Compresh plugs into your existing tools without rewriting your code. There are three integration paths — pick the one that matches how your stack already calls the model.
Swap your base_url to api.compre.sh/v1. Compresh sits between your client and the model provider. Works with anything that speaks OpenAI- or Anthropic-compatible HTTP.
Compresh exposes a compress tool over the Model Context Protocol. The agent decides when to call it. Opt-in, model-controlled.
Plugin SDK / lifecycle hooks fire on every model call. The host rewrites the prompt before it goes out — deterministic, no agent discipline required. Only a few hosts expose this surface today.
Support matrix
Each cell shows the current status of that host × method combination. TESTED means end-to-end verified and ready to install. SOON is a planned guide we have not yet validated — assume it does not work until we publish a verified guide. — means the method is not applicable to that host (no API surface to integrate with).
Hook coverage is rare because it requires the host itself to expose a per-turn middleware API. OpenClaw, Claude Code, and CrewAI's step_callback are the only hosts with a usable hook surface today. Most editors and SDKs do not.
| Host | Proxy | MCP | Hook |
|---|---|---|---|
| Hosts with hook surface | |||
| OpenClaw | SOON | TESTED | TESTED |
| Claude Code | — | SOON | SOON |
| CrewAI | SOON | — | SOON |
| MCP-aware hosts (no hook surface) | |||
| Claude Desktop | — | SOON | — |
| Cursor | SOON | SOON | — |
| Continue.dev | SOON | SOON | — |
| Cline | SOON | SOON | — |
| Zed | SOON | SOON | — |
| Proxy-only — SDKs, libraries, local LLM | |||
| OpenAI SDK | SOON | — | — |
| Anthropic SDK | SOON | — | — |
| LangChain | SOON | — | — |
| LiteLLM | SOON | — | — |
| OpenRouter | SOON | — | — |
| Aider | SOON | — | — |
| Ollama | SOON | — | — |
| LM Studio | SOON | — | — |
Compresh works with anything that speaks OpenAI- or Anthropic-compatible HTTP. If your stack lets you change the API base URL, the proxy path applies. Open an issue on GitHub and we'll fast-track verification.
Which path should I pick?
- Use Proxy if you call the model directly from your own code (OpenAI/Anthropic SDK, LangChain, LiteLLM, Aider).
- Use MCP if your agent host already speaks MCP and you want the model to decide when to compress (Claude Desktop, Cursor, Cline).
- Use Hook if you want deterministic per-turn compression without trusting the model to call a tool (OpenClaw, Claude Code).
A single host can support multiple paths — for example, Cursor has both a Proxy and an MCP option. Pick whichever matches how your team prefers to manage middleware.