@glrs-dev/cmprss

Provider-agnostic context-compression proxy for AI coding agents. Bedrock-first.

v0.1 status: passthrough proxy only. cmprss wrap claude starts a local proxy, translates Anthropic Messages requests to Bedrock Converse outbound (with SigV4 via AWS SDK), and streams responses back. Compression lands in v0.2.

Install

bun add -g @glrs-dev/cmprss
# or
npm i -g @glrs-dev/cmprss

Usage

# Wrap Claude Code with Bedrock backend (default region us-east-1)
cmprss wrap claude

# Wrap opencode TUI
cmprss wrap opencode

# Different region / port
cmprss wrap claude --region us-west-2 --port 8788

# Pass args through to the wrapped agent
cmprss wrap claude -- --version
cmprss wrap opencode -- /path/to/project

cmprss uses the AWS SDK default credential chain (env vars, profile, SSO cache, IRSA, IMDS). No bearer token plumbing.

Model selection — cmprss does not pick

cmprss is model-agnostic. Pick models in your agent's UI as normal — the proxy maps each request independently to the right Bedrock inference profile for the configured region.

A single opencode/claude-code session typically uses several models (main + summarizer + planner). Each one passes through cmprss; each one gets its own mapping.

Anthropic-API names → Bedrock inference profiles (e.g. claude-sonnet-4-5-20250929 → us.anthropic.claude-sonnet-4-5-20250929-v1:0). Already-formatted Bedrock IDs pass through unchanged. Unknown models return a 400 with the list of recognized names — open an issue if you hit something we should add.

opencode caveat: only opencode's anthropic/* provider routes through ANTHROPIC_BASE_URL. Its amazon-bedrock/*, openai/*, google-vertex/* etc. providers go direct to their respective backends and bypass cmprss.

Status

Roadmap lives in the design plan; this is v0 of the vertical slice.