copt - Bayesian Context Optimizer

A local MCP server that uses Bayesian estimation to optimize your IDE's context window.

Reduces token usage and cost while preserving response quality — without touching your API keys or sitting in the request path. copt plugs into Cursor, Claude Code, Windsurf, and any other IDE that supports MCP.

Quick Start

# Install
curl -fsSL https://raw.githubusercontent.com/jmouchawar/copt/main/scripts/install.sh | bash

# Initialize — auto-detects and configures Cursor, Claude Desktop, Windsurf
copt init

# Your IDE will launch copt mcp automatically. To test it manually:
copt mcp

copt init writes the MCP server entry to your IDE's config file and requires no API key. Your IDE's existing AI credentials are used as-is.

How It Works

Your IDE's AI calls the optimize_context MCP tool before each LLM request. copt breaks the context into semantic chunks (system prompt, current file, error context, conversation history, etc.) and uses a Beta-Bernoulli Bayesian model with Thompson Sampling to select the most relevant subset within the token budget.

• No outbound requests: copt never calls an LLM. The IDE handles all API calls.
• Learns from feedback: record_feedback calls update the posteriors; future selections improve.
• Starts conservative: default priors favour including context until evidence says otherwise.
• Conjugate updates: O(1) per observation — no MCMC, no GPU.

MCP Tools

Tool	Description
optimize_context	Trim messages to the most relevant context chunks
record_feedback	Record whether a response was helpful, updating Bayesian posteriors
get_status	Return current posterior means for all (task type, category) pairs

Commands

Command	Description
copt init	Initialize config and register copt with detected IDEs
copt mcp	Start the MCP server (stdio) — normally launched by your IDE
copt status	Show metrics summary
copt status --posteriors	Show posterior convergence details
copt config	Show current configuration
copt config --check	Exit 0 if initialized, 1 if not
copt reset --confirm	Clear all learned posteriors and request logs
copt dashboard	Start the web analytics UI (default http://127.0.0.1:3847)
copt serve	Start the legacy HTTP proxy (requires API key)

IDE Setup

copt init handles this automatically. If you need to set it up manually:

Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "copt": { "command": "copt", "args": ["mcp"] }
  }
}

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "copt": { "command": "copt", "args": ["mcp"] }
  }
}

Claude Code (CLI)

claude mcp add copt -- copt mcp

Windsurf

Add to ~/.codeium/windsurf/mcp_settings.json:

{
  "mcpServers": {
    "copt": { "command": "copt", "args": ["mcp"] }
  }
}

Configuration

~/.copt/config.yaml — all fields are optional:

max_context_tokens: 8000     # token budget for optimization (0 = no limit)
relevance_threshold: 0.3     # Thompson Sampling minimum score
data_dir: "~/.copt"          # SQLite database location
log_level: "info"

The following are only needed for proxy mode (copt serve):

listen_addr: "127.0.0.1:8090"
upstream_url: "https://api.openai.com"
api_key: "sk-..."            # optional; can also use COPT_API_KEY env var
passthrough: true            # set to false to enable optimization in proxy mode

Dashboard

View request metrics, token savings, quality scores, and Bayesian posterior convergence in your browser:

copt dashboard
# Open http://127.0.0.1:3847

Use --listen to change the bind address:

copt dashboard --listen 127.0.0.1:9000

The dashboard reads the same SQLite database used by copt mcp and copt serve, so it works regardless of which mode you use.

Proxy Mode (advanced)

If you prefer the transparent HTTP proxy approach — useful for tools that don't support MCP — copt serve still works:

copt serve --api-key sk-...
# Point your tool at http://localhost:8090/v1

Installation

• Script: curl -fsSL https://raw.githubusercontent.com/jmouchawar/copt/main/scripts/install.sh | bash
• Go: go install github.com/jmouchawar/copt/cmd/copt@latest
• From source: git clone https://github.com/jmouchawar/copt && make build

Building

make build    # Build binary to ./build/copt
make test     # Run tests
make install  # Install to $GOPATH/bin