Key numbers from today's stories, April 7, 2026

Block Released Goose: An Open-Source AI Agent That Works With Any LLM

Block, the company Jack Dorsey runs, open-sourced an AI agent called Goose. It reads files, writes code, runs tests, and installs dependencies without manual intervention. [1] It is not locked to a single model. Goose works with Claude, GPT, Gemini, and local models through Ollama. [2]

The extension system is the interesting part. Goose supports over 70 extensions through the Model Context Protocol (MCP). That means it can connect to databases, APIs, file systems, and developer tools without custom integration code. It ships as a desktop app, a CLI, and an API. [3]

My Take

The model-agnostic approach is the right call. Most AI agent tools lock you into one provider's API. Goose lets you swap models without rewriting your workflow. The MCP extension system is where the real leverage sits. If your team already uses MCP servers for other tools, Goose plugs into that ecosystem immediately. The 70+ extension count is high enough to cover most development workflows out of the box. Worth evaluating if you are building agentic pipelines and want to avoid vendor lock-in on the model layer.

Claude Code Shipped Four Releases in One Week (v2.1.89 Through v2.1.92)

Anthropic pushed four Claude Code releases between v2.1.89 and v2.1.92 in a single week. The headline features span usability and performance. [4]

The /powerup command launches an interactive tutorial that walks new users through Claude Code's capabilities. The MCP tool result character limit jumped to 500,000 characters, up from the previous cap. The /cost command now shows per-model breakdowns so you can see exactly which model is burning your credits. File-diff operations on large files run 60% faster. [5] [6]

My Take

Four releases in one week signals that Anthropic is in a ship-fast cycle on Claude Code. The 500K character MCP limit is the most practically significant change. That ceiling was a real bottleneck for anyone running MCP tools that return large payloads, like database queries or file reads on big codebases. The /cost per-model breakdown is overdue. If you run mixed-model workflows (Opus for planning, Sonnet for execution), you need visibility into where the spend actually lands. The 60% file-diff speedup matters most on monorepos where large-file edits were noticeably slow.

RAGFlow v0.24.0 Added Persistent Memory to Open-Source RAG Agents

RAGFlow, the open-source retrieval-augmented generation engine from InfiniFlow, released v0.24.0 with memory management for agents. [7] Agents can now maintain context across multi-turn conversations and pick up where they left off between sessions. The release also includes batch metadata management and a chat-style agent interface. [8]

The memory problem is one of the most common complaints about RAG-based agents. They retrieve relevant documents but forget what happened three messages ago. RAGFlow's approach stores conversation state alongside the retrieval index so the agent has both document knowledge and session history. [9]

My Take

Memory is the gap that separates a RAG demo from a RAG product. Most open-source RAG engines handle retrieval well but treat every conversation as stateless. RAGFlow adding persistent memory at the engine level means you do not have to bolt on a separate memory layer yourself. If you are building internal tools where users return to the same agent across days or weeks, this is the feature that makes the difference between "search engine with a chat box" and "assistant that actually knows your project." Free and open-source lowers the barrier to trying it.

Open-source agents are getting model-agnostic. Developer tools are shipping weekly. RAG engines are learning to remember. The building blocks keep getting better.

Sources

${sourcesHtml}

Originally published on chento.io