I’m enjoying using OpenClaw more than ever because it’s actually cheaper than ever. I have it using the official Claude Code backend now with a custom system prompt, so my regular Max tokens work with it.

Super easy to install but may be buggy rn. https://github.com/zeulewan/glueclaw

Background: Anthropic changed their billing, which made API-based use a lot more expensive. That’s all good, they’re rolling out Mythos, which likely needs a lot more compute, but I found that OpenClaw wouldn’t work through a custom system prompt even though other things still would.

What I found out: By binary searching the OpenClaw system prompt and scrubbing out mentions of OpenClaw, I was able to get it working again.

Technical info: It spawns the real Claude CLI as a subprocess with `--system-prompt` and `--output-format stream-json` and translates OpenClaw’s gateway tools through the MCP loopback bridge. The system prompt is scrubbed of Anthropic’s detection triggers and response tokens like `HEARTBEAT_OK` are translated back so the gateway still works. It creates a model provider called `glueclaw` and three models: `glueclaw-opus`, `glueclaw-sonnet`, and `glueclaw-haiku`.

PRs welcome. MIT licensed.

Note: This project uses only official, documented Claude Code CLI flags. No reverse engineering, no credential extraction, and no API spoofing. I don’t believe this is against the TOS, but I’d love for Anthropic to clarify the policy here and let us know whether OpenClaw through the official Claude tooling is all good.

     const cleanPrompt = (context.systemPrompt ?? "")
          .replace(/personal assistant running inside OpenClaw/g, "personal assistant running inside GlueClaw")
          .replace(/HEARTBEAT_OK/g, "GLUECLAW_ACK")
          .replace(/reply_to_current/g, "reply_current")
          .replace(/\[\[reply_to:/g, "[[reply:")
          .replace(/openclaw\.inbound_meta/g, "glueclaw.inbound_meta")
          .replace(/generated by OpenClaw/g, "generated by GlueClaw");