A resident AI agent that lives on your machine, not in someone's cloud

What it does

AnyModelBot is an Electron-based desktop agent app (Windows, macOS, Linux, plus Docker for servers) where each “bot” is a persona with its own default model, workspace, and memory. Bots can read and write files, run commands in a sandbox, and drive your desktop — and they ask before stepping outside their workspace. You can reach them from the app or from WeChat, Feishu, Telegram, Slack, and Discord, and they keep working in the background, surfacing only when they need a decision from you.

The interesting bit

The design borrows from DeepSeek’s Super Agent prototype and runs on the DeepSeek Harness, so the agent loop isn’t homegrown glue — there’s a lineage. And the model layer is deliberately decoupled: one Modelflare key routes to 70+ models across 13 providers (GPT, Claude, Gemini, Grok, DeepSeek, Kimi, GLM), with per-request usage and cost logging, so you’re not married to any single vendor’s subscription.

Key highlights

• Local-first: conversations, memory, files, and API keys stay on your device; only model requests transit Modelflare.
• Three permission tiers per bot: read-only, workspace (default, asks before escaping its folder), and full access without sandbox.
• Background tasks, scheduled rules, and approval prompts work out of the box — no extra configuration.
• Model picker shows context length, capability flags, and reference price next to each model.
• MIT-licensed, with SHA256 checksums published per release.

Caveats

• The README is candid that model requests are logged by Modelflare — “local-first” applies to everything except the model hop, and retention details live in their trust center.
• The app is tightly coupled to Modelflare as its model gateway; the pricing pitch (discounted GPT/Claude tiers, cache-hit guarantees) is prominent in the README, so treat this as an open-source client with a commercial backend.
• The README is largely in Simplified Chinese; there’s an English version, but the primary docs skew Chinese-first.

Verdict

Worth a look if you want a persistent, self-hosted personal agent with real file and desktop control and the freedom to swap models per conversation. If you want fully self-contained inference or zero third-party routing, this isn’t that — the model layer is Modelflare, full stop.