Skip to main content
When you report a routing problem, include the session ID (/router-session in Claude Code, or /status) and roughly when it happened. That is what lets Weave find the exact decisions behind it.

​
Billing and access

Your organization’s balance hit zero and the request presented no subscription credential that could cover the turn. An org admin can top up at Organization settings → Weave Router. Turn on auto-recharge to avoid a repeat.
Same cause, softer landing: the balance is empty, so the turn ran on your Claude or ChatGPT plan and paid-model fallback is switched off until credits are added. Routing quality is reduced in this state because the router can only choose plan-covered models. Codex shows the same warning with ChatGPT (Codex) in place of Anthropic.
You've reached your monthly Weave Router spend limit means the engineer’s monthly spend limit was reached. The org limit and per-key caps have their own messages: Your organization has reached its monthly Weave Router spend limit and This router key has reached its spend cap. Limits bound paid spend: after the org limit or a per-key cap, subscription-covered turns keep flowing unless subscription routing is Use prepaid credits only; after a per-engineer limit, they keep flowing only when subscription routing is set to Strict pass-through until low. Raise the limit, switch to that mode, or have them run npx @weave-os/router off --claude (then quit and reopen the client) to use their subscription directly in the meantime.
Turns served on your Claude or ChatGPT subscription don’t consume credits — pass-through is working as intended. If you want to price the workload without the subscription, switch subscription routing to Use prepaid credits only.

​
Routing behavior

Move the quality-versus-price dial toward price in Models & routing → Routing priority and save. If your team is on Claude subscriptions, the router deliberately favors Claude models — that quota is prepaid, so keeping work in-family is usually the cheapest option. If the mix still looks wrong after a change, send Weave a couple of session IDs.
An org admin can uncheck it in Models & routing → Models, or uncheck the whole provider under Providers, then save. The router stops selecting it, including as a failover target for providers.
This usually means a session was pinned to a model the client can’t handle — most often after resuming a conversation that started before the router was installed. Clear it with /unforce-model, or start a fresh session. If it repeats, send the session ID to Weave.
Claude Code’s native percentage and the router’s approximate request estimate measure different things. A served-model window can also differ by provider and from your client’s local budget. Context is omitted when fields are invalid or unavailable, the router snapshot is older than five minutes, or an asynchronous fetch fails. Codex uses a terminal title and emits no status message in the conversation. Refresh happens on lifecycle events, so rerun the installer if your managed helper predates this support. Existing custom status lines and unrelated hooks are preserved.
The router filters candidates using a conservative per-request estimate. If every model is excluded, it admits the widest eligible candidates so the provider can decide actual fit. A genuine overflow is returned in the client’s native form so the harness can compact. In Claude Code, try /compact; you can also start a new session or use /force-model <model-id> to select a larger window.
In Strict pass-through until low, Anthropic-model turns still forward your subscription token — so seeing a Claude model other than the one you picked is not the same as being billed per API call. If the requested model is being substituted below your threshold, capture a screenshot plus the session ID and send it to Weave.
Grab the session IDs and the time window and contact Weave. Unexpected cache misses have caused this before and are treated as a bug, not expected behavior.

​
Install and client issues

The status line script needs jq on PATH. Install it, then run claude --debug and check stderr for errors invoking cc-statusline.sh.
Claude Code reads its router setting at launch — fully quit and reopen it. Codex and opencode pick the change up on their next run. Cursor is toggled in Settings → Models, not by the installer.
Remote Control requires Claude Code to reach the Anthropic API directly. It is unavailable when ANTHROPIC_BASE_URL points to a host other than api.anthropic.com, including Weave Router. An eligible claude.ai subscription login does not remove this custom-endpoint restriction.Turn routing off for your install scope:
Fully quit and reopen Claude Code after the toggle: it reads the configuration at launch. Requests then go directly to Anthropic and bypass Weave Router; they are not routed, measured, or governed by Weave Router. The toggle preserves your Router configuration. Weave Router does not proxy Remote Control or automatically switch routing for it.When finished with Remote Control, restore routing with the matching command:
Fully quit and reopen Claude Code again to resume routing.If Remote Control still fails with routing off, follow Anthropic’s current requirements and troubleshooting. Remote Control requires an eligible Pro, Max, Team, or Enterprise subscription and claude.ai sign-in; API-key authentication is not supported. Team and Enterprise organizations also need an Owner to enable Remote Control. Organization policy, cloud-provider routing, and network access can cause independent failures.
On npm 6 and older, use npx --package @weave-os/router -y -- weave-router --claude, or upgrade with npm i -g npm@latest.
They are different products. Remove the router with npx --package @weave-os/router -y -- weave-router --uninstall --scope user, then follow the Claude Code plugin install.
Rotating or revoking a key invalidates the old token immediately. Re-run the installer on each machine with the new key.

​
Reporting

Data normally lands within about an hour. Beyond that, contact Weave with the time range you’re missing.
Savings are measured against the model your agent requested, not against another vendor’s pricing. See how savings are calculated.
Still stuck? Email support@workweave.ai or message your shared Slack channel with the session IDs.
Last modified on October 17, 2026