densely — Claude Code MCP server: lossless context compression
Packs large logs, JSON and dumps into 2–8x fewer tokens. Expands them back byte for byte, with a sha256 check.
claude mcp add densely -- uvx denselyWhat densely does
densely runs text through lzma, splits the result into 16-bit chunks and maps them onto English carrier words. One such word costs a single token, so two bytes of compressed data fit into each one. A neural backend is available too: an integer arithmetic coder takes next-token probabilities from Qwen2.5-Coder-0.5B.
The agent gets three tools. compress_file(path) is meant to be called instead of reading a big log or JSON, and returns a preview plus the dense payload. compress_text(text) does the same for tool output already in hand. expand(payload | payload_file, start_line, end_line) gives the original back after verifying the hash; the line range keeps the agent from paying for the whole thing. A .dense sidecar lands next to the original — plain text you can commit or archive.
Install with pip install "densely[mcp]"; the neural backend needs the [neural] extra, which pulls in torch and transformers. For hands-off use there is a PostToolUse hook that compresses Bash output and reads of non-code files from 5000 tokens up, with DENSELY_HOOK_MIN_TOKENS as the knob. Point ~/.claude/settings.json at an absolute path to densely-hook, since hooks run outside your PATH. Code files the agent is editing are left alone.
What it helps with
- fit a huge log into context without losing lines
- restore the original JSON from a payload byte for byte
- pull just one line range out of a payload
- compress Bash output automatically instead of calling a tool by hand
- see how many tokens a session saved
Example prompts for the agent
- Compress
install.logand show me the preview, don't read the whole thing - Expand lines 200 to 260 from
payload.dense - How many tokens have we saved this session?
How to install densely
Platform CLI
claude mcp add densely -- uvx denselyuvx runs the command named after the package; if the README names another one, use uvx --from densely <command>
Platform CLI
codex mcp add densely -- uvx denselyuvx runs the command named after the package; if the README names another one, use uvx --from densely <command>
Config file
{
"mcpServers": {
"densely": {
"command": "uvx",
"args": [
"densely"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from densely <command>
Config file
{
"servers": {
"densely": {
"type": "stdio",
"command": "uvx",
"args": [
"densely"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from densely <command>
Platform CLI
gemini mcp add densely uvx denselyuvx runs the command named after the package; if the README names another one, use uvx --from densely <command>
Step by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about densely
How do I install densely in Claude Code?
Run claude mcp add densely -- uvx densely. The compression itself needs the package: pip install "densely[mcp]".
What does densely do?
It shrinks text to 2–8x fewer tokens and rebuilds it losslessly, checking sha256 on every expand. The MCP server exposes three tools: compress a file, compress text, and expand.
How is this different from dropping content from context?
The data stays in the context itself and comes back byte-exact, with no external cache and no expiry. Compression below the entropy of the data is impossible, so the gain stops at that bound.
