A local-first desktop AI agent application (Windows / macOS / Linux, also available via Docker and command line): each “bot” has its own persona, default model, workspace, and memory, and can be accessed within the app or via WeChat, Feishu, Telegram, Slack, and Discord; it can read and write files, execute commands in a sandbox, control the desktop, launch background tasks and scheduled rules, and requires approval cards for operations outside the workspace. The agent runtime is the DeepSeek Harness, and the design of background tasks and the “Needs You” inbox draws inspiration from DeepSeek’s Super Agent prototype.
Star count 518 (+107 / 24h)
Reasons
- It turns the Super Agent concept, officially open-sourced in the previous period, into an application that ordinary users can install: completed background tasks send results back to the original conversation, rules can initially run in “observe-only” mode to record without executing, proactive messages only push results, pending decisions, and failures, do-not-disturb mode is enabled by default at night, and each task and rule has a budget cap, pausing to ask for confirmation when nearing the limit.
- Permissions are divided into three tiers (read-only / workspace / full access), with write access restricted to its own folder by default; conversations, memory, files, and keys are stored locally, and the local console listens only on 127.0.0.1.
- Database built with 10-27, collecting 107 new stars over the past 24 hours (indicating attention level only); this is the first inclusion in this column.
Usage
- Download the installer for your platform from GitHub Releases (SHA256SUMS are provided for verification); on first launch, log in or register with Modelflare in the browser, after which the model list appears automatically; for NAS / servers, use docker run -d --name amb -p 3600:3600 -v amb-data:/data ghcr.io/modelflare-ai/anymodel-bot, then retrieve the access token using docker exec amb cat /data/ui-token, and do not expose port 3600 to the public internet.
- Running from source requires Node.js 22.13+ and Go 1.26.8+: npm install, npm run setup (compiles the built-in gateway), npm start, local interface at http://127.0.0.1:3600/。
Risks
- Models can only be called via Modelflare (a paid API gateway operated by Havenbyte LLC in the US), billed per token, and the README states that Modelflare retains request logs; prompts, file contents, tool results, and desktop operation screenshots all leave the local machine. The README does not provide an option to connect directly to the official DeepSeek API, and the keys entered are for Modelflare.
- The README contains extensive Modelflare top-up promotions and registration links with tracking parameters, indicating the project itself serves as a customer acquisition channel for the gateway; pricing multipliers, cache hit compensation, and similar claims are vendor statements. Desktop operations execute automatically by default, and are in Beta on macOS; GitHub does not recognize a license type (the README states MIT).