AI Agent GitHub Digest — 2026-10-21

🌏 中文版

Today's Highlights

Today's three trending repos pull the agent ecosystem in two directions at once: a major lab stepping into the coding-CLI fight directly (Moonshot's Kimi Code), and a multi-agent framework treating governance as a first-class feature rather than something bolted on after the fact (open-multi-agent) — while, on the other side, "getting past a site's defenses" gets packaged as a standard MCP server (stealth-browser-mcp). What agents can do keeps stretching toward both ends: writing more code, and getting past more gates.

Trending Repos

MoonshotAI/kimi-code ⭐ 7,900

GitHub　·　TypeScript　·　MIT

• What it is: Moonshot AI's (Kimi's) own terminal coding agent — it reads and edits code, runs shell commands, searches files, and fetches web pages, installed as a single binary with no Node.js required.
• Why it matters: A major Chinese lab shipping its own product to compete directly with Claude Code, Codex, and Cursor. Its differentiator is video input — drop in a screen recording or demo clip and the agent watches it directly; the README's own examples include turning a reference clip into a LUT, a long video into a short cut, or a screen recording into working code. It also speaks the Agent Client Protocol natively, so kimi acp can plug straight into Zed or JetBrains as the underlying agent, and MCP servers are configured conversationally via /mcp-config instead of hand-editing JSON.
• Tech stack: TypeScript core + pi-tui (terminal UI framework) + Agent Client Protocol + MCP
• Getting started: Low — one official install script (curl ... | bash), no Node.js needed; log in with Kimi Code OAuth or your own API key.

open-multi-agent/open-multi-agent ⭐ 7,000

GitHub　·　TypeScript　·　MIT

• What it is: A self-hosted multi-agent runtime that can pause before planning, dispatching a task, or calling a tool for human approval, and records every run so it can be inspected offline afterward.
• Why it matters: Unlike other multi-agent orchestrators covered here recently — the desktop-app Orca, the enterprise-facing Paperclip — OMA puts governance first. Its README states plainly that it makes no analytics, license, update, or phone-home request: no telemetry, no hosted control plane. verifyRun() checks a run's recorded content for internal consistency offline, though the project is upfront that this doesn't prove the journal was never edited. It also has a "governance floor": declare which roles and execution order a run must satisfy, then check separately whether it actually did. External agents — Claude Code, Gemini CLI, Codex — can be put on the same task graph through process and ACP backends.
• Tech stack: TypeScript + Node.js 20+ + built-in adapters for Anthropic, OpenAI, Azure, Bedrock, Gemini, Grok, DeepSeek, and more + optional OpenTelemetry
• Getting started: Low — npm create oma-app@latest runs a local demo in one command with no API key needed; wiring up real approval flows and governance rules takes longer with the docs.

vibheksoni/stealth-browser-mcp ⭐ 3,200

GitHub　·　Python　·　MIT

• What it is: An MCP server that lets Claude Code, Cursor, or Claude Desktop drive a real Chrome, Chromium, or Edge instance (via nodriver and the Chrome DevTools Protocol) through Cloudflare's JS challenge page and ordinary login walls.
• Why it matters: A fundamentally different approach from the usual Playwright MCP — it never touches WebDriver, so the browser's fingerprint (TLS/HTTP2 handshake, navigator.webdriver, GPU rendering) all comes from a genuine Chrome binary, which gives it a higher pass rate on Cloudflare's non-interactive JS challenge than typical automation. Its 98 tools cover network interception and rewriting and pixel-accurate DOM cloning. The project is also upfront about its limits: it has no answer for IP reputation or interactive CAPTCHAs (Turnstile, image puzzles), and GPU-less Linux boxes still leak a SwiftShader software-renderer fingerprint.
• Tech stack: Python 3.10+ + nodriver (CDP) + FastMCP
• Getting started: Medium — you git clone, build a venv, and pip install yourself rather than running one npx command; routing through a proxy and trimming the 98 tools down to what you need both take some time with the docs.

Notable Releases

pydantic-ai v2.58.0

Release Notes

• Key changes: Fixes two security advisories — GHSA-xv4x-g8mx-ch56: provider-native MCP configuration could expose server URLs, authorization tokens, and custom header values in OpenTelemetry request traces even with include_content=False; GHSA-5pfm-2g3g-gh85: with a non-empty message_history and the default manage_system_prompt='server', the Vercel AI and AG-UI adapters could let a client execute an available tool with its own arguments without a model request at all (tools requiring approval aren't affected). On the feature side: timeout= to bound a whole agent run, DelegationTasks.steer to redirect a delegated task mid-flight, and new ElevenLabs Agents realtime voice model support.
• Compatibility Notes: Filed as "Compatibility Notes" rather than breaking changes, but they do change existing behavior — RunUsage/UsageLimits now count a resumed response's usage only in the run that produced it, RealtimeConnection lifecycle events move to a new public pydantic_ai.realtime.lifecycle module, and read_file now returns images and PDFs instead of just describing them.
• What this means for you: If you use provider-native MCP or the Vercel AI/AG-UI adapters, upgrade to v2.58.0 (or v1.107.8) now and rotate any credentials that may already have leaked into your tracing system. Everyone else should check the compatibility notes against their own code to make sure nothing's behavior quietly changed underneath them.

Today's Takeaway

I used to think multi-agent frameworks mostly competed on doing more. But open-multi-agent's verifyRun() is a reminder that "can do it" and "can do it and leave evidence behind" are two different things — automation with no audit trail means that when something goes wrong, you can't tell which agent did what, at which step. That may be the problem multi-agent systems actually need to solve first, ahead of one more layer of clever planning logic.

References

• MoonshotAI/kimi-code — GitHub
• open-multi-agent/open-multi-agent — GitHub
• vibheksoni/stealth-browser-mcp — GitHub
• Stealth Browser MCP: get your AI agent past Cloudflare — Roundproxies
• pydantic-ai v2.58.0 Release Notes — GitHub
• GHSA-xv4x-g8mx-ch56 — MCP request telemetry security advisory
• GHSA-5pfm-2g3g-gh85 — UI adapter history security advisory