If you run an AI agent heavily enough, you know the moment: it's Tuesday, your Claude account hits its weekly limit, and your assistant quietly falls back to a different provider — or just gets dumber. Meanwhile the second account you also pay for sits there with a full quota, because there's no way to tell one running agent "carry on, but on the other login."
We had exactly that problem with our OpenClaw assistants, so we built multi-clawd — and this week we opened it up. It's free, MIT-licensed, and if you run OpenClaw with more than one Claude subscription, it may be the missing piece.
GitHub: https://github.com/Drakon-Systems-Ltd/multi-clawd
npm: https://www.npmjs.com/package/@drakon-systems/multi-clawd
What it does
multi-clawd registers each of your own Claude Code logins as a first-class OpenClaw CLI backend — full skills, MCP tools, and native compaction intact on every account — and then pools them behind a single backend id. Route your model at clawd/claude-fable-5 and every launch runs on the first account in the pool that isn't nearly maxed out.
The part that took the real engineering: it rotates before the limit, not after the error. A transparent shim watches the Claude CLI's own rate_limit_event stream, keeps per-account health state, and hands over at 85% utilization — so the switch happens on a healthy turn, not as a recovery from a failed one. Sticky selection stops it flapping back and forth mid-conversation, and when the home account's usage window resets, the pool comes home.
When every account is exhausted, an optional degradation ladder steps down a model tier on the same provider instead of hard-failing to the next one — with pin rules for lanes that must never degrade.
Setup is a conversation, not a runbook
The install path we're most pleased with: if you already run an agent (that's the target audience, after all), you paste one line to it —
Read https://raw.githubusercontent.com/Drakon-Systems-Ltd/multi-clawd/master/SETUP-AGENT.md and follow it to set up multi-clawd on this machine. I own a second Claude account — ask me when you need me to log in.
— and your agent does the whole thing: install, guided wizard, config merge with backup, watchdog scheduling, health check. The guide bakes in the guardrails (never print tokens, never overwrite a working account, and the OAuth login is an explicit ASK-YOUR-HUMAN checkpoint, because your agent should never touch your credentials).
Prefer to drive it yourself? The CLI is five commands:
npm i -g @drakon-systems/multi-clawd
multi-clawd update # install/update the OpenClaw plugin
multi-clawd setup # guided wizard: accounts, pool, watchdog
multi-clawd login claw2 # launches the right Claude sign-in per account
multi-clawd doctor --probe # health check + one live end-to-end turn
multi-clawd explain # your whole setup in plain English
explain is the one people seem to love: it narrates your accounts, your fallback chain rung by rung, and live health with reset times — no JSON spelunking.
The honest bits
Young project, moving fast — v1.0 to v1.4.2 shipped in two days, and the README keeps an honest known-limitations section. Two things worth knowing up front:
- On current OpenClaw builds an idle plugin backend can occasionally be evicted by an upstream bug (openclaw#107408 — our fix is in review as PR #107596). The setup wizard schedules a watchdog that self-heals it in the meantime.
- Switching accounts mid-conversation costs the Claude CLI its native session once per rotation; as of v1.4.2 the turn reseeds seamlessly from OpenClaw's own history and stays on Claude. We found that one the fun way, in production, this Monday morning.
Everything it does is disclosed surface-by-surface in the README's security table: it reads your own OAuth token files (or pulls them from your secret manager — never logged, never transmitted), spawns the claude CLI, and writes one watchdog unit. No network egress of its own, no telemetry.
To be clear about what this is: pooling capacity across your own subscriptions, on your own machine — the accounts you already pay for, working like one.
If that's your Tuesday problem too, give it a spin — issues and PRs welcome. 🦞
Top comments (0)