clumsies

Persistent, deterministic, and user-managed project memory for vibe coding.

v0.18.0-alpha — functional but not battle-tested. The MCP server, stats engine, and Claude Code plugin are all part of this release line.

Every AI agent compresses memory when context runs low. Claude Code at ~80% capacity, Cline at ~80%, Amazon Q at ~80%. The agent decides what matters and what gets dropped. This works for generic knowledge — common patterns, standard library usage, widely known conventions.

But your project has rules that are specific to you. "All Zig code must use explicit allocators, never GeneralPurposeAllocator." "Commit messages follow this exact subsystem:subject format." "External API calls must go through the retry wrapper." These rules matter to your project, but from the compression algorithm's perspective, they are edge cases — not frequent enough to be prioritized, not generic enough to be preserved. They get silently dropped, and the agent keeps working confidently with outputs that look correct but violate constraints you thought were non-negotiable.

You have no way to know what was compressed away. The agent does not tell you.

clumsies puts your critical constraints in a place the agent's compression algorithm cannot touch: a .prompts/ directory in your project that the agent queries but never manages.

You decide what goes in .prompts/. You write rules, workflows, and context as markdown files. The agent discovers them through a structured protocol (memory.search, memory.load), follows them, and declares which ones it actually used (memory.refer). Over time, the trace data tells you which constraints are effective and which are dead weight — so you can refine what you wrote.

The agent has no authority to compress, summarize, or delete anything in .prompts/. It can only read.

CLI + Registry. Manage a personal prompt library. Register constraints refined through real use, store them in a git-based registry, import them into any project.

MCP Server. Structured protocol for agents to discover constraints (memory.search), load them (memory.load), and declare references (memory.refer — batch support). Every interaction produces a trace log.

Stats engine. Aggregates trace data: which constraints are hot, which are cold, how coverage changes across versions.

Claude Code plugin. Hooks and skills that solve MCP's passive nature. Startup hook loads your meta-prompt automatically via CLI pipe. Stop hook reminds the agent to declare constraint references.

Install the CLI:

curl -fsSL https://raw.githubusercontent.com/lilhammerfun/clumsies/main/install.sh | sh

Import a starter bundle into your project:

cd your-project
clumsies get opus-coding --registry https://github.com/lilhammerfun/clumsies-registry.git

This creates .prompts/ with coding rules, workflows, and a META_PROMPT.md. Launch Claude Code with the plugin (marketplace distribution planned):

claude --plugin-dir /path/to/clumsies/cc-plugin

On session start, the plugin loads META_PROMPT.md and generates slash commands for your workflows. Give the agent a task — it will search and load constraints from .prompts/.

Check what happened:

clumsies stats

.prompts/
├── META_PROMPT.md     # protocol bootstrap — loaded on session start
├── rule/              # constraints — coding rules, project context, etc.
├── workflow/           # ordered procedures — commit messages, architecture, etc.
├── context/           # reference material — research, specs, documentation
└── ...                # whatever else you need

META_PROMPT.md tells the agent what clumsies is, how to use the protocol, and what the priority model looks like. Everything else is discovered through memory.search and loaded on demand.

curl -fsSL https://raw.githubusercontent.com/lilhammerfun/clumsies/main/install.sh | sh

git clone https://github.com/lilhammerfun/clumsies.git
cd clumsies
zig build -Doptimize=ReleaseFast

Current version: v0.18.0-alpha

The MCP server and stats engine are functional but not yet tested at scale. Trace data quality depends on agent compliance — which is what the Claude Code plugin is designed to improve.