DEV Community

Malachi
Malachi

Posted on

One Modelflare sign-in, 70+ models: the exact switching path inside AnyModelBot

#ai #llm #selfhosted #opensource

Hi, I'm Malachi, a co-developer of AnyModelBot. This post is only about how to see and switch models. No speed contest, no "model A beats model B" table, no invented latency numbers. If you want benchmarks, run your own prompts and read the public price list.

What you get after one sign-in

AnyModelBot is a local-first agent app (Windows / macOS / Linux, plus Docker). Models come from:

  1. Modelflare — one account, 70+ text / image / video models from many providers, pay per token, no AnyModelBot subscription; or
  2. Your own official API keys — OpenAI, Anthropic, Google Gemini, xAI, DeepSeek, Kimi, GLM, Qwen, MiniMax, Mistral, etc. Those calls go straight to the provider. Image / video generation still needs Modelflare today.

You do not run models locally inside this app. Network is required for inference.

Step 0 — Install and open the console

Desktop: download an installer from Releases (verify with SHA256SUMS if you care).

Docker:

docker run -d --name amb -p 3600:3600 -v amb-data:/data ghcr.io/modelflare-ai/anymodel-bot
docker exec amb cat /data/ui-token
# open http://<host>:3600/?token=<token>
# do not expose 3600 to the public internet
Enter fullscreen mode Exit fullscreen mode

Step 1 — Sign in once (or paste a key)

Desktop: first launch opens a browser to Modelflare. Approve, return to the app. No key copy-paste required for Modelflare.

Docker / server: browser callbacks do not reach the container. Sign in to Modelflare in any browser, then paste the landing link back into the app's first-run screen. Alternatively, paste an official provider API key and skip Modelflare for text models.

After connect, the model list fills automatically.

Step 2 — Read the picker before you click

Open the model button next to the composer (or Settings → Modelflare / Models).

For each row you should see, at least:

  • context length;
  • capability tags (vision / tools / reasoning, depending on the model);
  • a reference input/output price (from the public catalog — your bill is authoritative).

Also useful in Settings:

  • Balance, top-up link, low-balance reminders, last days of usage;
  • Per-bot budget keys when you want a spending ceiling on one persona;
  • Three API domains — modelflare.dev, origin.modelflare.dev, cf.modelflare.dev — compare response times yourself. AnyModelBot never auto-switches domains.

Step 3 — Switch for this conversation only (/model)

In the app or on a connected chat channel:

/model
Enter fullscreen mode Exit fullscreen mode

Pick another model for this thread. Typical pattern:

  1. Draft or brainstorm on a cheaper default.
  2. /model → a stronger coding or long-context model for the hard part.
  3. /model back when you are done.

Memory files do not rewrite themselves just because the model changed. Scopes are team / bot / group, not "which provider answered."

Other slash commands that stay useful: /reset (conversation state as documented in-app); on channels, /bot <name> to move the thread to another bot (accounts are shared; each bot keeps its own memory and workspace).

Step 4 — Set a default per bot

Create or open a bot → set its default model. Persona + default model + workspace + memory is the unit you live with day to day.

Permission presets stay independent of the model: Read-only / Workspace (default) / Full access.

Step 5 — Point background chores at a cheaper model

Long chats use a context steward (compact or hand-off summary). Memory tidying and similar background work can use a separate, usually cheaper model than the one you chat with. That picker is not the same control as per-conversation /model or the bot default — set it once and forget it unless costs surprise you.

We will not claim a dollar amount saved. Watch the in-app usage panel for a few days and adjust.

Honest limits (so you do not hit them mid-demo)

  • No offline / local LLM runtime (no Ollama-style "models on disk" path in product scope).
  • No standalone mobile app; channels talk to a host that must stay on.
  • Feishu channel chat: DMs only for now. Telegram / Slack / Discord channel chats: DMs only for now.
  • Computer use on macOS is Beta.
  • Rules do not run while the host is asleep — use Docker if you need closer to always-on.
  • Official keys: text models yes; image/video generation still Modelflare-only.

Minimal checklist

  1. Install desktop or Docker; open console.
  2. Modelflare sign-in (paste landing link on Docker) or official API key.
  3. Open picker; read context + reference price; note balance.
  4. Chat → /model → harder question → /model back.
  5. Set bot default; optionally set background / hand-off model cheaper.
  6. Optionally try another API domain in Settings and compare — manually.

That is the whole path. One account (or your own keys), visible prices, conversation-level switch, bot-level default, background-level thrift.

— Malachi, co-developer of AnyModelBot

Top comments (0)