Paperclip

🚀 Releases

Paperclip AI released `v2026.427.0` yesterday, a significant update introducing multi-user access with authenticated control planes, company memberships, and invite flows (#3784). This version also enables agents to post structured proposals as interactive cards within issue threads, facilitating explicit plan approvals and decisions (#4244, #4381). Enhancements to run liveness continuations and runtime lifecycle recovery mean long or interrupted agent work can now resume with full context, and a new watchdog system surfaces stalled runs for operator intervention (#4083, #4419, #4209). Additionally, Paperclip AI pushed several canary updates, including `v2026.427.1-canary.2` from cryppadotta, which hardens recovery issue handling, redacts sensitive retry failure details, and defaults new-hire approvals to opt-in (#4600). Cryppadotta also released `v2026.427.1-canary.1`, which prevents stale company skill refreshes from continuing against deleted or missing companies (#4601), and `v2026.427.1-canary.0`, ensuring that stored company selections in the UI are validated against loaded companies to prevent transient invalid states (#4602). Finally, cryppadotta updated Paperclip's READMEs to include an official X (formerly Twitter) link for the `@papercliping` account (PAP-2475).

Why it matters: These releases significantly enhance Paperclip AI's robustness, multi-user collaboration capabilities, and agent interaction model, directly impacting how engineering teams can deploy and manage zero-human companies.

🧠 Builder Signals

Thariq from Claude Code, Anthropic, clarified yesterday that the Claude Opus 4.5 model has been available in Pro plans since January, addressing confusion stemming from outdated support articles. He responded to users @synthwavedd and @scaling01, pointing to a Wayback Machine link to demonstrate the previous state of the documentation and apologized for the oversight. The previous articles did not reflect the inclusion of Opus in Pro plans, leading to user questions about its availability.

Why it matters: This clarification ensures that users of Claude Code and Opus 4.5 in Pro plans are aware of their full access, preventing frustration and enabling them to utilize the latest model capabilities.

💡 Ideas & Patterns

Paperclip Creator dotta yesterday highlighted the official Paperclip X account, urging followers to engage with the `@papercliping` handle due to its low follower count, signaling an intent to build community presence. Dotta also characterized Paperclip as "the ai agent orchestrator for humans to do 100x more work." Meanwhile, The Startup Ideas Podcast (SIP) outlined a 4-project starter pack for "Codex" to help users grasp its potential, suggesting projects like building games for AI play, conducting deep research into spreadsheets and presentations from a single prompt chain, developing mobile apps from existing website code, and automating daily annoying tasks. Steph Ango, CEO of Obsidian, engaged in a discussion about domains, asking if any app or tool used a ".md" domain (referencing Markdown) before Obsidian, and also clarified Obsidian's use of alternative domains and end-to-end encrypted Sync data.

Why it matters: These signals offer insight into Paperclip's community building and strategic positioning, provide actionable patterns for leveraging advanced AI agents like Codex in various engineering workflows, and clarify Obsidian's domain strategy and security features important for data privacy.

🤖 Model Drops

Nous Research co-founder Teknium announced yesterday that Hermes will now, by default, utilize native vision capabilities if the main agent model supports it, eliminating the need for a separate auxiliary vision model unless explicitly configured otherwise. This functionality, enabled by a pull request from teknium1, automatically attaches user-sent images as OpenAI-style content parts for vision-capable models, allowing them to process real pixels instead of a text description. Users can update their Hermes instance with `hermes update` to apply the change immediately. Separately, Teknium also shared a peer-reviewed CMU study (ICSE 2026) titled "Inside GitHub's Fake Star Economy." The study found millions of fake stars across thousands of repositories, with AI/LLM repos being a significant non-malicious category. It highlighted that stars are sold for $0.03-$0.85 each and are explicitly used by VCs as a sourcing signal, with the median star count at seed rounds being 2,850.

Why it matters: The native vision integration in Hermes streamlines multimodal AI applications, potentially improving performance and simplifying configurations for engineering teams. The expose on GitHub's fake star economy provides critical context for evaluating project traction and open-source contributions, informing decisions about dependencies and investment.

0xJeff echoed the concerns about fake GitHub stars yesterday, reinforcing that VCs use stars as a primary signal for sourcing and investment. He suggested that a fork-to-star ratio below 0.05 on a repository with over 10,000 stars indicates gamed traction.