Exodia is a zero-waste development configuration for agentic AI coding assistants like Claude Code, Cursor, Windsurf, Copilot, OpenAI Codex, and Antigravity. It combines five MCP servers plus a CLI proxy to reduce token bloat, prevent context collapse, and cut costs — particularly valuable for developers on the Claude Pro $20/month plan who need to make every token count.
Paste this into Claude Code, Codex, or any agentic coding assistant:
Clone https://github.com/reni10/exodia-stack and run ./install_exodia.sh to install the EXODIA token optimization stack. Then copy the contents of policy.md into my CLAUDE.md file.
That's it. Your agent clones the repo, registers the MCP servers, installs RTK, and configures the policy. No manual steps.
Before adding any MCP servers, install this first. Every MCP server you add dumps its full tool schema into the context window on every message. With 5 servers, that's 20-50k tokens of tool descriptions eating your context before you even type anything. mcp2cli eliminates this overhead entirely.
| Piece | What It Does | Token Savings |
|---|---|---|
| mcp2cli | Claude Code plugin that converts MCP servers into CLI tools so their schemas aren't loaded into context. Without this, the other 5 pieces ironically add their own token bloat. | 96-99% on tool schema overhead |
Note: mcp2cli is a Claude Code plugin. Cursor, Windsurf, and other agents benefit from the other 5 pieces but don't support this plugin.
With Piece Zero handling the overhead, these five components do the real work. Each delivers value independently:
| Piece | What It Does | Token Savings |
|---|---|---|
| jCodeMunch | Structural project indexing via tree-sitter AST. Returns exact symbols instead of entire files. | 80-95% on code reading |
| Serena (LSP) | Semantic navigation — Go-to-Definition, Find-All-References via Language Server Protocol. | 60-70% on code navigation |
| Codebase-Memory | Persistent knowledge graph of your codebase architecture. 64 languages, sub-ms queries. | 70-90% on architecture discovery |
| AI-Distiller | Real-time compression of verbose terminal outputs. | 50-80% on log/output parsing |
| RTK (Rust Token Killer) | CLI proxy that compresses command output before it hits the context window. Single Rust binary. | 60-90% on command outputs |
The Claude Pro plan has a usage-based rate limit per 5-hour window. Without optimization, agents waste most of your token budget on reading files and parsing output rather than writing code. EXODIA flips that ratio.
| Operation | Without Exodia | With Exodia |
|---|---|---|
| Reading a function | Read entire 500-line file (~3,800 tokens) | Retrieve exact symbol (~700 tokens) |
| Finding all usages | grep + read 5 files (~12,000 tokens) |
LSP Find-All-References (~500 tokens) |
| Build error debugging | 500 lines of raw output (~10,000 tokens) | Compressed summary (~800 tokens) |
git push output |
~200 tokens | ~10 tokens |
pytest (33 tests) |
~756 tokens | ~24 tokens |
The result: significantly more messages per rate-limit window, and longer productive sessions before you hit the wall.
git clone https://github.com/reni10/exodia-stack.git && cd exodia-stack && ./install_exodia.sh- Copy the contents of
mcp_config.jsoninto your agent's MCP settings - Install RTK:
cargo install rtk - Copy
policy.mdinto your project's.cursorrulesoragents.md
- Python 3.8+ with
uvx(for jCodeMunch and Serena) - Node.js 18+ with
npx(for Codebase-Memory and AI-Distiller) - Rust toolchain (for RTK) — or use the pre-built binary
Prefix verbose commands to compress their output:
rtk npm run build
rtk pytest -v
rtk cargo buildBefore EXODIA can protect your context window, it needs to build its indexes:
- jCodeMunch: Ask your agent to run
mcp_jcodemunch_index_folder— creates the symbol map (classes, functions, types). - Codebase-Memory: Ask the agent to scan the project — builds the architectural knowledge graph.
- Serena: Ask the agent to perform a reference check on a core module — verifies the LSP is mapping correctly.
Once indexed, all three track changes incrementally. You only do this once per project.
Before asking an agent to "find where X is used," prompt it: "Use jCodeMunch/Serena to look this up." This prevents fallback to expensive grep + full file reads.
When a build fails, the agent normally dumps the entire error log into context. With RTK, it only sees a compressed summary of what actually went wrong.
Instead of the agent reading 20 files to understand how your app is structured, Codebase-Memory provides the dependency graph and module relationships in a single query.
- Serena works best with Python, TypeScript/JavaScript, Go, Rust, and Java. Other languages have partial LSP support.
- jCodeMunch indexing can take a minute on very large repos (100k+ files). Incremental updates are fast after that.
- RTK requires Rust to compile from source. Pre-built binaries are available for macOS and Linux.
- AI-Distiller adds the most value on very verbose outputs (build logs, test suites). For short commands, the overhead isn't worth it.
- The full stack adds 4 MCP server processes to your system. On resource-constrained machines, start with just jCodeMunch + RTK for the biggest impact with the least overhead.
- mcp2cli is Claude Code only — other agents don't support plugins but still benefit from the 5 MCP servers.
EXODIA is a curation of the best zero-waste tools in the agentic ecosystem. Full credit to the original creators:
- mcp2cli by myeolinmalchi
- jCodeMunch by J. Gravelle
- Serena by Oraios AI
- Codebase-Memory by DeusData
- AI-Distiller by Jan Reges
- RTK by RTK AI
MIT. Go forth and build without bloat.