Open Source Agent Tools: November 6, 2026

1. Poki Lets Claude Code, Codex or Its Own Agent Operate a Real iPhone Over MCP Without a Jailbreak

Poki, published by Joseph Zhou with v0.1.0 on November 5 and starred about 290 times on GitHub by the next morning, is a local MCP server plus a standalone agent that operate a physical iPhone or an iOS Simulator from a Mac. Claude Code and Codex connect with claude mcp add or codex mcp add and use the user’s own subscription, while Poki Agent runs tasks from the terminal or from iMessage and Telegram channels and remembers past tasks so repeats run faster. A physical phone needs no jailbreak or extra hardware: it is paired over USB, Developer Mode and UI Automation are turned on, and Apple’s Developer Disk Image is mounted, and the bundled native bridge uses the open-source idevice tools. The project publishes its own test log: on simulators, batching commands cut a form task from 8 round trips and 31.6 seconds to 3 and 20.0 seconds with Claude Code, and on one physical iPhone the final build dropped 0 of 24 touches on a search field, while taps next to a bottom floating bar in Settings still fail. Physical support has been verified only with the Chinese interface in portrait, and the README warns that Poki will send messages or make payments if asked and that screen contents go to the configured model service. It needs macOS 15 or later, full Xcode, Rust, Node.js 24 and ffmpeg, or a prebuilt Apple Silicon package. License: AGPL-3.0-only. Source Release

2. LATCH Defines a Five-Operation HTTPS Contract in Which Agents Quote, Get Approval for a Digest, Then Commit

LATCH, published on November 4 by Anton Balanov (author of the Shim harness covered on November 4), is a small JSON-over-HTTPS contract between an agent and a service: a card at /.well-known/latch.json and one POST endpoint with ask (reads without side effects), quote (an exact offer plus a SHA-256 digest, without executing it), approve (a local user or trusted-host decision bound to that digest), commit (the digest, the approval and an Idempotency-Key), status (the task plus a provider receipt) and cancel. The reference client, server and CLI are a single ESM file for Node.js 20 or later with no runtime dependencies, an independent PHP 8.2 provider verifies the same digests, an optional MCP stdio bridge exposes it to MCP hosts, and the CLI refuses to commit unless the user repeats the exact reviewed digest after --yes. It does not replace MCP, A2A, OpenAPI or agent payment protocols, has no registry, account, telemetry or payment rail, and its documentation states the limits plainly: a digest binds content rather than a person, a receipt is the provider’s assertion rather than physical evidence, idempotency cannot undo an external effect, and the default store is in-memory and unauthenticated. The wire format is experimental version 0.1 (implementation 0.2.0), and the demo’s 3D-print orders are simulated. License: MIT. Source

3. NestJS Publishes @nestjs/mcp, Turning Decorated Provider Methods Into MCP Tools Behind Nest Guards and OAuth 2.1

The NestJS project published @nestjs/mcp on November 4 (versions 0.0.1 to 0.0.3 on npm that day), an official module in which MCP tools, resources and prompts are methods of ordinary Nest providers declared with decorators and a schema in Zod, ArkType, Valibot or JSON Schema, and every call runs through Nest’s guards, interceptors, pipes and exception filters. Importing McpModule.forRoot() serves an endpoint at /mcp over Streamable HTTP from an existing Express, Fastify or Node adapter application, statelessly so it can run behind any load balancer, for clients of both the 2026-07-28 and the 2025 protocol revisions, and over stdio from a standalone application context. The module can act as an OAuth 2.1 resource server with protected resource metadata, WWW-Authenticate challenges, per-tool scopes and a user-supplied token verifier, ships an in-memory test client, is built on the official MCP TypeScript SDK (@modelcontextprotocol/server) and requires Zod 4.2 or later. License: MIT. Source

4. WebsiteBench Has Agents Rebuild 70 Websites Through a Browser Alone; the Best Model Passes 26.2% of Hidden Tests

WebsiteBench, published by TIGER-AI-Lab on GitHub and Hugging Face on November 4 and on PyPI on November 5, gives a coding agent up to 16 hours to reverse-engineer a full-stack copy of one of 70 expert-built local websites that reproduce features of widely used services such as Amazon, reaching the reference only through a browser (navigate, click, type, screenshot) and never its HTML, scripts or server code. Its 46,916 hidden tests (302 to 2,016 per task) fall into request, browser and system groups, and a case passes only when its functional checks pass and every visual checkpoint reaches an RGB SSIM of 0.99. On the project’s nine-model leaderboard, Claude Opus 5.5 leads with 26.2% of hidden tests passed while averaging about 2,235 turns and $305 per task, GPT 6.1 Sol follows at 20.2% for about $36, and the best open-weight model, GLM 5.3 Flash, reaches 14.7% for about $4.90; the authors attribute failures to underexploration, not implementing behavior the agent observed, and stopping early, with most runs ending well before the 16-hour limit. Runs use the Harbor framework (pinned at 0.25.0) with the agent and verifier in separate Docker environments of 8 CPUs, 8 GiB of memory and 20 GiB of disk each, the full task set is a 9.7 GB download, and the paper is listed as coming soon. License: Apache-2.0 for the code, task definitions and verifier; the reference sites’ content is for non-commercial research only. Source Leaderboard Dataset

5. Hermes Voice Puts a Hermes Agent Profile Behind Home Assistant Assist and a Repurposed Amazon Echo

Hermes Voice, released as v0.1.0 by the majorminorlabs GitHub organization on November 5, connects Home Assistant’s Assist pipeline (speech-to-text such as Whisper, text-to-speech such as Piper) to one profile of Nous Research’s open-source Hermes Agent, with an old Echo Dot or Echo Show running the community echo-dot-assist firmware as the microphone and speaker. Its Home Assistant conversation agent sends only the new user turn with a stable X-Hermes-Session-Id, so Hermes keeps the history, the profile’s SOUL.md personality and its long-term memory; it streams the reply so speech starts early, cleans it for speech and speaks a short fallback instead of error text. Sessions can be per conversation (a new one after five minutes of silence) or per device (reset after 30 idle minutes by default). The repository also ships a Hermes skill with which the user’s own agent prepares a voice profile and installs, configures and verifies Home Assistant, plus a reproducible training pipeline and model for a “hey hermes” microWakeWord wake word. Version 0.1 gives Hermes no Home Assistant device control; the README suggests giving the profile Home Assistant’s MCP server for that. License: MIT. Source Release