| PID | Type | Name | RSS | State | Uptime | Actions |
|---|
Message Queue
Channels
Anthropic Credentials
checking…Fleet Default Model
The warden pins every fleet inference request to this
model. Accounts with
model-selection enabled keep the model they choose per group;
everyone else runs this default. Set on the deployer via
WARDEN_DEFAULT_MODEL.
Sync Flip-and-Retry
When ON, a rate-limited inference (429 / auth failure) is flipped onto the next usable credential and re-issued in place — the agent never sees the limit unless every credential is exhausted (then it gets the contact-support reply). When OFF, the limit is forwarded and the flip applies to the next request. Costs a buffered retry only when a limit is hit.
Credential Flip Mode
--legacy marks a limited credential un-usable for hours and skips it (can strand a good account). pipe stops guessing: on a failure it walks every credential trying your real request, skipping only a dead-and-unrefreshable one, and pauses inference until a good credential is found. If all are rate-limited the caller gets a "system under load, try again in ~10 min" message. One walk at a time — storm-proof.
Overload Throttle
--Anthropic server-side limiting (HTTP 529 overload / Retry-After 429 — not your usage cap) trips a global cooldown. While cooling down, inference is queued and drained at a safe rate so we ramp gently instead of hammering a busy upstream. State is shared via Redis.
Fair-Share Auto-Block
--Automatically blocks machines that use far more than their fair share of inference, and unblocks them when they calm down. Everything is based on the last 24 hours of traffic, measured in requests. Machines you block by hand always stay blocked — this never touches them.
Details
Warden Credential Accounts
Log in via the embedded terminal (any account), then click Capture current login to add it here. To switch accounts, log in again and capture again. Use copies that account onto the live credentials file the warden reads — it takes effect immediately.
| account | tier | token | captured | actions |
|---|---|---|---|---|
| loading… | ||||
Credential Usage
Live window utilization per account from Anthropic (the data
behind Claude Code's /usage).
Binding = the limit closest to capping;
critical means it will 429.
Auto-refreshes every 10 min; expired tokens are refreshed
the warden-safe way (rotated + stored).
| account | 5h | 7d | binding | resets | status |
|---|---|---|---|---|---|
| loading… | |||||
Fleet — Warden Metering & Idle State
--| machine | state | served model | reqs | in tok | out tok | cache rd | last infer | block |
|---|---|---|---|---|---|---|---|---|
| loading… | ||||||||
Warden events
Fleet Analytics
--Top systems by requests
Traffic by day
Traffic by hour of day (UTC)
By model
By credential
Fleet events by day
Loading tasks...
Loading hook prompts...
Loading containers...
Loading groups...
Add Web Group
Share a folder with an existing group or use a new one
Select a group to manage its environment variables.
Per-group environment variables are injected into that
group’s agent container (the agent
can read them) and encrypted at rest in the
database. They override global vars of the same name (e.g.
OPENROUTER_API_KEY,
ANTHROPIC_BASE_URL). Existing values are shown
masked — leave a field untouched to keep it. Takes effect
on the group’s next message.
Loading swarm teams...
Select a web group to start chatting
Select a group to browse files
Select a file to preview
Published Artifacts
Knowledge Vaults
Linked Peers
Share this code with the other system. Expires in 5 minutes.
Paste the link code from the other system.
Loading peers...
Claude Code Router
checking...Transparent proxy that routes Agent SDK requests to any model provider. Config changes hot-reload — no restart needed.
Model Routing
Providers
Request Log
Host Database
messages.db — the orchestrator's own DB (groups,
messages, tasks, sessions, hooks, artifacts…). Edits here
affect the live daemon. Writes are confirmed; schema changes
are blocked.