Glue Claude back into OpenClaw.
GlueClaw will no longer work after June 15 2026 as per changes to -p mode billing. Enjoy the last month of OpenClaw with Claude!
Uses the official Claude CLI and scrubs out Anthropic's detection triggers from the system prompt due to Anthropic not allowing its use. Tested with Telegram. As far as I can tell all functions work such as heartbeats.
Requires OpenClaw 2026.5.x+, Claude Code logged in with Max, and Node.js 22+. Non-destructive, won't touch your existing config or sessions.
npm install @zeulewan/glueclaw-provider && bash node_modules/@zeulewan/glueclaw-provider/install.shgit clone https://github.com/zeulewan/glueclaw.git && cd glueclaw && bash install.sh
See installation docs for uninstall and details.
Uses the official Claude CLI:
claude --dangerously-skip-permissions -p \
--output-format stream-json \
--verbose --include-partial-messages \
--system-prompt <scrubbed prompt> \
--model <model> \
--resume <session-id> \
"<user message>"
The only way this breaks is if Anthropic changes how --system-prompt or --output-format stream-json work, which would affect all Claude Code integrations.
| Model | Claude Model | Context |
|---|---|---|
glueclaw/glueclaw-opus |
Opus 4.6 | 1M |
glueclaw/glueclaw-sonnet |
Sonnet 4.6 | 1M |
glueclaw/glueclaw-haiku |
Haiku 4.5 | 200k |
Switch in TUI: /model glueclaw/glueclaw-opus
| Env var | Default | Description |
|---|---|---|
GLUECLAW_REQUEST_TIMEOUT_MS |
120000 |
Maximum time (in milliseconds) to wait for the Claude CLI subprocess to complete a single request before it is terminated. Increase if long-running tool calls or extensive reasoning trip the default 120s limit. Invalid or non-positive values fall back to the default. |
Example (10 minute timeout):
export GLUECLAW_REQUEST_TIMEOUT_MS=600000- Tested with Telegram and OpenClaw TUI
- Switching between GlueClaw and other backends (e.g. Codex) works seamlessly via
/model - The installer does not patch OpenClaw's dist. GlueClaw starts the MCP loopback in-process when available.
- Multi-agent setups are isolated end-to-end: each agent gets its own Claude project storage and its own session-id cache, anchored at the agent's
workspaceDir. See the multi-agent guide.
Uses only official, documented Claude Code CLI flags. No reverse engineering, no credential extraction, no API spoofing. Use at your own risk. Not affiliated with Anthropic. Compatible with OpenClaw's plugin allowlist system.