Decompose — Claude Code MCP server: classify text, no LLM
Splits text into semantic units tagged by authority, risk and attention. Useful when an agent has to read a spec or contract before reasoning over it.
claude mcp add decompose -- uvx decompose-mcpWhat Decompose does
Decompose is an MCP server that breaks text into individual units and labels each one: authority (mandatory, prohibitive, directive), risk (safety_critical, compliance, financial), type, an irreducible flag and an attention score from 0.0 to 10.0. Formal references such as standards and codes land in entities. The whole thing runs on regex and heuristics, so no model call happens.
Two tools come with it: decompose_text handles text you pass in, decompose_url fetches a page and processes its content. There is also a CLI (decompose --pretty, --compact) and a Python library exposing decompose_text and filter_for_llm. That last function drops boilerplate and keeps mandatory, safety-critical and financial units, which matters if you trim context before sending it to a model.
No API key, no GPU. The server works offline and returns the same output for the same input every time. Classification patterns were built for construction documents — specs, contracts, RFIs, inspection reports, pay applications — so expect noise on prose or source code.
What it helps with
- split a spec into individual requirements
- drop boilerplate before sending text to an LLM
- sort contract clauses by risk level
- pull out cited standards and codes
- count how many units actually need analysis
Example prompts for the agent
- Decompose
spec.txtand show only the units with safety_critical risk - Run
decompose_urlon this link and list the mandatory requirements - Filter the document with
filter_for_llmand tell me the token reduction
How to install Decompose
Platform CLI
claude mcp add decompose -- uvx decompose-mcpuvx runs the command named after the package; if the README names another one, use uvx --from decompose-mcp <command>
Platform CLI
codex mcp add decompose -- uvx decompose-mcpuvx runs the command named after the package; if the README names another one, use uvx --from decompose-mcp <command>
Config file
{
"mcpServers": {
"decompose": {
"command": "uvx",
"args": [
"decompose-mcp"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from decompose-mcp <command>
Config file
{
"servers": {
"decompose": {
"type": "stdio",
"command": "uvx",
"args": [
"decompose-mcp"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from decompose-mcp <command>
Platform CLI
gemini mcp add decompose uvx decompose-mcpuvx runs the command named after the package; if the README names another one, use uvx --from decompose-mcp <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 Decompose
How do I install Decompose in Claude Code?
Run claude mcp add decompose -- uvx decompose-mcp. That registers the server in your MCP config, and the agent gets the decompose_text and decompose_url tools.
What does Decompose do?
It splits text into semantic units and gives each one an authority category, a risk level, a type, an irreducible flag and an attention weight. Classification runs on regex rather than a language model.
Does it need an API key or a GPU?
Neither. Everything is computed locally with heuristics, so it runs air-gapped and costs nothing per call.
