Labsco
MCP SERVER

Daimonos

by beardfaceguy

Replaces your agent's file, search, exec and git tools with ones that return compact JSON instead of raw terminal output.

Codebase Context Packing & Compression
Summary
The saving is real and the project publishes the arithmetic.

The premise is simple enough to check: an agent pays tokens for every line of terminal noise it reads and then carries in context. On identical coding tasks the project reports output tokens falling from 5,842 to 3,198 and tool calls from 17 to 14 on average, with remote benchmarks on AWS showing a 20.3% cost reduction. Those are the author's own measurements on their task set, and the benchmark scripts are in the repository to re-run. The part that needs no benchmark is the read deduplication — re-reading an unchanged file returning a flag instead of the file is straightforwardly free.

What it is

A single Rust binary that acts as an agent-optimised tool layer. When an agent runs a test suite it normally gets back hundreds of lines of progress bars and passing test names; Daimonos returns the summary and the failures. It runs as an MCP server your IDE spawns and exits when you close the project — no daemon, no background service. Linux and macOS; Windows is planned.

What you get
  • Core tools that are always available: `read_file` with offset and limit and content-hash deduplication, `write_file` with auto-mkdir, `edit_file` with diff confirmation, `search` in content or file-discovery mode, `exec` with semantic filtering and bounded capture, `batch` for several operations in one round-trip, and `workspace_info` for project type, git status and directory listing
  • Native plugins that appear automatically when the CLI is on `PATH` — `git` (status, log, diff, branch, add, commit, push, pull, checkout), `cargo` (test, build, check, clippy, fmt, add), `gh` (PR view, list, create, diff, checks, api) and `docker` (ps, logs, exec, images, inspect, stop, compose)
  • Semantic compression for commands with no native plugin, so pytest, make, pip install and eslint return a summary plus failures rather than the full stream
  • Read deduplication: re-reading an unchanged file returns `{"unchanged":true}` instead of the whole content
  • Managed subprocess execution — output is bounded while it is read rather than after full buffering, Unix process groups are owned so descendants are retired on cancellation, child environments are restricted to an explicit allowlist, and background output goes to private bounded artifacts
  • Workspace snapshots to checkpoint before a risky edit and roll back on failure, Starlark scripting to bundle several tool calls into one script, and per-tool-call token analytics with cross-session history
Requirements

No account and no key. Download a prebuilt binary for Linux x86_64, Linux aarch64, macOS Apple Silicon or Intel Mac, or build from source with `cargo build --release`; put `daimonos` on your `PATH`. Configuration is one entry in your MCP client — for Cursor, `daimonos` with `--mcp -w /path/to/your/project` — and it starts when the IDE opens the project. Setup guides ship for Cursor, GitHub Copilot, Claude Code, Windsurf, Cline, Gemini CLI, Zed and others. Everything else is tunable through a single TOML config file. MIT licensed.

Setup effort

One command — curl -L https://github.com/beardfaceguy/daimonos/releases/latest/download/daimonos-x86_64-linux.tar.gz | tar xz