FAQ & troubleshooting.

Setup

The model never calls tools / ignores my vault. Agent mode needs a model that supports tool calling — most cloud models do; on local providers pull a tool-capable one (ollama pull qwen2.5:7b-instruct). Agent mode must also be enabled (Settings → Agent → Enable). If your "Agent model" override doesn't support tools, Curtis falls back to the chat model with a notice. → AGENT.md

The model picker is empty or missing a model I know exists. Most providers auto-discover their catalog from /v1/models at runtime; the built-in list is only a fallback. If discovery lags a provider's real lineup, use Add model on the provider card — manually added ids are always offered and never pruned. → PROVIDERS.md

A provider returns 400 "unsupported field". Strict APIs reject unknown request fields. Curtis gates known ones per provider; if a field still leaks through, check Copy last request JSON on the provider card to see exactly what hit the wire, and remove overrides (or use per-provider resets) until it's clean. → PROVIDERS.md

Where did my API key go / where is it stored? In the OS keychain (Obsidian 1.13+), never in data.json. ChatGPT sign-in tokens live there too. → PROVIDERS.md

Chats & data

Where are my conversations stored? Every chat is a markdown note in AI/Conversations/ — plain files, synced and searchable like any note. History from before this system (localStorage) imported itself once on upgrade. Deleting the file deletes the chat. → IMPORT.md

I deleted a chat / broke a conversation file by editing it. There is no database and no trash for conversation files — the file is the storage. Edit carefully: the <!-- curtis:msg --> markers carry machine metadata; keep them intact. Restore from a vault sync/backup if you have one, or re-import a .curt export. → IMPORT.md

How do I move chats to another vault or machine? Copy the AI/Conversations/ folder, or use Export all chats as .curt (zip) and import the zip on the other side — re-importing skips what's already there. Moving your whole setup (settings, prompt library, harnesses too)? Use Export full backup instead — it restores section-by-section, API keys excepted. → IMPORT.md, IMPORT.md

Privacy

What leaves my machine? By default: only what you send to the AI provider you configured. Everything networked beyond that is opt-in and listed in one sentence in llms.txt — web tools (DuckDuckGo/Jina), MCP server URLs, GCP, ChatGPT sign-in, dictation (Whisper), link favicons (one toggle, off = fully local rendering), and the opt-in outbound webhook URL (metadata-only events, off by default). Local providers + web tools off + favicons off = nothing leaves the machine. → PROVIDERS.md, VOICE.md, LINK_PREVIEWS.md

Can a local-only agent leak my notes to the cloud? Not through chat — but through shared memory it could, which is exactly what the per-agent memory: off tri-state severs. A memory-off agent's facts never enter the shared memory file. → AGENTS.md

Features behaving unexpectedly

The mic button does nothing. Your environment lacks MediaRecorder support or mic permission. All desktop Obsidian is fine; iOS 16+ and modern Android need explicit per-app permission (granted once, it persists). → VOICE.md

Text-to-speech is silent or uses the wrong voice. Voice lists load asynchronously — the settings dropdown fills in when the OS delivers them. A voice picked on one machine that doesn't exist on another silently falls back to auto. → VOICE.md

A scheduled run didn't fire. Curtis is an Obsidian plugin — nothing runs while Obsidian is closed. Due-checks run every 30 seconds while it's open; a daily job missed while closed fires once on next launch. Mobile needs the app in the foreground. → SCHEDULED_RUNS.md

Claude Desktop can't connect to the Curtis MCP server. The server binds 127.0.0.1 only, requires the bearer token from Settings → MCP server, and refuses non-loopback hosts and browser origins. Desktop only. The settings page has a copy-ready claude mcp add command. → MCP_SERVER.md

My arena run cost 4× what a normal message costs. Arena streams one prompt to every selected model — 2–4 columns means 2–4 full request streams, each billed by its provider. The picker warns; /stats shows spend. → ARENA.md

Autocomplete is sending my notes somewhere? Only if you turned it on — it's off by default. When enabled, each suggestion sends the text around your cursor to its configured provider. Dedicated model picker, throttled, cached. → AUTOCOMPLETE.md

Semantic search returns nothing / the agent says the index is missing. Build it: Settings → Vault retrieval → Rebuild index. If you changed the embedding provider or model, the old vectors don't mix — rebuild again. → VAULT_RETRIEVAL.md

Still stuck

Check the dev console (Ctrl+Shift+I) for errors, then file an issue with your Obsidian version, plugin version, provider + model, and the console output. Security issues go to [email protected] — never a public issue. → CONTRIBUTING.md