AI Agent GitHub Digest — 2026-09-01

🌏 中文版

Today's Highlights

Today's thread is "long-running personal agents" moving from demo projects to something people actually depend on — nanobot and CowAgent both bet on a self-hosted personal-assistant approach built around a small core, multiple chat channels, and long-term memory. The infrastructure underneath is catching up too: conductor wires a durable-execution graph engine to native MCP tool calls, letting an agent's think-act loop survive a crash or even a weeks-long human approval wait, while context-mode goes straight at a pain point every heavy MCP user has hit — tool calls flooding the context window.

Trending Repos

HKUDS/nanobot ⭐ 47,574

GitHub · Python · MIT

• What it is: an ultra-lightweight, self-hosted personal AI agent framework that runs as a WebUI, terminal, or chat app, packing tool use, long-term memory (Dream), MCP integrations, model routing, multi-agent delegation, scheduled automation, and an OpenAI-compatible API into a small, readable core.
• Why it's worth a look: created in February 2026, it's already at 47.5k stars within half a year — remarkable growth for a framework whose selling point is a small core rather than a long feature list. It isn't trying to replace heavyweight orchestration frameworks like LangGraph or CrewAI; it's aimed at individuals who want to self-host a persistent personal assistant wired into several chat platforms at once.
• Tech Stack: a Python core plus a Bun-built TUI/WebUI, an OpenAI-compatible API, native MCP integration, and swappable LLM providers.
• Getting Started: Low — a one-line curl install script or pip/uv, with a guided walkthrough for non-technical users.

conductor-oss/conductor ⭐ 32,152

GitHub · Java · Apache-2.0

• What it is: a Netflix-originated open-source durable-execution / event-driven workflow engine now explicitly positioned as an AI agent orchestration layer, with native LLM tasks and MCP tool calling (LIST_MCP_TOOLS, CALL_MCP_TOOL).
• Why it's worth a look: unlike most code-first agent frameworks, conductor expresses orchestration as a versioned JSON graph — every step of an agent's "discover tools via MCP → LLM reasoning → call the tool → repeat" loop is durably persisted, can resume after a crash, and can pause for weeks awaiting human approval before picking up exactly where it left off. A different angle for teams that need production-grade durability instead of running the agent loop in memory.
• Tech Stack: a Java server with polyglot workers (Java/Python/Go/JS/C#/Ruby/Rust), five persistence backends (Redis/Postgres/MySQL among them), and built-in MCP task types.
• Getting Started: Medium — npm install -g @conductor-oss/conductor-cli && conductor server start gets a server running in under a minute, but a production deployment still means choosing a persistence and message-broker backend.

mksglu/context-mode ⭐ 20,281

GitHub · TypeScript · ELv2

• What it is: a tool that sandboxes raw MCP tool-call output and enforces routing to stop an AI coding agent's context window from being flooded by dumped data — the project's own example: a single Playwright snapshot can cost 56 KB.
• Why it's worth a look: this is a pain point nearly every heavy MCP user has hit — the project's numbers show 40% of the context window gone after 30 minutes, which forces a compaction that makes the agent forget which files it was editing or what it was asked to do. It hit #1 on Hacker News, and works across 17 different agent platforms via MCP plus hooks, positioning itself as a vendor-agnostic context-management layer rather than a single tool's plugin.
• Tech Stack: TypeScript, shipped as an npm package that's both an MCP server and a set of hooks, compatible with Claude Code, Cursor, Codex, Zed, and more.
• Getting Started: Low — install via npm as an MCP server/plugin.

zhayujie/CowAgent ⭐ 46,740

GitHub · Python · MIT

• What it is: formerly chatgpt-on-wechat (2022), now reborn as a full "Agent Harness" — it plans tasks, runs tools and skills, and self-evolves through nightly memory and knowledge distillation, all while bridging a dozen-plus chat channels (WeChat, Telegram, Slack, Discord, and more).
• Why it's worth a look: as one of the earliest "chatbot wrapper" projects, CowAgent's reinvention adds a three-tier memory architecture (conversation context → daily memory → core memory) and a nightly "Deep Dream" distillation pass, plus a skill marketplace — a real-world case study of how a simple wrapper project grows into full agent infrastructure as the ecosystem matures.
• Tech Stack: a Python core with swappable LLM providers (Claude, GPT, Gemini, DeepSeek, Qwen, GLM, and more), native MCP tool integration, and a macOS/Windows desktop client.
• Getting Started: Low — a one-line installer script or Docker compose; the rest is configured through the Web console.

Notable Releases

agno v3.0.4

Release Notes

• Key changes: KnowledgeManagementTools' constructor flags were renamed, and ingest_path flips from on-by-default to opt-in — the tool reads any path the server process can read, and under scope="shared" whatever it loads becomes readable by every agent on that knowledge base, so registering it is now a deliberate choice; the agno.tools.knowledge_management import path also moved to agno.tools.knowledge.
• Breaking Changes: the old enable_ingest / enable_remove flags are no longer recognized and are silently ignored if passed; the old agno.tools.knowledge_management import path is gone (a compatibility shim keeps FileGenerationTools's old path working).
• What it means for you: if your agent uses KnowledgeManagementTools to read and write a knowledge base, upgrading means path ingestion is off by default — you'll need to explicitly pass ingest_path=True to restore the old behavior. This turns "the agent can read any path the server can" from a default into a deliberate opt-in, which is a real security fix worth upgrading for rather than pinning around.

Today's Takeaway

I used to think self-hosted personal agents were mostly weekend demo projects. But watching CowAgent evolve from a 2022 WeChat bot into a three-tier memory architecture with nightly distillation, alongside nanobot hitting 47.5k stars in half a year, makes it clear this path already has real users depending on it long-term — "small core + long-term memory + multiple channels" is turning into a validated product shape, not just a demo.

References

• HKUDS/nanobot
• nanobot README
• conductor-oss/conductor
• conductor README
• mksglu/context-mode
• context-mode README
• zhayujie/CowAgent
• CowAgent README
• agno v3.0.4 Release Notes
• GitHub Trending — Daily