; /ultracave is even shorter
Say stop caveman to go back to normal

All 3 use cases GitHub

Headroom headroomlabs-ai/headroom

Headroom is an open-source tool that compresses tool outputs, logs, files, RAG chunks and chat history on your machine before they reach an LLM, available as a Python or TypeScript library, a local proxy, and an MCP server.

Best for: Developers and platform teams who run coding agents or LLM apps and want fewer tokens sent and billed per request.

Needs: Python 3.13 with uv, or pip, to install the headroom-ai package; An account with an LLM provider such as Anthropic or OpenAI; A coding agent or app that sends prompts through an API

Start here: Cut Claude Code token use with one wrap command Intermediate, 30 minutes, works with claude-code

Install the CLI with uv tool install --python 3.13 "headroom-ai[all]"
Run headroom doctor to confirm routing works
Start a wrapped session with headroom wrap claude
Work as usual, then check results with headroom savings

All 4 use cases GitHub

Context Mode mksglu/context-mode

Context Mode is an MCP server and hook set for AI coding agents that keeps raw tool output out of the context window, persists session memory in SQLite, and routes work across many clients.

Best for: Developers who use AI coding agents on large codebases and notice their context window filling up during long sessions.

Needs: Claude Code v1.0.33+ for the plugin install, or Node.js >= 22.5 (or Bun) for the Gemini CLI setup; An existing AI coding agent session in a project you want to work on

Start here: Keep large tool output out of Claude Code's context Beginner, 15 minutes, works with claude-code

Run /plugin marketplace add mksglu/context-mode in Claude Code.
Run /plugin install context-mode@context-mode.
Restart Claude Code or run /reload-plugins.
Run /context-mode:ctx-doctor and confirm all checks show [x].

All 5 use cases GitHub

RTK rtk-ai/rtk

RTK is a single Rust binary that filters and compresses the output of 100 plus common dev commands before it reaches an AI coding agent's context, with hooks for many agents and a savings dashboard.

Best for: Developers using coding agents who want to cut token usage from noisy command output.

Needs: macOS, Linux or Windows; A supported coding agent

Start here: Cut your coding agent's token use from command output Beginner, 10 minutes, works with claude-code

Install: brew install rtk on macOS or Linux, winget install rtk-ai.rtk on Windows, or the install.sh script
Check it: rtk --version, then rtk gain must show the savings dashboard, not command not found
Hook it into Claude Code (or Copilot) with rtk init -g, then restart the tool
Test with git status in the agent; it is rewritten to RTK git status

All 3 use cases GitHub

Repo Context MCP nduc99911/repo-context-mcp

repo-context-mcp is a local MCP server and CLI that gives AI coding agents a repository map, fast code search and token-budgeted context packs instead of dumping whole repos into prompts.

Best for: Developers using Codex, Claude Code, Cursor or Cline on large repositories or monorepos.

Needs: Node.js 18+

Start here: Give your coding agent a repo map and code search Intermediate, 15 minutes, works with codex, claude-code, cursor

Install Node.js 18+
Add a server with command npx, args -y Repo Context MCP, and REPO_CONTEXT_ROOT set to your repo
Restart your MCP client
Ask the agent to call repo_map before editing

All 3 use cases GitHub

Cache Tax karanb192/cache-tax

Cache Tax is a Claude Code mod that keeps the prompt cache warm during breaks and shows the estimated rewrite cost before a cold send.

Best for: Developers who use Claude Code in long sessions and want to avoid paying to rewrite a large cached context after a break.

Needs: Claude Code 2.1.287 or later; A one-hour cache (included subscription usage defaults to one hour; API-key, usage-credit and cloud-provider sessions default to five minutes and need promptCacheTtl set to "1h"); Claude Code left running during the break

Start here: Keep Claude Code's cache warm during a break Intermediate, 15 minutes, works with claude-code

Run claude plugin marketplace add karanb192/claude-code-mods in a terminal.
Run claude plugin install cache-tax@claude-code-mods, then restart Claude Code or run /reload-plugins in an open session.
If you use API-key, usage-credit or cloud-provider billing, set promptCacheTtl to "1h" so the main cache lasts an hour.
In a warm session, run /keepwarm 90m before your break, or /keepwarm for the six-hour default.

All 3 use cases GitHub

Use cases for cutting coding agent token usage

Cut your coding agent's token use with one skill Caveman
Shrink what your agent reads: logs, test output and web pages Caveman
Add Caveman to Gemini CLI as an extension Caveman
Cut Claude Code token use with one wrap command Headroom
Shrink messages in a Python app before each LLM call Headroom
Route any OpenAI-compatible client through a local proxy Headroom
Keep large tool output out of Claude Code's context Context Mode
Count or summarize code with one script, not many reads Context Mode
Pick up a coding task after the chat compacts Context Mode
Cut your coding agent's token use from command output RTK
Set up RTK for Codex, Gemini CLI, Cursor and other agents RTK
Use the Apache-only RTK edition for license-sensitive teams RTK
Give your coding agent a repo map and code search Repo Context MCP
Build a token-budgeted context pack for a prompt Repo Context MCP
Generate repo maps as JSON for CI and scripts Repo Context MCP
Keep Claude Code's cache warm during a break Cache Tax
See the rewrite price before a cold send Cache Tax
Keep every session warm with always mode Cache Tax

Questions

What is the best open-source tool to cut Claude Code and coding agent token usage?

Which repo for cutting coding agent token usage is easiest to start with?

Caveman: its starter use case "Cut your coding agent's token use with one skill" is rated Beginner and takes about 5 minutes.

How do I get started with Caveman?

Install for every agent with npx skills add JuliusBrussee/caveman -g; Or in Claude Code: /plugin marketplace add JuliusBrussee/caveman then /plugin install caveman@caveman; Type /caveman ($caveman in Codex) if it does not start on its own; /ultracave is even shorter. Full steps and a copy-paste agent prompt: https://dedupedai.com/use-cases/juliusbrussee-caveman/cut-coding-agent-tokens-with-the-caveman-skill/

Are these cutting coding agent token usage repos free?

They are open-source projects on GitHub (check each license; some, like fair-code projects, limit commercial use). Some also sell a hosted version or need a paid model API key; the requirements for each repo are listed on this page.

How is this list of cutting coding agent token usage repos chosen?

Deduped tracks the hottest AI repos on GitHub and writes step-by-step use cases from each README. For this page we picked the repos that do this job, ordered by fit and then GitHub stars. Repos with a risk flag or a retired page are left out. The page is rebuilt with the site; stars are as of the last build in 2026.

What is the best open-source tool to cut Claude Code and coding agent token usage?

Comparison

When to pick which

Caveman JuliusBrussee/caveman

Headroom headroomlabs-ai/headroom

Context Mode mksglu/context-mode

RTK rtk-ai/rtk

Repo Context MCP nduc99911/repo-context-mcp

Cache Tax karanb192/cache-tax

Use cases for cutting coding agent token usage

Questions

More answers