gramatr

Intelligence layer for AI coding agents. Pre-classifies every request so Claude Code, Codex, and Gemini CLI spend tokens on your work, not on routing overhead.

What it does

gramatr sits between you and your AI agent. Before the agent sees your prompt, gramatr's decision router (BERT classifiers running in <5ms) pre-classifies:

• Effort level — instant, fast, standard, extended, advanced, deep, comprehensive
• Intent type — search, retrieve, create, update, analyze, generate
• Skill matching — which of 25 capabilities are relevant
• Memory tier — what context to pre-load from your knowledge graph
• Reverse engineering — what you want, what you don't want, gotchas

The agent receives this as a pre-computed intelligence packet, saving ~2,700 tokens per request that would otherwise be spent on the agent figuring out what you need.

Install

npx gramatr install claude-code

That's it. One command installs hooks, registers the MCP server, and configures your AI agent.

Non-interactive (CI, remote provisioning)

npx gramatr install claude-code --yes --name "Your Name" --timezone "America/Chicago"

Supported platforms

| Platform | Status | Install | |---|---|---| | Claude Code | Stable | npx gramatr install claude-code | | OpenAI Codex | Stable | Auto-detected during install | | Google Gemini CLI | Stable | Auto-detected during install | | Claude Desktop | Coming soon | — | | ChatGPT Desktop | Coming soon | — |

How it works

You type a prompt
  |
gramatr hooks intercept (UserPromptSubmit)
  |
Decision router classifies in <5ms (BERT on GPU)
  |
Intelligence packet injected into agent context
  |
Agent receives pre-classified request
  = skips routing overhead
  = starts working immediately
  = saves ~2,700 tokens per request

The flywheel

Every interaction trains the classifier. Memory powers routing, routing generates patterns, patterns improve routing. The system gets smarter with use.

Architecture

gramatr is a thin client + smart server:

• Client (this package): 29 files, ~290KB. Hooks into your AI agent, forwards to server.
• Server: Decision routing engine (BERT + Qwen), knowledge graph (PostgreSQL + pgvector), pattern learning.
• Protocol: MCP (Model Context Protocol) for Claude Code/Codex, REST API for web/mobile.

The client never stores intelligence locally. The server delivers everything: behavioral rules, skill routing, agent composition, ISC scaffolds, capability audits.

What gets installed

~/gmtr-client/          # Client runtime
  hooks/                # 8 lifecycle hooks
  core/                 # Shared routing + session logic
  bin/                  # Status line, login, utilities
  CLAUDE.md             # Minimal behavioral framework
~/.claude/settings.json # Hook configuration (merged, not overwritten)
~/.claude.json          # MCP server registration
~/.gmtr.json            # Auth token (canonical source)

Commands

gramatr install          # Interactive target selection
gramatr install claude-code  # Install for Claude Code
gramatr detect           # Show detected AI platforms
gramatr doctor           # Health check (coming soon)
gramatr uninstall        # Clean removal
gramatr --version        # Show version

Requirements

• Node.js 20+
• One of: Claude Code, OpenAI Codex, Google Gemini CLI

Links

• gramatr.com
• GitHub

License

MIT