One Claude is never enough
An OpenClaw plugin that registers your extra Claude Code logins as first-class CLI backends. Route through the pool to try another eligible account on the same model before a configured tier drop, keeping the Claude Code harness under the host's tool policy.
Features
Accounts register as real CLI backends (claw2/…, claw3/…) — resolvable in model refs, fallback chains, and per-agent overrides. No API keys, no baseUrl hacks.
Exhaust the account, not the model. Route through the clawd pool to try another eligible login on the same model before a configured tier drop.
Account hops keep the Claude Code subprocess, native compaction, skills and OpenClaw MCP bridge. Runs with an exact tool cap restrict tools and disable hooks, plugins, memory, slash commands and Chrome to honour the host policy.
For Claude Code launches, setup-tokens are read at launch and passed through the child process environment, not command-line arguments. Use a token file or a gateway secret reference rather than placing the token in your config.
If the plugin loader hands over an empty config block, registration re-reads the resolved runtime config directly rather than treating that empty block as the account list.
Native accounts use the default Claude config directory or OS keychain without copying credentials. Extra accounts can use an isolated configDir with its own login, a token file, or a gateway secret reference — a separate token file is not required for a directory login.
How it works
npm i -g @drakon-systems/multi-clawd && multi-clawd update. The CLI installs the plugin, offers a gateway restart, and runs a health check. The registry package is prebuilt; OpenClaw 2026.8.1 or later and the Claude CLI on PATH are required for the plugin.
multi-clawd setup walks through accounts, isolated login dirs, token storage and the pool. It backs up openclaw.json before touching it, merges rather than overwrites, and supports --dry-run.
multi-clawd login claw2 launches the sign-in flow for that account: native login, isolated directory login, or setup-token flow. Native and directory logins show the signed-in email; token accounts get storage guidance. Gateway secret references are supported for OpenClaw launches.
multi-clawd chain audits recognised Claude references in configured chains, per-agent overrides and stored session pins for pool bypasses. multi-clawd doctor checks installation, credential evidence and routing; add --probe for live account turns. Unreadable or unsupported stores are reported, not a guarantee of complete coverage.
Release v2.3.0 · 20 October 2026
The 2.3.0 release adds an optional reset-aware pool strategy that uses weekly quota before it resets. If you are upgrading from the 1.7-era CLI, these are the main changes along the way — including the 2.0 token-source behaviour change.
A new optional pool setting, strategy: reset-aware, picks the healthy account whose seven-day quota window resets soonest, so allowance that is about to refresh gets used first. It applies to both CLI and direct routes, and the explain output shows the strategy and why the chosen account was picked. The default stays ordered, so existing configs behave as before.
A new optional pool setting, weeklyUtilizationThreshold, lets day, week and month quota windows hand over at their own fraction, for example using more of a weekly allowance before rotating. It defaults to utilizationThreshold, so existing configs behave as before; paid overage still uses the general threshold when rotateOnOverage is enabled. The release also documents and pins how long an account stays available after its stored login goes missing.
The plugin requires OpenClaw 2026.8.1 or later. On supported hosts, verified Claude launches fix OpenClaw 2026.10.2 setup authority; Windows keeps the legacy launch, as do hosts where the claude executable cannot be resolved or that lack the required launch capabilities. Side-question turns have no in-turn retry. Login-backed models now appear available in the Control UI when supported by credential evidence, with cache-boundary fixes in 2.1.2. Version 2.1.3 fixes source installs on npm 10 (Node 22) by shipping a lockfile.
Native and directory login checks need an unexpired access token, or an expired credential backed by a successful CLI turn after expiry within 24 hours of the check. Credential-file presence alone is not proof. After its initial exclusion, a recorded auth failure stays unverified until observed success clears it; usage-poll data is not login proof. Doctor also ignores an inherited CLAUDE_CONFIG_DIR when checking the native account.
Doctor's split-store check only reads and prints: for native and directory OAuth logins it checks the expected credentials file and, on macOS, Keychain items under that account's service. It lists recognised copies and which one the CLI reads when that can be determined; it does not move or delete credentials. Cleanup guidance calls for proving the CLI's copy before moving a stray file aside, and a fresh login before removing a stray Keychain item.
The pool can retry once on an eligible sibling when a recognised model-limit refusal arrives before output is forwarded. Resumed conversations can hand over their pre-attempt transcript when it can be verified. Native, configDir and oauthTokenFile accounts support this retry; oauthTokenRef accounts rotate on the next launch only. In 2.0, an unresolved declared token refuses that account even if it has a configDir, rather than falling back to the directory's login. Token files must contain only the token, not captured sign-in output.
Eligible native and configDir logins with credential files can poll provider usage, with a default two-minute interval, an 85% rotation threshold and 95% operator alerts. Setup-token accounts and keychain-only native logins are not polled; token accounts retain stream-based rotation. Failed polls leave existing health evidence unchanged, and expired credentials skip the tick. The 1.10.1 fix stores poll results separately so they cannot overwrite the shim's rejection or credential-failure evidence. multi-clawd usage shows figures on demand.
Accounts can also serve OpenClaw's anthropic/* API route through explicitly configured setup tokens or existing profiles. multi-clawd direct sync stores missing profiles and applies pool-health ordering; existing sessions keep their pinned account until it cools down. Native rotating OAuth grants are not copied into direct profiles. This is the API route, not the Claude Code harness. Later 1.9 fixes improve resume handover and keep the shim outside OpenClaw's temporary plugin captures when the runtime copy succeeds.
Chain audits recognise agents.entries overrides and read supported SQLite session stores with a JSON fallback; SQLite inspection needs Node 22.13 or later. Doctor reports resolved login identities, duplicate pool logins and non-Claude auth-profile problems where it can inspect them. Token-account identity is not knowable at rest. Optional live probes target individual accounts. Catalog fixes add the documented context sizes for Sonnet 5, Fable 5.1, Opus 5.5 and Sonnet 5.5.
Short-window warnings without percentages can trigger rotation; numeric usage takes precedence. Model-scoped and recognised family-wide limits no longer need to bench unrelated models. Successful turns clear contradicted model rejections, and a wholly exhausted pool can re-test an overdue rejection. Purchased-overage utilization does not trigger rotation unless you opt in with rotateOnOverage; exhaustion alerts are rechecked on launch.
The CLI can import stable oauthTokenFile setup tokens into Hermes' native Anthropic pool, with dry-run support and an optional distribution strategy. It does not import native or configDir rotating grants or resolve oauthTokenRef secrets. Native logins need no copying: Hermes' own claude_code credential source reads the native ~/.claude/.credentials.json directly. As of Hermes 0.20.6 that source does not read an arbitrary configDir, so a configDir login needs its own oauthTokenFile to reach Hermes; otherwise it stays OpenClaw-only. Re-sync after token-file changes. Validated against Hermes 0.20.6, this uses Hermes' own tools and reactive failover, not the Claude Code harness or multi-clawd's proactive usage rotation.
Your fallback chain stops meaning “smaller model” and starts meaning “next account”.
"agents": { "defaults": { "model": {
"primary": "claude-cli/claude-fable-5", // main login
"fallbacks": [
"claw2/claude-fable-5", // 2nd login — same model
"claw3/claude-fable-5", // 3rd login? go on then
"anthropic/claude-opus-4-8" // only NOW drop a tier
]
} } }4
Claws
0
API keys required
CLI
Claude Code harness
MIT
Licence
Stop letting your second Claude account sit idle. Install multi-clawd and make your failover chain pool the capacity you already pay for.