惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

罗磊的独立博客
爱范儿
爱范儿
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
让小产品的独立变现更简单 - ezindie.com
让小产品的独立变现更简单 - ezindie.com
博客园_首页
博客园 - 叶小钗
酷 壳 – CoolShell
酷 壳 – CoolShell
Apple Machine Learning Research
Apple Machine Learning Research
云风的 BLOG
云风的 BLOG
量子位
博客园 - 三生石上(FineUI控件)
Stack Overflow Blog
Stack Overflow Blog
小众软件
小众软件
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报
V
V2EX
人人都是产品经理
人人都是产品经理
V
Visual Studio Blog
Jina AI
Jina AI
L
LangChain Blog
M
MIT News - Artificial intelligence
MongoDB | Blog
MongoDB | Blog
Last Week in AI
Last Week in AI
Martin Fowler
Martin Fowler
WordPress大学
WordPress大学

Hacker News: Show HN

PurrrrrFocus: Pomodoro Timer App - App Store Workflow Engine — Multi-Step Orchestration for Bun RapidPhoto: Pro Photo Editor App - App Store GitHub - DheerG/swarms: Achieve extraordinary results with claude code across a variety of tasks SPICE simulation → oscilloscope → verification with Claude Code — Lucas Gerads Show HN: VCoding – A 5 MB native Windows IDE with no dynamic dependencies Show HN: LLMs don't hallucinate because they're bad at math, it's the format GitHub - Agent-FM/agentfm-core: AgentFM is a peer-to-peer network that turns everyday computers into a decentralized AI supercomputer. AgentFM lets you run massive AI workloads directly across a global mesh of idle CPUs and GPUs. Show HN: Tracking Top US Science Olympiad Alumni over Last 25 Years GitHub - Potarix/agent-hub: One place to talk to all your agents Show HN: Runtime security for AI agents(injection,tool abuse, data exfiltration) GitHub - dubeyKartikay/lazyspotify: Terminal Spotify client for macOS and Linux GitHub - the-banana-tool/king-louie: Easy to use GUI Personal AI Assistant. Win/Linux/Mac. Show HN I made my vacation rental bookable by AI agents–no Airbnb, 0% commission GitHub - basteez/jsf-autoreload: maven plugin to enable hot reload on jsf projects uvm32/hosts/host-gdbstub at main · ringtailsoftware/uvm32 GitHub - labsai/EDDI: Config-driven engine that turns JSON into production-grade AI agents. Multi-agent orchestration, 12+ LLM providers, MCP/A2A protocols, RAG, persistent memory, and enterprise compliance (EU AI Act, GDPR, HIPAA). Built on Quarkus. GitHub - glitchnsec/fortyone-oss: AI Executive Assistant Platform Quickstart | Alien GitHub - muxshed/shed: One stream in, or many. Every destination, simultaneously. No cloud middleman, no per-channel fees, no limits. GitHub - ocrbase-hq/ocrbase: 📄 PDF/IMG ->.MD/JSON Document OCR API for PaddleOCR and GLMOCR. Self-hostable. GitHub - impactjo/home-memory: MCP server that lets your AI assistant remember everything about your home. GitHub - Sets88/dbcls: DbCls is a powerful terminal database client that supports various databases GitHub - neptun2000/heor-agent-mcp GitHub - SeanFDZ/macmind: Single-layer transformer in HyperTalk for the classic Macintosh RollQuation: Math Puzzles - Apps on Google Play GitHub - dropbox/witchcraft Show HN: Agent-cache – Multi-tier LLM/tool/session caching for Valkey and Redis GitHub - opentalon/opentalon: OpenTalon is an open-source platform built from the ground up in Go as a robust alternative to OpenClaw LinkedIn™ 职位抓取工具 - Chrome 应用商店
MCPlexer — Cross-Harness AI Runtime
maxrev17 · 2026-06-25 · via Hacker News: Show HN

open source · cross-harness AI runtime

Cross-harness runtime for serious AI work.

MCPlexer gives Claude Code, Codex, OpenCode, Cursor, Grok, Pi, Gemini, and other MCP clients the same operating layer: delegations, workers, tasks, memory, browser control, approvals, audit, workspaces, sandboxing, and routed MCP servers.

$ one slim surface · mcpx__search_tools + mcpx__execute_code + secret refs

liveDelegation ledger and worker review loop liveAI-first tasks, memory, mesh, and skills wiredBrowser control available across harnessesApprovals, audit, workspaces, and guards active

Built for the harnesses and tools agents actually use

  • Claude Code
  • Codex
  • OpenCode
  • Cursor
  • Grok
  • Pi
  • Gemini CLI
  • Windsurf
  • custom MCP clients
  • local daemon
  • AGPL source

Full feature list

One runtime for the whole agent operating model.

The important part is not a single dashboard. It is that the same primitives are available across harnesses: work can be delegated, remembered, audited, approved, routed, restricted, and continued.

Harness

One slim tool surface

Every harness sees discovery, batched execution, and secret refs instead of loading a giant static tool list into context.

Delegations

Cross-harness delegation

Delegate between models and harnesses with ease: keep the parent strategic while workers inspect files, run tests, draft changes, and return for scored review.

Workers

Durable AI workers

Manual, scheduled, mesh-triggered, and delegation-backed workers with model profiles, caps, approvals, and live run output.

Tasks

AI-first task manager

A work ledger designed for LLMs: stable IDs, leases, status events, offers, assignments, context packets, and attachments.

Memory

Cross-harness memory

Persistent facts and notes survive model swaps, client swaps, and machines, with consolidation, invalidation, and sharing.

Browser

Remote-capable browser control

Drive visible Chrome, Playwright sessions, and desktop surfaces through the same routed MCP layer, including worker and peer workflows.

Workspaces

Directory-scoped routing

The real working directory decides which servers, tools, auth scopes, routes, and approval policies are available.

Restrictions

Sandboxing and tool gates

Capability presets, allowlists, deny rules, shell guard, sanitizer guard, schedule guard, sandbox guard, and downstream sandboxing.

Approvals

No self-approval shortcut

Sensitive calls pause for review in another session or the dashboard, with OS notifications and a worker approval queue.

Audit

Everything leaves evidence

Tool calls, worker dispatches, model sends, denials, approvals, mesh triggers, memory events, and route decisions are searchable.

GitHub, Linear, Slack, Gmail, Calendar, Postgres, Vercel, WordPress, WooCommerce, Reddit, Telegram, browsers, and custom servers.

Secrets

Auth scopes, not pasted tokens

OAuth and API keys are encrypted at rest, injected only into matching routes, and kept out of transcripts through secret refs.

Mesh

Agent-to-agent coordination

Agents discover peers, exchange findings, offer tasks, send targeted questions, and coordinate across paired machines.

Search, install, publish, compose, and sync skills so agents load only the guidance they need for the current job.

Models

Model routing and ranks

Capacity routing, provider profiles, cost dashboards, review scores, and leaderboards show which cheap models actually work.

Dashboard

Installable operations view

PWA dashboard for approvals, audit, live signals, workspaces, routes, servers, model providers, backups, and recovery.

Plan

Delegate the work that burns context

Frontier sessions define architecture, scope, and acceptance criteria. Workers do broad inspection, implementation, test loops, and log triage in bounded contexts.

Run

Keep agents working across sessions

Tasks, memory, skills, and mesh give agents durable coordination primitives instead of making every chat rediscover the same project state.

Control

Expose power without losing the boundary

Workspace routing, approvals, audit, tool restrictions, sandboxing, and encrypted auth scopes make powerful MCP servers usable in real repos.

Browser control

Beats Claude Chrome because it is a shared runtime capability.

A browser tied to one chat helps that chat. MCPlexer achieves browser control with brw — a real, visible Chrome window for agents — exposed as routed infrastructure, so the capability works across any harness, durable workers, and paired machines while staying auditable.

Explore brw
  • Cross-harness: Claude Code, Codex, OpenCode, Cursor, Grok, Pi, Gemini, and workers can all use it.
  • Remote-capable: a paired machine or worker can use the browser near the login, network, display, or local service.
  • Semantic-first: snapshots, find, click text, type, press, tabs, and screenshots as fallback.
  • Audited and scoped: browser actions are still routed tool calls with approvals and workspace policy.
  • Visible or clean-room: use real Chrome tabs for human-visible state, or Playwright sessions for automation.

MCP server fabric

Wire up broad capability without flooding every context.

Downstream servers sit behind routes, auth scopes, workspace policy, approvals, and audit. Agents discover what they need when they need it.

Code

GitHubLinearClickUpVercel

Comms

SlackGmailCalendarTelegramOpenWA

Data

PostgresSQLitefetchcustomer MCP

Browser

brwPlaywrightHammerspoon

Content

WordPressWooCommerceObsidianExcalidraw

AI

LM Studiomodel providersskillsmemorytaskmesh

AI-native setup

Configure it by talking to your agent.

Inside ~/.mcplexer, the agent can provision servers, import OpenAPI specs, write routes, set auth scopes, inspect audit, and manage approvals. In ordinary project repos, it only sees the slim universal surface.

> Wire GitHub, Linear, Gmail, Calendar, Postgres, browser control, and a cheap worker model into this workspace. Require approval for write tools and give Codex the same surface Claude Code gets.

Power with boundaries

Serious agents need serious constraints.

Human approvals with self-approval blocked

Searchable audit with redaction

Encrypted auth scopes and secret refs

Sandboxing and tool restrictions

Read our security and design notes

Build agent systems that continue across tools, models, and machines.

MCPlexer is open source infrastructure for people pushing past one model, one chat, and one tool list.