Cline desktop agent imports Claude Code and Codex sessions

Overview

10 stories in this issue. The first 3 are today's priorities.

Hot model news

1. Top · Cline ships a desktop coding agent that imports Claude Code sessions
2. Top · Apple's rebuilt Siri runs on Google's Gemini, but not in the EU
3. Top · SealGate lets Grok drive a Flipper Zero from an Android phone

Global AI news

4. OpenAI acquires Glass Imaging and its former Apple camera engineers
5. Agent-net open-sources Webagent, a Go harness under Apache-2.0
6. Ninth Wave builds open finance onboarding on Bedrock AgentCore
7. Grab standardizes more than 500 internal agent services with LLM-Kit

Regional and early signals

8. StepFun ships five StepAudio 3 models, claiming top rankings
9. Infinigence, Tsinghua and SJTU open-source APXInf for on-robot inference
10. Feishu and Doubao Work put a team agent inside group chats

Jiufeng graphic based on the sources cited in this issue.

Hot model news

Cline turns its open-source coding agent into a standalone desktop app that can pick up tasks from other agents and continue them on a different model.

Cline founder Saoud Rizwan (@sdrzn) launched Cline Desktop on September 14th, a beta for macOS and Windows that works with Cline accounts, outside API providers, or locally hosted models. The launch thread says users can import tasks from Claude Code, OpenAI Codex and other coding agents, then continue those sessions with a different model.

The desktop app also schedules recurring jobs such as pull request reviews, security scans and documentation updates, and a built-in marketplace manages plugins, Model Context Protocol servers and reusable agent skills. Outside VS Code, users can point Cline at a workspace and let the agent read files, make edits, execute commands and track the current Git branch.

Limitations: The release is a beta. Session imports, scheduling and marketplace distribution all arrive with it, and the reporting gives no general-availability date and no detail on how ClinePass quotas differ from bring-your-own-model use.

Image source: GitHub; mirrored on Jiufeng R2.

Source: RuntimeWire · Cline on X · GitHub

After years of delay, Apple shipped its overhauled assistant as a beta built on Gemini models.

According to Apple, "Siri AI" is now available as a beta, in English only for now. It runs on Google's Gemini models and processes data partly on the device and partly through Private Cloud Compute. The assistant can read screen content and personal context from messages or photos, and carry out tasks across different apps. TechCrunch editor Ivan Mehta, who has been testing iOS 27 since the beta, describes the new Siri as a step forward.

Limitations: The service will not launch in the EU or China for now, and English is the only supported language at this point. Early tests praise the practical improvements but fault ongoing misunderstandings.

Source: The Decoder · TechCrunch

A September 14th demo routes a cloud chatbot into hardware sitting next to the phone.

SealGate founder Eito Miyamura (@Eito_Miyamura) posted a video showing Grok on an Android phone operating a Flipper Zero, which then controlled a hotel-room television over infrared. The connection runs over Model Context Protocol, and SealGate's Android client is published as open source on GitHub. Miyamura previously worked on self-driving infrastructure at Wayve; his London startup SealGate, formerly EdisonWatch, sells a control layer for monitoring and restricting how AI agents reach company data and external tools.

Limitations: The public evidence is a single demo covering one scenario, infrared TV control. Per the report, the setup depends on Grok's custom connectors, which require a paid tier, while Android runtime permissions and on-device authorization prompts still apply; the gateway monitors tool calls, assigns risk scores and enforces deterministic policy. The report notes that once cloud agents can operate nearby hardware, permission controls and auditable tool calls become part of device security.

Source: RuntimeWire · Eito Miyamura on X · GitHub

Global AI news

The team behind iPhone Portrait Mode technology joins OpenAI's hardware group.

The Wall Street Journal reported on September 14th that OpenAI acquired Glass Imaging in recent months, in a deal valuing the company above $300 million. Glass Imaging was founded in 2019 by former Apple camera engineers Ziv Attar and Tom Bishop, and built a neural image processor after years spent extracting better photographs from the constrained lenses and sensors inside smartphones. Attar had already founded LinX Imaging in 2011 and ran it until Apple acquired it in 2015; LinX developed dual-camera systems for depth effects and low-light image fusion.

Limitations: The deal terms and timing come from the Journal's reporting. OpenAI has not said publicly which product line the team joins, and the idea that cameras will serve as a primary device input is the report's framing, not an official statement.

Source: RuntimeWire

Fill in a declarative JSON spec, pick one provider per slot, and run webagent serve.

Agent-net released Webagent, an open source harness for standing up public-facing business agents. Instead of writing orchestration code, a business fills in a declarative JSON spec and picks one provider for each pluggable slot:

• Language and license: Written in Go, shipped under Apache 2.0, and the repo builds green
• Structure: 1 Brain (an LLM plus instruction) over 9 pluggable slots, each a service provider interface with a registry in spi/ and a designated default
• Running today: Live Slack, WhatsApp and HTTP agents backed by MCP tools
• Design basis: DESIGN.md frames the project around a research finding that architecture, not model capability, decides agent success, citing arXiv 2511.19477

Limitations: The project is still labeled v0. The browser action provider, OAuth-gated MCP, OTel export, and the AgentNet identity and billing layer are listed as not yet built.

Source: MarkTechPost · arXiv 2511.19477

Validating bank APIs against the FDX standard and scoring readiness is handed to a managed multi-agent system.

Ninth Wave built Compass, an AI onboarding assistant on Amazon Bedrock AgentCore that validates bank APIs against the Financial Data Exchange (FDX) standard, maps fields, and scores readiness for production connectivity. Ninth Wave provides secure data connectivity between financial institutions and third-party applications, connecting to aggregators such as Plaid, Finicity and MX and accounting systems including Intuit QuickBooks, Xero and Sage, normalizing bank APIs to FDX so a bank integrates once and reaches the whole open finance network. AgentCore supplies the managed agent runtime, while the Strands Agents framework handles orchestration logic.

Limitations: This is a customer story on AWS's own blog. The post says the same validation, mapping and scoring work has traditionally required weeks of specialist effort across email threads and spreadsheets, but gives no measured before-and-after figures for Compass and no third-party verification.

Source: AWS Machine Learning Blog · Strands Agents

One internal framework absorbs scattered agent services and cuts new-agent deployment from two weeks to one hour.

InfoQ reports that Grab has implemented LLM-Kit, a framework that standardizes over 500 internal agent services. The system unifies service integration, evaluation and secret handling, and reduces the time needed to deploy a new agent from two weeks to one hour.

Limitations: LLM-Kit serves Grab's own internal agent services, and the two-weeks-to-one-hour figure is Grab's own account. This item rests on a single InfoQ report with no independent reproduction of those numbers.

Source: InfoQ

Regional and early signals

A single release covers realtime speech, ASR, TTS, sound generation and music, with several first-place claims.

On September 15th, StepFun (阶跃) released the StepAudio 3 family: Realtime, ASR, TTS, Gen and Music, covering realtime voice interaction, speech recognition, human-grade speech generation and music creation. Per the third-party Artificial Analysis rankings cited by StepFun:

Realtime supports native full-duplex conversation, deciding when to answer and when to wait and handling interruptions; reasoning and speech generation run in parallel, while tool calls and long tasks execute asynchronously without breaking the current conversation. TTS uses a streaming architecture so generation and playback overlap, and produces paralinguistic cues such as laughter, hesitation, repetition and self-correction. Gen merges voice, sound effects, ambience and background music into one generation pass.

Limitations: The ranking claims come from StepFun and are relayed by a single outlet with no independent reproduction. Open weights, license, pricing and availability dates are not mentioned. (Chinese-language source)

Source: Leiphone

An embodied-AI inference engine whose publisher reports Pi 0.5 end-to-end latency under 26 ms on Jetson Thor.

Infinigence AI (无问芯穹), Tsinghua University and Shanghai Jiao Tong University open-sourced APXInf, an on-device inference engine for embodied AI. It supports mainstream compute platforms including RTX 4090, Jetson Orin and Jetson Thor, and covers the path from model development through validation to deployment on the robot. The stated optimization approach cuts end-to-end inference latency across pipeline, graph, kernel and quantization layers, generates fused kernels with a technique called Agent4Kernel, and builds a customized runtime per model to strip redundancy.

The team reports that on Jetson Thor, FP8 end-to-end inference latency for Pi 0.5 drops from 278 ms to under 26 ms, a 10.7x reduction, with frequency reaching 38.46 Hz.

Limitations: The "SOTA on Pi 0.5" claim and those latency figures are the publisher's own numbers, carried by a single outlet with no third-party reproduction, and the open-source license is not specified. (Chinese-language source)

Source: QbitAI

An agent with its own organizational identity joins a group and works within the permissions granted to that group.

Feishu and Doubao Work released a team agent that holds an independent organizational identity, can be added to group chats, takes tasks from different members, and uses documents, meetings and tools within its granted scope. In hands-on use it tracks tasks, handles Feishu documents, runs data analysis, sets scheduled monitoring and reminders and follows up with owners. Permissions apply per group chat, resource and application authorization, and Feishu's existing permission rules still hold. It can also be asked to look up an earlier discussion in another group and bring the conclusion back, and it records stated collaboration preferences into memory. The report compares it to Claude Tag, the shared agent that lives in Slack.

Limitations: The product is still in an early enterprise co-creation phase and is not open to all users. The account is a hands-on piece from a single Chinese outlet, with no numbers on participating companies, general availability or pricing. (Chinese-language source)

Source: ifanr