Skip to content

ecosystem.md

Latest commit

 

History

History
83 lines (57 loc) · 4.58 KB

ecosystem.md

File metadata and controls

83 lines (57 loc) · 4.58 KB

Context-Efficient AI Coding — Ecosystem Map

Token Pilot solves one half of the context problem: the tokens that come into Claude's context from reading code. A full coding session has at least four independent places where tokens pile up — the tools below each own one of them.

Use them together. They compose cleanly and do not overlap.

Where your tokens actually go

┌──────────────────────────────────────────────────────────────────┐
│  A COMPLETE CODING SESSION                                        │
│                                                                   │
│  INPUT side                          OUTPUT side                  │
│  ──────────                          ───────────                  │
│  1. Reading code files               4. Claude's response prose   │
│  2. git diff / log                   5. Chain-of-thought noise    │
│  3. Running shell commands                                        │
│  4. Keeping context across sessions                               │
└──────────────────────────────────────────────────────────────────┘

The stack

Tool Owns Savings Mechanism
token-pilot code reads, git, search 60-90% input MCP tools + PreToolUse hooks
caveman Claude's response prose ~75% output System-prompt skill (terse style)
ast-index code indexing (underlying) foundation for structural reads Native Rust indexer
context-mode shell / python / js execution 90%+ on big stdout Sandbox — only stdout enters context
cavemem cross-session memory context across restarts Persistent structured recall

Combined footprint: a session that spends ~70% on reading code, ~25% on shell output, and ~5% on remembered context could see ~85-90% total reduction when the right tool owns each segment.

What token-pilot does not do

To keep the boundaries clear:

  • token-pilot does not change Claude's response style. If answers feel long, that's OUTPUT. Install caveman for terse-speak.
  • token-pilot does not execute code. If npm test or long python output floods your context, install context-mode.
  • token-pilot does not remember across sessions. If you're re-explaining context every morning, install cavemem.
  • token-pilot does not index source. It uses ast-index under the hood — installed automatically, but also standalone.

Installing the full stack

Each tool is independent — install whatever you need.

# Claude Code (plugin system)
claude plugin marketplace add Digital-Threads/token-pilot
claude plugin install token-pilot@token-pilot

claude plugin marketplace add JuliusBrussee/caveman
claude plugin install caveman@caveman

# token-pilot bootstraps ast-index automatically.
# context-mode can be installed via its own plugin route.

For other clients (Cursor, Codex, Cline, Windsurf, …) each tool has its own install matrix — follow each project's README.

Why not one meta-plugin?

We considered shipping ai-coding-savings-pack that bundles all of them. Tradeoffs:

  • Pro: one-command install.
  • Con: blast radius. If any component ships a regression, the whole pack looks broken. Each tool has its own release cadence and support surface; coupling them hides those boundaries.

For now the recommendation is install what you need, individually. Revisit bundling after real combined-usage data shows it pays off.

Measuring the combined effect

Each tool ships its own telemetry, read as-is:

npx token-pilot tool-audit            # input savings (per tool, cumulative)
npx token-pilot doctor                # ecosystem coverage check — see what's installed

The doctor command checks which ecosystem tools are active in the current environment and prints gaps. That's the cheapest way to see "am I leaving savings on the table?"

Credits

Everything listed here is MIT or Apache-licensed open source maintained by small teams. If you use them, please star the repos — all of this is volunteer work.