With no API key configured, the system enters dry-run mode: LLM calls return simulated responses, the whole pipeline still runs, at zero cost. Learn the system first, connect a model later.

Pasted keys are stripped of stray whitespace, quotes, Bearer prefixes, full-width characters and variable-name residue; the provider is auto-detected; a minimal test request runs on save and tells you immediately: valid / invalid / out of quota. You never discover a bad key mid-task.

Set a daily spending cap. At 80% you get an alert; at 100% the breaker trips — running tasks are suspended instead of burning money. Worst case is capped. A runaway bill is structurally impossible.

When a model/provider keeps failing it gets circuit-broken; requests reroute to a backup key → backup provider → cheaper model. Retries use exponential backoff with jitter (1s→2s→4s…). No retry storm can ever eat your quota.

Every completed step snapshots its state. After a power cut, restart or network outage, the task resumes from the last checkpoint — paid intermediate results are never recomputed. You never pay twice for the same work.

A task that hangs past its timeout is flagged as a zombie, reclaimed, and re-queued (resuming from checkpoint). Nothing silently occupies resources forever.

All text bound for the cloud passes three detection layers (regex / entropy / NER): emails, IDs, API keys are replaced with placeholders and restored after the model responds. At the highest sensitivity level, egress is blocked outright. Your unpublished ideas never streak across third-party servers.

External content (papers, web pages) is wrapped in a data zone declared as "material, not instructions" before entering prompts; outputs are scanned for suspicious links. Malicious instructions hidden in documents can't hijack the agent.

Destructive or outbound operations never auto-execute — they queue for your explicit approval, then resume seamlessly from checkpoint. Denied means safely stopped.

With multiple users, retrieval is ownership-filtered before anything reaches the model context. Every call's model, tokens, cost, and I/O digests land in an append-only audit trail — you can always answer "which call produced this conclusion?"

A: Self-hosted backends need the litellm dependency: run pip install litellm on the server. The Docker image ships with it preinstalled.

A: Usually the daily budget breaker has tripped (check the Dashboard) or all workers are busy with long tasks. Raise the cap in Settings or wait for a worker to free up; suspended tasks resume automatically.

A: Everything lives in the data/ directory next to the deployment (SQLite database, PDFs, admin_password.txt). Back up that one directory and you have backed up everything.