AI accounts

Mimir has no model of its own. Each bot uses an AI account you choose, on the bot's page.

Account	How Mimir uses it	Cost
Claude	Runs the official `claude` CLI (`claude -p`) with your signed-in subscription	Your plan
ChatGPT	Runs the official `codex` CLI (`codex exec`) with your ChatGPT login	Your plan
Grok	Runs Grok Build (`grok -p`) on your SuperGrok or X Premium+ plan	Your plan
OpenRouter	One key for hundreds of models (only ones that can use tools are offered)	What OpenRouter charges
xAI API	A per-token key from console.x.ai	Per token
Local model	Ollama (`localhost:11434`) or LM Studio (`localhost:1234`); nothing leaves your computer	Free
OpenAI-compatible	OpenAI or any compatible service, with a key and base URL	Per token

Web search. Bots on Claude search the web on their own (Claude's built-in search). For bots on any other account, set up Brave Search or your own SearXNG once in Settings → Web search, and every bot can search (how).

Signing in #

Claude / ChatGPT: install the CLI and sign in once in a terminal (claude, then codex login). The app shows whether it's installed and signed in. On a server, see Self-hosting. Since April 2026 Anthropic has at times counted Claude-plan use outside its own apps as paid "extra usage". If a reply fails because of that, the bot says so plainly; turn on extra usage in your Claude account, or use another account (ChatGPT's login is supported for tools like this).
Grok: Settings → AI keys → Sign in to Grok (a device code). Mimir keeps that login in its own Grok home (~/.mimir/grok), so your personal Grok plugins, hooks and tool servers never load into a bot.
OpenRouter: Connect OpenRouter signs you in through the browser and saves the key (OAuth PKCE with a one-off localhost callback). Or paste a key.
Keys (OpenAI-compatible, xAI) go to the OS keychain (or the server's private database).

How tools reach the model #

The CLIs (Claude, Codex, Grok) start Mimir as an MCP server for each run, so the model sees Mimir's tools (memory, routines, web pages, meetings…) the same way it sees any tool. API accounts get the same tools as function-calling definitions. Either way, every call passes through Mimir's rules (see Tools, approvals and safety).

Local models #

The app finds a running Ollama or LM Studio and lists its models. Models that can't use tools are flagged: they can chat but can't remember things or run routines. Ollama's default 4k-token context drops older messages in a long conversation (instructions and memory are kept); start it with OLLAMA_CONTEXT_LENGTH=16384 for more.

What a run costs #

The run log shows each reply's model, tokens and cost. For Claude it's the API list-price equivalent that claude -p reports (subscriptions aren't billed per token); OpenRouter runs show what OpenRouter charged; other key and local runs show tokens only.