What changed in AI, and what it means for your work.

Poki is an open-source (AGPL-3.0) tool that lets Claude Code, Codex or other MCP clients operate an iPhone paired to a Mac, with no jailbreak and no extra hardware. It needs macOS 15 or later and full Xcode; community ports for Linux and Windows exist but are not tested by the author. Its own agent defaults to GPT-6.1 Terra and can take tasks over iMessage or Telegram. The author warns that it will send messages or make payments if asked, that screen contents go to your model provider, and that physical iPhone support is verified only on phones using the Chinese interface.

Why it matters: Handy for testing iPhone apps or repetitive phone-only tasks, but early. Try it on a spare device with no payment or banking apps signed in, because the agent acts on the real phone.

Stroq is an open-source local tool for Claude Code, Codex, Cursor and similar agents. It reads the last 30 days of sessions on your machine, flags commands that came from text the agent read, and checks whether known keys left the machine (names only, never values). There is no account and no telemetry. `stroq init` installs hooks that can deny curl-to-shell and outbound calls carrying known secrets, and ask before a command repeats suspect text. `stroq attack` replays 36 known attacks in a sandbox; on 0.30.0 the project says 0 of 36 passed through. HN (90 points, 34 comments): No consensus. One camp says auditing is too late and you need a sandbox or isolated user instead. Another says audit logs are needed to keep access policies honest over time. The author clarifies that init hooks prompt before risky commands, and that Stroq is a filter, not a sandbox. Sorted take: run the secret scan and hooks, but still isolate the agent; do not treat rules alone as a containment boundary.

Why it matters: If your coding agent can shell out or browse, check whether keys already left past sessions, then add a pre-exec guard. Pair it with a sandbox or separate user so a missed rule cannot wipe the machine.

GitHub's guide says writing code got faster while the work around it did not. It cites more than 1.2 billion commits in one October week, over 400% more than a year earlier, and a Faros AI study where pull requests spent 441% longer in review as teams moved from low to high AI adoption. The post walks through Plan, Build, Verify and Observe with Copilot: gather context into plans and issues, run tasks in cloud sandboxes with session continuity, use stacked PRs and Copilot review, then investigate production with observability connectors. It stresses that judgment still sits with the developer.

Why it matters: If AI raised your commit count but slowed reviews, shrink PRs, automate routine checks, and put agents on triage and incident drafts before you buy another coding seat.

A GitHub issue reports Claude Code putting the user's email into outbound User-Agent headers during web fetches, which the reporter links to a surge of scam emails. An Anthropic Claude Code engineer said models sometimes add a contact email because sites like Wikipedia ask scripted clients to identify themselves, and that a built-in cc-plugin-privacy mod in v2.1.317+ blocks the behaviour by default. HN (357 points, 116 comments): No consensus. One camp treats it as a serious PII leak and wants the email removed from the system prompt, not papered over with a plugin. Another camp focuses on harness design, arguing models will always improvise unless deterministic checks stop them. Sorted take: treat agent network access as a data-exfil surface; pin Claude Code versions and keep privacy plugins on.

Why it matters: If your coding agent can browse, your login email may already have left the machine. Audit versions, disable auto web tools on shared accounts, and use masked emails for agent accounts.

Rivet (YC W23) launched an open-source actor orchestrator for agent workloads, aimed as a self-hostable alternative to heavy sandboxes and Cloudflare Durable Objects. Its README cites 4.99 ms p99 cold starts, about 72 KB RAM per actor versus 512 MiB for a typical sandbox, and S3-backed SQLite with roughly 1.07 ms p99 durable writes. Amp reportedly migrated from Durable Objects to Rivet and deploys on-prem. HN Launch (71 points, 38 comments): No consensus. Some self-hosters praise the Durable Objects alternative; others say the product pitch is unclear and question whether orchestration latency is the real agent bottleneck. Sorted take: worth a look if you are drowning in idle sandboxes, but validate cold-start and isolation claims on your own workload first.

Why it matters: Agent infra cost is often idle sandboxes and storage. A lighter actor model can cut spend if your agents need state more than full VMs.

Agent Autoloop is an open-source tool that runs command-line agents such as Claude Code or Codex on a schedule or on events like a new GitHub issue, using rules you write in a JSON file. Each agent finishes a round and exits; when it reaches something like deleting data, restarting a service, spending money or sending messages out, it sends you a card on Feishu, DingTalk, WeCom, Slack, Discord or a desktop notification, then resumes the same session after you answer. It runs on one Mac or Linux machine with Node 24, with no cloud server needed. The README is mainly in Chinese with an English version linked, and Windows support is not out yet.

Why it matters: If you already use a coding agent for routine jobs like morning reports or issue triage, this pattern lets it work while you're away but still stop for your approval on anything risky. Try it on a non-critical folder first.

A GitHub issue opened on 25 October reports that opening the Postman desktop app on a Mac automatically installed Postman plugins into Codex and Claude Code, a remote Postman MCP connection, the Postman CLI, and a startup hook telling the coding agent to load Postman's API skill before any non-trivial API task. Postman's own log recorded it as an automatic upgrade for an existing user, and the reporter found installer code that tries to approve its own Codex hook. A second user saw the same on version 12.35.1 on 3 November; neither report claims evidence that data was uploaded. A reply in the thread speaking for Postman said the tools are bundled to make setup easier, are mentioned on the download page and in the update flow, can be uninstalled inside the app, and that Postman will make opt-in clearer for existing users and review the hook behaviour. On Hacker News (67 top-level comments when we read it) none defended the auto-install, Bruno was the most named replacement (8 comments), and one commenter in healthcare said their security team had decertified Postman. Sorted take: a desktop app can now rewrite your coding agent's instructions, so review agent configs the way you review browser extensions.

Why it matters: If your developers use Postman, check ~/.codex/config.toml, ~/.claude/settings.json and Cursor's plugins for Postman entries and remove anything nobody approved. Add agent config files to your security checks, since other apps can change them.

DeepSeek has published dsh-superagent on GitHub under the MIT licence: a coding agent for a single GitHub repository, packaged as a bundle for DeepSeek Harness. It listens to GitHub and, optionally, Feishu; when someone with permission asks it for work, it opens a work item, does the job in its own session and directory, and replies in the original conversation, for example with a pull request. A web panel shows what needs a person, and the agent is built to act only when asked or when a rule the team has written allows it. DeepSeek labels it a research prototype, not an official product, says it may still contain bugs, and notes it differs from its internal version.

Why it matters: If you are planning an agent that works for a whole team, its documents on authorisation and on staying quiet unless someone must act are a useful template. Treat the code as reference only, as DeepSeek advises, and review it before pointing it at a real repository.

Mika is a new open-source Android app (Android 11 and up, Apache-2.0 licence) that takes a one-sentence request and operates your phone's apps and the web on a hidden background screen while you keep using the phone. Its demos include summarising a group chat, finding a flight and adding a product to a cart while stopping before checkout. You connect your own model API key or sign in with a ChatGPT, xAI or OpenRouter account, the README says there is no Mika server, and controlling phone apps needs a one-time wireless debugging pairing. It says it hands logins, verification codes and payments back to you. The repository was created on 29 October and has around 3,800 GitHub stars; it has not been independently tested.

Why it matters: If you repeat the same phone chores, such as checking prices or catching up on a busy group chat, this is a free way to try a phone agent. Test it on a spare device first and keep it away from banking and work accounts until you trust it.

Grim Prompter is a new open-source skill that has Claude Code, Codex or another coding agent review your codebase and write a threat model. It includes 36 architecture workflows and 29 method prompts, with a quick mode, a deeper six-stage review, and a mode that checks how a branch changes an earlier model. Its runbook asks for read-only inspection and file-and-line citations for each finding, and says builds, tests, installs and network access need your explicit permission. It is new and not independently tested.

Why it matters: If your small team has no security engineer, this gives a first-pass threat model for an app or API; run it with read-only permissions and treat the findings as a checklist to verify, not a sign-off.

GitHub has made Copilot Fusion generally available on all Copilot plans in VS Code, the GitHub Copilot app, GitHub Copilot CLI and JetBrains IDEs. Instead of choosing a model, you pick an effort level (Standard, High or Max) and Fusion routes each task: to one model, to a cheaper model whose draft escalates to a stronger one if a quality check fails, or to a draft that a second model reviews before revision. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, at about half the average API cost per task; these are GitHub's own figures, not an independent test. You are billed for each model it uses at standard rates, with a 30% discount on model costs for paid plans for a limited time.

Why it matters: If your developers spend premium model credits on routine edits, try Fusion Standard as the default for a week and compare spend. Admins can limit which models it may use.

GitHub's 2026 Octoverse report says that by September more than 40% of pull requests on GitHub started with a detectable agent's first draft, up from 17% in March, about 110 million in September alone. GitHub also reports that 55% of Copilot's monthly active users now use agentic features, up from 35% a year ago, and that 61% of Copilot users who chose models themselves used more than one. India has 32 million developers on GitHub, second only to the United States, and added more than 10 million this year.

Why it matters: Handing whole coding tasks to agents is now routine, which makes review the bottleneck. If your team ships software, agree a review rule for agent pull requests before the volume grows.

A gist from Flamerix AI says that about every 30 seconds, Claude Code sends each background agent's full transcript back to that agent's model just to write a 3 to 5 word status line in the agents panel. In their own telemetry over one week, these agent_summary calls were 16.4% of Claude Code spend and 19.7% of API requests. They say it has been on by default since version 2.1.232 in August 2026, has no setting that turns off only this, and they share an unofficial workaround script that an update can silently break. On Hacker News (224 comments), a commenter who says they work on Claude Code replied that only actively working background agents get these calls, that across users who send telemetry they were about 2% of cost and just under 4% of requests last week (about 6% of cost in the terminal), and that a fix for the most expensive case shipped in 2.1.311+ with a cheaper mode in progress. Of 49 top-level comments, 3 suggest prompt caching keeps the cost low, while at least 15 blame Claude Code's bloat or the incentives of a model maker building its own tool. No consensus on how big the cost is for a typical user.

Why it matters: If your team runs many background agents in Claude Code on API billing, check your OpenTelemetry export for agent_summary requests before assuming the bill is all real work. Update to the latest version rather than installing unofficial patches on work machines. Sorted take: the share depends heavily on how you work, so measure your own usage instead of trusting either figure.

GitHub says Copilot now supports server-to-server authentication for GitHub Apps installed on organisations, in public preview. Services and agents built with the Copilot SDK can call Copilot on behalf of an organisation using a short-lived installation access token, without a signed-in user's credentials. Developers give the app the read-only Copilot Requests organisation permission; usage is billed to the installing organisation and follows that org's model policies, budgets and spending controls. Owners can spot app-attributed usage via the app_client_id field in AI usage CSV reports and stop access by suspending or uninstalling the app.

Why it matters: If you are wiring Copilot into internal bots or CI agents, move those calls to an org-installed GitHub App so spend and model policy sit with the company, not a personal account.

GitHub says GPT-6.1 Terra is now generally available in GitHub Copilot for Pro, Pro+, Max, Business and Enterprise users, aimed at everyday agentic coding and terminal work at a lower cost than GPT-6.1 Sol. In early testing, GitHub says it solved noticeably more terminal and C# coding tasks than GPT-5.6 Terra. You can pick it in the model picker across Visual Studio Code, Visual Studio, Copilot CLI, cloud agent, the Copilot app, github.com, GitHub Mobile, JetBrains, Xcode and Eclipse; rollout is gradual. It is billed at provider list pricing under usage-based billing, and Business and Enterprise admins can manage access in Copilot model policy settings.

Why it matters: If you already pay for Copilot usage-based models, try Terra on one well-scoped coding or terminal task and compare cost and quality with Sol before making it the default.

Hob is a new MIT-licensed desktop app from Frigade for Mac, Windows and Linux that tries to do what the Claude and ChatGPT desktop apps do with your files, but runs the model on your own computer. It checks your hardware and picks a model for you, Google's Gemma 4 on most computers and OpenAI's gpt-oss 120B on workstations, then works on folders, spreadsheets, PDFs and photos, with charts, sandboxed Python and approval before it changes anything in your folders. The project says there is no telemetry and that it only goes online to download models and, while Web access is on, to search, fetch data and install Python packages; you can turn that off. It is early and not independently tested. On Hacker News (24 comments), the lean was friendly. One commenter said there are many apps like this but upvotes every one because local AI matters more than what big companies are doing, and others suggested alternatives such as Open WebUI, Osaurus and Orca. No commenter reported testing it; one said they had not tried it yet, and requests included storing models on a separate drive and pointing it at a model on the local network. Sorted take: worth a try on a spare laptop for private documents, but expect local models to be weaker than frontier cloud ones on hard tasks.

Why it matters: If you handle client files you can't send to a cloud AI, try Hob on one folder of non-critical documents and see if a local model is good enough for summaries and spreadsheets.

PaperPilot is an open-source PDF toolkit written in Rust that runs as an MCP server for Claude Desktop, Cursor and other MCP clients, plus a CLI, a browser WASM build, a desktop app and a local REST API on localhost:7823. It merges, splits, rotates, compresses and OCRs PDFs on your machine; the project says files never leave the device, and the browser demo at usepaperpilot.com processes files in memory with no upload. Current v0.6 also covers watermarks, signature checks, PII scan, deep sanitize and Excel to JSON or CSV. It is pure Rust with no Chromium or C++ runtime.

Why it matters: If Claude or Cursor already handles your docs workflow, add the paperpilot-mcp binary so the agent can merge and compress PDFs without sending them to a cloud service.

Lampline is a new open-source project (Apache 2.0) for building AI phone agents on your own machine or servers. It can import agents from Google's Dialogflow CX and CX Agent Studio, and connects to phone lines through Twilio, Telnyx or a SIP trunk. Its Evals tab rings any agent with a phone number, including ones on Vapi, Retell or an old IVR, plays the customer with an AI Test Caller, and grades the call against checks you write. A free mode runs everything locally (Qwen3 4B in Ollama plus local speech models) after a download of about 3 GB; the README claims about $0.02 to $0.17 a call with your own keys, against $0.50. It is a young project, first published on 19 October, and not independently tested.

Why it matters: If you run a phone line with a bot or IVR, the AI Test Caller is the part to try first: script a few hard calls (noise, accents, interruptions) and see where your current agent breaks. Only test agents you own, as the project itself says.

OpenShell is an open-source tool from NVIDIA (Apache 2.0) that runs each AI agent in its own sandbox. You write a policy saying which files, system calls and network destinations each agent may use, and OpenShell enforces it at the kernel level. Agents never see your real credentials: OpenShell adds them only to requests going to approved endpoints. Before a policy change is applied, it uses formal verification to flag risky new access, such as reaching a new host with credentials, so a person can review it. It runs on Linux, Apple Silicon Macs, or Windows with WSL 2 (experimental), using Docker, Podman or host virtualisation. It sends anonymous telemetry by default, and the README explains how to turn that off. On Hacker News (284 points, 142 comments from 29 top-level commenters when we read it), one camp liked it: one person said a large share of the demos at GTC Europe used it, and one said their team has been using it for a while. A second camp objected to telemetry being on by default. A third asked why not just use containers or plain Docker, and some doubted any tool can give agents wide access and still keep them safe. No consensus. Sorted take: the useful idea here is credentials the agent never sees. Whatever sandbox you choose, look for that.

Why it matters: If you let coding or ops agents run commands with company keys, a sandbox with per-agent rules limits the damage from a bad step or a prompt injection. It suits teams with someone comfortable with Linux and Docker.

OpenAI has an official command-line tool, openai, for calling its REST API without writing an app. From a terminal or a shell script you can generate and save images, upload and download files, submit batch jobs and wait for the results, and export data as JSON for other scripts. With a separate admin key it can also pull your organisation's usage. It is open source (Apache 2.0) and installs through Homebrew, Go or prebuilt binaries for macOS, Linux and Windows. The README says to use Codex CLI instead when you want an agent to work on code; this tool is for direct API calls. On Hacker News (229 points, 109 comments from 16 top-level commenters when we read it), the biggest camp was confused: several asked how it differs from Codex, and two said the README never says plainly what it is for. The most-replied comment warned that OpenAI tends to deprecate its tools within a couple of years, and two others said a tool like this is trivial to build yourself. A smaller camp welcomed it, with one commenter explaining that Codex is an agent while this is a thin layer over the API. No consensus, though the lean was sceptical. Sorted take: handy for scripted batch jobs and usage checks. Keep those scripts simple in case the tool changes.

Why it matters: If you run batch jobs, make images in bulk or want a quick usage check, a few lines of shell script can now do it without custom code. You need an API key to use it.

stalemap is a small open-source tool (MIT licence) for JavaScript and TypeScript projects. It checks which of your dependencies changed after your AI model's training cutoff and writes a short list of those changes into AGENTS.md or CLAUDE.md, so the agent stops writing code for old APIs. It can also run your type-check after each edit and show the agent the matching change note, and it offers an MCP search tool and a Claude Code plugin. In the author's own blind, pre-registered test of 66 tasks, claude-sonnet-5-5 agents passed 89% of sessions with stalemap, against 68% with no help and 73% with a plain type-check after each edit. The project is early and maintained by one person, and the README says its setup for Cursor, Codex and OpenCode has not yet been tried in real installs.

Why it matters: If your coding agent keeps using old function names after a library upgrade, run npx stalemap on one project to see what it flags before adding it to your workflow.

DeepSeek has released DeepInfer on GitHub under the Apache 2.0 licence. It says this is the inference framework that powers its large-scale production services. This release covers offline inference only, supports one model (DeepSeek-V4.1-Flash) and runs on NVIDIA SM100 GPUs or Huawei Ascend 950 chips. Features include structured outputs (JSON schema, regex or grammar), deterministic sampling and batch-invariant results on NVIDIA. Parts you need for a full service are not included: a KV cache system, prefill and decode disaggregation, and load balancing across replicas. The README says it was originally built mainly on vLLM, and DeepSeek's speed charts were measured on 4 SM100 GPUs for prefill and 64 for decode. On Hacker News (2,108 points, 63 top-level comments when we read it) there is no consensus. One camp calls it a blow to US labs' pricing and valuations. Others push back: one commenter says more than 30 providers already serve DeepSeek with existing engines, and three point out that the KV cache and other serving pieces are missing. Three asked how it compares with vLLM or SGLang, and no top-level comment gave numbers. Huawei Ascend support drew attention, though one commenter said US rules keep those chips out of reach for American buyers. Another noted that a US host running DeepInfer would be running a DeepSeek product, which matters given the new federal removal order. Sorted take: for most teams this is not a drop-in tool. It needs top-end GPUs and you would still build the cache and routing yourself. The useful signal is price: cheaper serving code for DeepSeek models could mean lower prices at the hosts you already use.

Why it matters: If you pay for DeepSeek through a third-party host, watch for price or speed changes as hosts test this engine. If you run models in-house, treat it as reference code for now, since it needs SM100-class GPUs and is not a complete service.

Antirion is a new open-source (Apache 2.0) gateway that sits between your apps and model providers such as OpenAI, Anthropic, Amazon Bedrock, Google Vertex AI, Gemini, Mistral and DeepSeek. It ships as a single Rust binary with a web console built in. Features include fallback and cost-based routing, budgets and rate limits per team, project or key, caching, and guardrails for things like personal data and secrets. The project's own benchmark, on a 4 vCPU AWS machine at 5,000 requests a second, shows 0.53 ms median overhead against 19.7 ms for Bifrost 2.2.4, while LiteLLM 1.103.1 returned errors on every request. These are the author's figures, not independent tests, and the repo is only a few days old.

Why it matters: If several people or tools in your company call AI APIs, a gateway gives you one place to cap spend and see who used what. Try it with a test key next to your current setup before you rely on the speed claims.

Isuru Wijesiri, who built AI copilots at WSO2, has published Harness Engineering 101, a free online book on building AI agents from first principles. Its core idea: an AI model only takes in a list of messages and returns text, and the "harness" is the program that builds and protects that list, so tools, the agent loop, caching, subagents, guardrails and RAG are all fixes to that one loop. Over 16 chapters it grows a roughly 300-line toy agent in plain Python, with appendices on routing between models, cost, retries and rate limits. The code is MIT-licensed and the text is free to share but not to reuse commercially; the author says he wrote the chapters himself and used AI tools to reorder and polish them. On Hacker News (128 points, 64 comments), the 18 top-level comments lean critical. Four dismiss it as AI-written, one suggesting you build your own harness from older guides instead. Six argue with the brain and body comparison, offering other analogies or objecting to treating models as brains. Two are positive, one saying tuning the harness matters as much as the model. Three share their own harness projects, and three ask other questions, including where to find advice on hardening a harness. No consensus. Sorted take: The idea is sound and worth an afternoon if you manage people building agents, because most agent bugs are about what goes into the message list; skim the contents page and read the chapters on context, caching and guardrails first.

Why it matters: If your team builds internal agents or buys agent tools, the chapters on context limits, caching and guardrails explain why agents slow down, cost more or go off track, which helps you ask vendors and developers better questions.

Scrawl is a free, open-source app by developer Kit Langton for showing a coding agent a problem instead of typing it out. Press a shortcut, talk, click around and draw on the screen, press it again, and Scrawl packs the recording into a small bundle: a contact sheet of marked-up key frames, the full-size frames, a time-stamped transcript of what you said and did, and a data file. An add-on skill lets agents such as Claude Code, Codex and OpenCode read the bundle and zoom in on the moment you point to, and it can also export a shareable video. It needs an Apple silicon Mac on macOS 26 or later, with an unsigned Windows preview, and asks for screen recording, microphone and accessibility permissions. It is MIT-licensed and has about 465 stars on GitHub.

Why it matters: If you report bugs or design fixes to developers or agents, a 30-second narrated recording with marks often says more than a long message; try it on your next UI bug, and check what is on screen before you send it.

whatsapp-mcp, a free open-source project under the MIT licence, runs a small server in your own Cloudflare account that links to WhatsApp the way WhatsApp Web does, so Claude Code or any other MCP client can search and read your chats. It is read-only by design: it cannot send messages, stores photos and voice notes only as labels, and serves one person's account per setup. The developer warns that it relies on an unofficial library, Baileys, which breaks WhatsApp's terms, and calls the risk low but not zero. Importing a large chat history needs Cloudflare's Workers Paid plan, which the README puts at a few dollars a month. On Hacker News (115 points, 66 comments), 7 of the 19 top-level commenters were wary on privacy or trust grounds, two of them pointing out that the people you chat with never agreed to have their messages passed to an AI company. Four asked about or described WhatsApp bans for unofficial tools, four shared other setups or services, two called the project useful, and two raised other points, including that it is not really free. Rough consensus: a useful idea, with real risks to privacy and to your account. Sorted take: Fine for experimenting on a personal number, but don't connect a WhatsApp number your business depends on; for customer chats, use WhatsApp's official business tools.

Why it matters: If WhatsApp is where your work happens, this lets an agent find an old message or summarise a busy group, but a ban could cut that number off from customers; test it only on an account you can afford to lose.

Graphicle, a free open-source tool from KarLex AI under the Apache 2.0 licence, maps a code repository into a graph of files, functions and calls, and keeps a searchable memory of project docs, decisions and past Claude Code and Codex sessions. Coding agents such as Claude Code, Codex and Cursor query it over MCP and get excerpts that cite exact files and lines. It runs on your machine with no account and no telemetry, apart from a daily update check you can switch off, and memories an agent proposes wait for your approval. Its makers' own tests are candid: on 100 coding tasks it tied plain Claude Code at finding the files a fix needed, and they say they do not claim it saves tokens.

Why it matters: If your team uses coding agents on one codebase, the memory of past decisions and sessions is the part worth trying; don't expect faster or cheaper code search from it yet.

Harness2, from a Google Research team and posted to arXiv on 9 October, improves the setup around an AI model, such as its prompts, tools, memory, skills and sub-agents, without changing the model itself. For each new task it tries a few edits, runs them, compares the results and keeps what worked, then saves the lessons for future tasks, all without needing an answer key. In the authors' tests with GLM-5 and Gemini 3.5 Flash, one round improved scores by up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench and 11.8 percentage points on LAB. The code is open source under the Apache 2.0 licence, works with OpenCode and Codex, and is not an officially supported Google product.

Why it matters: If your agents underperform, the setup around the model can matter as much as the model; this is a way to tune prompts and tools automatically, though the gains are the authors' own results.

A bug report open on OpenAI's Codex GitHub since 2 June says the Codex desktop app on macOS, now part of the unified ChatGPT desktop app, keeps writing crash dump files in the background with no limit, while the app itself keeps working. Sixteen users in the thread report the problem, most recently on 17 October with a current build. Reported sizes range from 4.9 GB in a day to one user's 323 GB across two folders, about 35% of their disk. One user traced it to the crash handler crashing and dumping itself about once a second, roughly 15 GB a day. The issue is labelled as a bug, and no fix has been announced in the thread.

Why it matters: If you use ChatGPT or Codex on a Mac, check the Crashpad folders under ~/Library/Application Support/Codex and ~/Library/Application Support/com.openai.codex/web; deleting the files in their pending folders frees the space.

Perplexity has released SPACE, short for Sandboxed Platform for Agentic Code Execution, under the Apache 2.0 licence. It runs code from agents or users inside isolated Firecracker microVMs that an application creates, runs commands in, pauses and deletes through an HTTP API, so you do not need to write Rust to use it. It includes a compatibility layer for the E2B Python SDK, warm pools so sandboxes start without a cold boot, and a reference deployment on AWS. It needs a Linux x86_64 host with KVM and does not run directly on macOS.

Why it matters: If your team builds agents that write and run code, this is a free way to host your own isolated sandboxes instead of paying for a hosted service, provided someone can run Linux servers.

Solid Glass is an open-source design system from DeepLethe, made as a single stylesheet for shadcn/ui projects on React 19 and Tailwind CSS 4, so existing components pick up the look without code changes. Beyond the theme, it adds components made for apps where an AI agent works, such as a status line for what the agent is doing, a message box, the steps it took, a diff view, a question card, a context-usage ring and a reasoning panel. It comes in dark and light themes, is MIT licensed and passed 1,700 GitHub stars within a week of its 12 October release.

Why it matters: If your small team is building an internal AI assistant or agent dashboard on React, this saves design time on the parts every agent app needs, like showing progress and tool steps.

Neckbeard is an open-source Claude Code plugin that keeps agents from inventing infrastructure. Humans maintain a versioned catalog of OpenTofu modules for a fixed shape (container image, HTTP, worker and cron processes, Postgres, object storage). The agent profiles the app, asks what code cannot answer, then plans and scaffolds from that catalog, with Infracost estimates before anything is created. The README quotes about $77 a month for the smallest Cloud Run reference app across three environments, versus hundreds for Kubernetes, and it refuses to overwrite files you have edited. On Hacker News (75 points, 56 comments), there is no consensus. Four commenters questioned the need for it: two said they would rather write their own Terraform or OpenTofu, one said you can just ask the agent, and one called it a bunch of Terraform modules. The author replied that the point is opinionated modules the agent does not have on its own, and agreed it targets apps moving off Vercel or Supabase onto containers or serverless. Two said a cheap Hetzner server would cost less, and the author said Hetzner support is next. One commenter worried an LLM could make cloud spend skyrocket; the author replied that the tool does not let the LLM do that. One disliked the AI-sounding README. A side thread joked about other hair-named plugins, and a commenter who said they created ponytail, the plugin that inspired it, welcomed it. Sorted take: useful if you are moving a Vercel or Supabase app onto AWS, GCP or Azure and want a reviewable bill before apply. Skip it if you already run mature infrastructure as code.

Why it matters: If coding agents keep proposing open security groups and mystery monthly costs, a catalog that shows the bill first is a practical guardrail. Try it on a greenfield service before trusting it on production accounts.

open_scion is an MIT-licensed, community project that reimplements ideas from Anthropic's internal Scion source-control system for coding agents. Classic Git hosting assumes humans who clone, work for hours, and push once. Agents spawn thousands of short tasks that want to land in seconds. open_scion moves history, blame, and landing to a server with sparse fetches, compare-and-swap lands instead of a merge queue, a changeset log, and storage in SQLite or PostgreSQL plus object storage. It sits on Gitaly and normal Git repos, keeps the server model-agnostic, and ships agent tokens, review rules, and optional LLM conflict resolvers. Status is 0.1.0, first public release, with known gaps listed in its docs.

Why it matters: If you run many parallel coding agents on one repo, Git clone and merge-queue pain is becoming the bottleneck. This is an early open attempt to try Scion-style landing without waiting for Anthropic to ship Scion publicly.

Outcrawl released Scout under the Apache-2.0 licence. You describe a task in plain words, and it uses a real browser to click, type and scroll until the job is done, or tells you what is missing and returns what it found. It has no built-in model: you plug in any service that speaks the standard chat completions API, including a model on your own computer, and can use different models for taking steps, planning, reading screenshots and extracting answers. It runs locally with Playwright, or on Outcrawl's cloud, which adds captcha solving and run replays, with 1,000 free credits a month. It is a TypeScript package for Node 22 or later, and can return answers in a fixed JSON shape.

Why it matters: If you have repetitive web tasks, such as checking prices, filling forms or pulling data from portals with no API, Scout is a cheap way to try automating one with your own model. Start on a test account before giving it real logins.

MailCheap, from GitHub developer marclou, is an open-source newsletter app. You can self-host it for free with your own Amazon SES or PostShiba sending account, or use the hosted version from $19 a month. The pitch is cost: you pay per email sent, not per subscriber, so 10,000 readers and four issues a month is 40,000 emails, about $4 to $6 on SES. It includes an editor, signup forms, confirmed opt-in, unsubscribe and bounce handling, list cleanup, staged sends and stats. Every dashboard action also works through a REST API and an MCP server, so an AI agent can draft and send issues. The repo says plainly there is no support.

Why it matters: If your company newsletter is on a per-subscriber plan, compare the bill. Self-hosting takes about an hour, mostly waiting for DNS, and lets your AI agent handle drafts and scheduling.

new-term-radar is an open-source (MIT) Python tool, created on GitHub on 12 October 2026, that writes a daily report of AI names starting to surge: new models, products and repos. It watches first mentions by X accounts you choose, models new to Hugging Face's trending top 50, GitHub repos under 14 days old averaging 200+ stars a day, new words in sitemaps you list, and your own Search Console queries. It then checks each lead in Google Trends 14 days later to score its own hit rate. A keyword table pairs each new name with the generic searches behind it and marks it prioritize, watch or skip. The repo had about 880 stars and 357 forks two days after launch. The free signals (Hugging Face, GitHub, sitemaps, Search Console) need no key. X reads, Trends, the keyword table and the name judge go through a paid AIsa key; the author puts the example setup at about $15 a month. It runs from the command line, as three Agent Skills, or as an MCP server. The author is candid about limits: in early runs most generic terms did not rise within five days, and Google Trends' latest two days are unreliable.

Why it matters: If you run an AI tool directory, newsletter or SEO site, try the free GitHub and Hugging Face signals first to see whether it surfaces names earlier than your current workflow before paying for the X and Trends reads.

Janas is an experimental, GPL-3 C toolkit whose Janas-LLM engine runs mixture-of-experts models larger than system RAM by keeping weights on NVMe and streaming the experts each token needs into a memory cache. Author Maurizio Cammalleri reports that on a 32 GB Debian laptop (Intel Core Ultra 9 185H, integrated Arc GPU), Qwen3-Next-80B (48.4 GB file) wrote at 24 tok/s without drafts, versus 2.8 for llama.cpp on the same machine, and read a 1,024-token prompt at 62 tok/s versus 16. With drafts in chat he reports about 37 tok/s for that model. He says Janas-LLM was faster than llama.cpp on every model compared in Q4_K_M on that laptop across 13 models. Linux and Windows x86-64 binaries are on GitHub releases; the README stresses it is early, API-unstable, and hard on free memory. On Hacker News (205 points, ~77 comments), many welcome another open local-inference engine and note SSD expert streaming is becoming common. A camp says llama.cpp already covers much of this and specialised engines often reimplement it for speed. Others debate whether 60 tok/s is enough for agents (prefill vs decode) and whether SSD wear or High Bandwidth Flash is the next step. No broad consensus beyond "useful experiment, measure on your own machine". Sorted take: promising for laptop MoE if you accept experimental software. Run janas-bench on idle hardware before you trust the headline speeds.

Why it matters: If you want a large open MoE on a 32 GB laptop without a big GPU, this is another option besides llama.cpp. Keep it to a test machine until the API settles.

GitHub published on 12 October 2026 a Copilot app beginners guide on giving Copilot lasting project context instead of re-explaining house rules every task. Four layers: repository custom instructions in `.github/copilot-instructions.md`, path-specific files under `.github/instructions` with an applyTo field, agent skills as folders under `.github/skills` with a SKILL.md for repeatable procedures such as quality checks, and custom agents under `.github/agents` for broader roles such as accessibility review. MCP servers connect Copilot to outside tools; the post walks through Playwright so Copilot can open the app in a browser and click through flows after a change. The advice is to start small: put one convention you keep repeating into copilot-instructions.md, then add skills and agents only for workflows you already repeat. Awesome Copilot is pointed to for community-ready examples to adapt.

Why it matters: If your team uses Copilot on a real codebase, a short instructions file and one skill for your test/lint script usually cut more mistakes than chasing a smarter model.

REA, short for Reverse Engineer Anything, is an MIT-licensed tool, first published on GitHub in April 2026, that connects an AI coding agent, through MCP, to reverse-engineering tools for native binaries, JavaScript and Electron apps, .NET assemblies and websites. The idea: point your agent at an app whose feature you like, and it inspects the app without source code, explains how the feature works with evidence, and can help build a version for your own project. It also works from the terminal, and analysis runs locally. Its latest release, rea-agents v6.4.0, shipped on 11 October 2026. Native analysis can use an existing Hopper, Ghidra or IDA install. Treat the legal and licensing side carefully: copying a competitor's feature by reverse engineering can raise intellectual property questions depending on what you reproduce.

Why it matters: If you investigate how an app works, for security, debugging or compatibility, this packages the usual tools behind an agent. Check the licensing and terms of anything you reverse-engineer before you reuse it.

A new open-source project, esp32-iphone-agent, lets an AI model operate a real iPhone with no Mac, Apple developer account or jailbreak. The MIT-licensed project uses AirPlay screen mirroring as the AI's "eyes" and a roughly $7 ESP32-S3 board plugged into the phone's USB port as its "hands", appearing as a standard USB mouse and keyboard. Any OpenAI-compatible vision model can be the brain, including one you run yourself. A planner model writes a plan from the goal and screen, and an actor model picks each tap, swipe or keystroke; it checks whether the screen changed after each action and retries missed taps. The README demos it drawing a smiley face and setting a repeating alarm. This is an early hobbyist project. It needs a computer on the same network running UxPlay. The README lists its limitations: iOS does not mirror some screens, such as passcode entry, Apple Pay and DRM video, so the agent cannot see them; mirroring stops when the phone locks; keyboard input is ASCII only; and landscape apps are not supported yet.

Why it matters: If you test mobile apps or automate phone tasks, this is a cheap way to try AI control without Apple's developer setup. Keep it to test devices, since it needs broad control of the phone.

QGIS, the free and open-source mapping software, has approved a new policy: "No LLM, AI or agent assisted code contributions are allowed" in its main repository. Contributors later found to have used AI face "code reversion and contributor rights sanctioning". The proposal was merged on 7 October 2026 after a vote. It replaces an earlier policy that allowed AI tools if a human reviewed everything. A reviewer on the pull request noted that under the old policy about 50 pull requests were tagged AI Assisted and about 40 were closed without merging. A maintainer said the project had already lost an extremely high value contributor. QGIS core developer Nyall Dawson described the change as a closed door that may be opened selectively later. On Hacker News (363 points, about 390 comments), opinion split. Of the 60 top-level comments, many backed the ban, citing overloaded reviewers and code quality, while many others criticised it as short-sighted or predicted AI-rewritten forks. A small group, about 5 of the 60, asked how "AI assisted" would even be defined or detected, a worry QGIS contributors also raised on the proposal itself. No consensus. Sorted take: this is a reviewer-capacity problem more than an anti-AI stance. Expect more volunteer projects to restrict AI contributions until they can review them properly.

Why it matters: If your team sends fixes to open-source tools you depend on, check each project's AI policy before submitting AI-written code. Some now ban it outright and will revert it.

Nous Research has put the official plugins for its open-source Hermes Agent in one GitHub repo (MIT licence, created 10 October 2026). Each plugin installs with one command from the Hermes plugin catalog, and every catalog entry is pinned to a reviewed commit, so you get exactly the code that was reviewed. The official set includes Home Assistant, Spotify, Snyk security scanning, Blender, a Memory Wiki that turns your local session history into browsable pages, and Telegram Business Mode, which drafts replies to your customers and sends nothing until you approve each one. There is also an experimental plugin that runs Hermes on your Claude Pro or Max subscription through the official Claude Code CLI instead of a per-token API key. Hermes itself works from the terminal, a desktop app, Telegram, Slack, WhatsApp and other chat apps, with whichever model you choose. Sorted take: The approve-before-send pattern is the right default for any agent that talks to customers.

Why it matters: If you handle customer chats on Telegram Business, the approval-first mode is a low-risk way to try an AI assistant on real messages. The reviewed, pinned plugins also lower the risk of installing something that changed after review.

Talorys is a new open-source personal assistant that installs with one command (npx create-talorys@latest) into your own Cloudflare account. It gives you chat with memory you can view and edit, tasks, notes, projects, and reminders or routines that run on a schedule. It is single-user, has no telemetry, and the developer runs no server or database. It uses Cloudflare's free tier, including Workers AI with the GLM-4.7 Flash model; the README is clear that free quotas exist and can change, and that tasks and reminders keep working when the daily AI allowance runs out. On Hacker News (239 points, about 120 comments), the biggest camp argued over the word "self-hosted", since it runs on Cloudflare rather than your own machine; others said a cloud account you control is close enough. Several asked what Cloudflare's free "neurons" buy, with one estimate of about $0.11 of usage a day. Most of the thread argued over that label. The two commenters who said they tried it were positive: one found setup easy but used 500 of the 10,000 daily neurons in a few minutes and wanted web search; the other called it a cool idea but said it needs semantic memory recall, after it suggested chicken despite a saved note saying they were vegetarian. No consensus on the label. Sorted take: owning the account matters more than the label. Use it for light personal tasks and reminders, and don't rely on free quotas that can change for anything critical.

Why it matters: If you want a private reminder-and-notes assistant without a subscription, this is a cheap one to try. Expect the free AI quota to cover only light daily chat.

SolveCraft is a new open-source parametric 3D CAD app, written in Rust under the MIT or Apache 2.0 licence, at your option, that aims to be a clean-room reimplementation of Autodesk Fusion 360. You sketch, add constraints and dimensions, and build solids on a feature timeline; change a parameter and the model rebuilds. It runs on desktop or in the browser, exports to STEP, IGES, 3MF, STL and OBJ, and imports STEP and IGES. It includes an MCP server, so AI agents such as Claude Code can build, measure and screenshot parts, either headless or in the running app while you watch. The README's demo shows an agent building an electronics enclosure, then the width changed from 80 to 100 mm and the model rebuilding. The developer calls it "early but usable for simple parts" and says all 76 reference parts built in Fusion rebuild with the same volume, area and topology.

Why it matters: If you design enclosures, brackets or jigs, you can try it free in the browser and let an AI agent draft simple parts you then adjust. Keep complex production work in your current CAD tool for now.

Chrome Relay is a new open-source tool (MIT licence) for Macs that lets coding agents such as Claude Code, Codex and Cursor drive your everyday Chrome, with your real logins, in hidden background tabs. It works through a Chrome extension and a local relay, so nothing steals focus. Each agent sees only the tabs it opened, and idle agent tabs close after 30 minutes. Only connections with a secret stored on your Mac and coming from an allowed agent app are accepted, and web pages are refused. Agents never type passwords: if a site needs a password, passkey or two-step check, the agent stops and asks you. Every action is logged to a local audit file, with typed text stored only as a character count. The author says plainly it is not a defence against malware already running as you. Setup needs macOS, Chrome, Node.js 20+ and loading the extension by hand.

Why it matters: If your agents keep stalling at logins for internal tools, this lets them work inside your signed-in browser. Use it only on accounts you are happy for an agent to act in, and check the audit log.