Table of Contents
🌏 中文版
Today's Highlights
Four AI Agent projects on today's trending list share a strikingly consistent theme — they all move agents away from "cloud black box + abstract API" toward "local machine + filesystem." Cursor open-sources its official plugin marketplace to standardize the ecosystem, apache/maka uses event logs to make every step auditable, magnitude makes local models work out of the box, and vercel/eve lets the filesystem itself serve as the interface. It looks like the competitive focus in the second half of 2026 is shifting from "how much can an agent do" to "can you understand, audit, and run it offline."
Trending Repos
cursor/plugins ⭐ 3,960+ (+470 today)
GitHub · TypeScript · MIT
- What it is: Cursor's official open-source plugin marketplace — a monorepo containing 30+ plugins, each defined by its own
plugin.jsonmanifest plus agent skills, Cursor rules, and MCP server definitions, with a root-level marketplace manifest listing them all. - Why it matters: This is the second major coding agent (after Claude Code plugins) to lay out its "plugin ecosystem" in an open-source standard file format. Plugins fall into three categories — dev tools (agent workflow, code review, CLI design), productivity (Gmail, Google Drive, Calendar), and integrations (GitHub, Salesforce, HubSpot, Zoom, Playwright) — effectively pushing agent capability extensions from proprietary formats toward "one manifest, run anywhere."
- Tech stack: TypeScript,
plugin.jsonmanifest +SKILL.md+ MCP server definitions. - Getting started: Easy — just place files in the right directory following the manifest format. The structure closely resembles Claude Code / Codex plugins.
apache/maka (Incubating) ⭐ 1,900+ (+360 today)
GitHub · TypeScript · Apache-2.0
- What it is: A "local-first AI agent workbench" now in the Apache incubator. It records model messages, tool calls, tool results, permission decisions, and termination events as an append-only event log, with Desktop, terminal, and CLI interfaces.
- Why it matters: It treats "auditability" as a first-class citizen — the event-sourcing architecture leaves a trace for every agent action, enabling replay and recovery. This is valuable for teams that need to explain "what exactly did the agent do, and who approved it." Tools are local file operations (Read / Write / Edit / Bash / Glob / Grep); models can connect to cloud APIs, local models, or compatible gateways, with data staying on your machine.
- Tech stack: Electron + React frontend; Node.js 22+ + SQLite + TypeScript backend, with an event-sourcing architecture spanning Runtime Host / AgentRun / model adapter / tool runtime.
- Getting started: Medium — the concepts are straightforward, but getting full value from the event-sourcing workbench requires understanding sessions, permissions, and the replay model.
magnitudedev/magnitude ⭐ 1,440+ (+130 today)
GitHub · TypeScript · Apache-2.0
- What it is: An open-source agent framework with "built-in inference" — it auto-profiles your hardware, recommends a suitable local model, downloads it, and runs everything locally without needing Ollama or a separate inference server.
- Why it matters: It removes the entry barrier for local agents. Previously, running an offline agent meant wiring up Ollama and configuring an inference server yourself. This packages "detect hardware → pick model → download → run" into an out-of-the-box experience — fully offline after first setup, zero API cost, zero data leakage. A practical option for anyone who wants to run agents on local data without touching the cloud.
- Tech stack: Bun runtime + TypeScript monorepo (Turbo), including CLI, desktop app, and web components, with skills-based extensibility.
- Getting started: Easy — designed for out-of-the-box use, with automated hardware profiling and model downloading.
vercel/eve ⭐ 4,700+
GitHub · TypeScript · Apache-2.0 (Public Beta)
- What it is: Vercel's "filesystem-first" durable agent framework — agent capabilities live in convention directories (
tools/,skills/,channels/,schedules/), with behavior defined through human-readable file structures likeinstructions.md, tool files, and skill programs rather than abstract APIs. - Why it matters: Unlike frameworks that emphasize code-centric or config-driven approaches, eve lets the filesystem itself be the interface — you can understand and modify what an agent does through ordinary file browsing, no API docs required. Combined with durable execution, agents can run long-term and recover from failures. Compared with munder-difflin and apache/maka, "filesystem as interface" is solidifying into a distinct design school.
- Tech stack: TypeScript / JavaScript + Zod schema validation, connecting to Claude / OpenAI and other models via AI gateway.
- Getting started: Easy — the convention directory structure is intuitive. However, this is a Vercel beta, so APIs and behavior may change before GA.
Notable Releases
pydantic-ai v2.32.1
- Key changes: A pure bug-fix patch — rejects calling
Agent.run_sync()from a synchronous callback (avoiding event loop conflicts);FunctionModelnow accepts any callable. - Breaking Changes: None. The previous v2.32.0 (covered yesterday) added instrumentation v6, xAI attachment search, and OpenRouter source attribution.
- Impact: If you're on pydantic-ai v2.32.0, this is a safe patch upgrade with no code changes needed. If you've hit the
run_syncevent loop issue, upgrade.
Takeaway
I expected the next step in the agent ecosystem to be about who has more powerful cloud orchestration or a more full-featured managed platform. But today's four trending projects — Cursor's open-source plugin manifest, apache/maka's append-only audit log, magnitude's out-of-the-box local models, and eve's filesystem-as-interface — all converge on "local, readable, auditable, offline-capable." This tells us the market is actually filling a gap that cloud agents skipped: when agents start touching real files, permissions, and money, what people want isn't "a smarter black box" but a transparent structure they can understand, shut down, and rely on.
References
Loading...