FAQ
Can I run sidanclaw self-hosted?
Not as a turnkey product today. The repo is open source. You could deploy your own instance to your own GCP project, but support and updates target the hosted product.
Which models do you use?
Four tiers. Standard (1 credit per message, Gemini Flash 3 on a tighter tool budget) handles routine queries; Pro (2 credits, the same Gemini Flash 3 with full headroom) is the default; Max (10 credits, Gemini Flash 3.5) is the premium tier for hard reasoning and deep agentic work; Research (20 credits, Gemini Pro 3.1 on a deep multi-step budget) powers deep web synthesis. The tier picker lives in the chat header. See Pricing for the full menu.
Does my data train any model?
No. sidanclaw doesn't train on your conversations. Inference goes through Google's Gemini API under terms that prohibit model training on customer data.
What happens when I hit my plan budget?
Billing is monthly per workspace. When you run out of credits before the cycle ends, the default policy is a hard limit: new messages return 429 (or force-downgrade to Standard, depending on your workspace setting) until current_period_end. Admins can opt into paid overage at $30 per 1,000 credits. A separate daily LLM-cost safety cap may also force Standard for the rest of the day if a single 24-hour window blows past 3× the workspace's monthly plan price.
Can I export my data?
An export tool is on the roadmap. In the meantime, the database schema is documented; if you self-host, you have direct access. For hosted users, contact us for a one-off export.
How is memory different from the knowledge base?
Memory is per-(user, assistant): facts about you. The knowledge base is per-workspace: facts about the world. Memory is auto-extracted from chat; the KB is curated by you or synced from a GitHub repo. The assistant uses both, but they answer different questions.
Can the API stream responses?
Not in v1. We return one JSON body when the assistant turn completes. SSE is on the roadmap.
Can the public API call my Google Calendar?
No. The API path deliberately omits MCP write-tools because there's no human in the loop to approve confirmations. The model has memory + knowledge base + read tools (web search, fetch) only. Per-key tool allowlists are on the roadmap.