Try
Useful tools & techniquesTheo says agents built a Rust TypeScript compiler
Theo announced tsc-rs, an open-source, drop-in replacement for the TypeScript compiler, type checker and language server, written in Rust. He says agents worked on it for five months; he spent about $400,000 on Codex without success, then about $20,000 on Opus to finish it in two weeks. He says he has not read the code.
Treat this as an anecdote, not a quality benchmark; inspect the code and run compatibility tests before trying it.
Claude SDKs add browser and computer-use loops
Anthropic says computer-use and browser-use toolsets are now built into its Python and TypeScript SDKs. The API exposes the clicks or typing Claude wants to perform, while the SDK handles the interaction loop that developers previously had to write themselves. The captured announcement does not give further implementation details.
If you are building a Claude browser agent, check the SDK toolsets before maintaining a custom interaction loop.
EmbeddingGemma 2 brings multimodal search on-device
Google released EmbeddingGemma 2, a 740-million-parameter model under Apache 2.0 that embeds text, code, images, video and audio into one space. Its 8K context supports up to 5.5 minutes of audio, 29 images or 58 video frames. Developers can get weights on Hugging Face and Kaggle; Google reports a 9.92-point gain over version 1 on MTEB Code.
Try it for local code search or multimodal RAG; truncated vectors can reduce vector storage by up to six times.
TikTok makes its Business MCP a connector
TikTok says its Business MCP is now available as a connector through Claude, Kimi, Manus, Perplexity, Replit, Snowflake, Tencent WorkBuddy and the IAB Tech Lab registry, letting agents manage ad workflows. The company also announced Shopping Assistant, Buy Direct checkout and Agentic Leads; a U.S. ad network spanning nearly 400,000 apps; and automated search-ad and creative-selection features.
If you build advertising agents, check the MCP connector; the other announced features are aimed at advertisers and sellers.
Know
What changedClaude plans include monthly API credits
Anthropic is rolling out monthly Claude Platform API credits: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team. The post says credits work with any model, including Haiku 5.5, in code or third-party harnesses; it points to separate terms but does not include them.
Check the full terms before shifting API workloads; the announced credit amounts may change your plan’s effective cost.
Haiku 5.5 cuts claimed model costs
Anthropic calls Haiku 5.5 its fastest, most capable small model and says it costs about 75% less to run than Haiku 4.5. It also claims gains in coding, computer use and knowledge work. A separate Claude Code post suggests using Haiku subagents for swarmable tasks, Sonnet for the main session and Opus as an advisor; the captured posts give no pricing or availability details.
Compare Haiku on routine subtasks before assigning expensive-model context to parallel work.
Liquid AI releases two open d1 models
Liquid AI released open-weight d1-3B for text and images, plus experimental d1-omni-600M for text with image or audio input. Liquid says they target real-time decisions from data centers to edge hardware such as RTX workstations and Jetson devices. The captured posts do not specify a license or benchmark results.
Consider these for local multimodal classification or routing, but validate performance on your target device.