New open dev tooling for LLM workflows

Large language model tooling is getting a new layer of open-source plumbing: wrappers, gateways, memory stores, and debugging apps are appearing around Anthropic’s Claude Code. (github.com) Claude Code itself is Anthropic’s terminal-based coding agent, and Anthropic now pitches it across the terminal, desktop, integrated development environments, Slack, and the web. Its product page says desktop users can manage “multiple parallel tasks,” a cue that outside developers are now extending with their own orchestration layers. (claude.com) One of the clearest examples is Claude Squad, an open-source terminal app that manages multiple Claude Code, Codex, Gemini, and Aider sessions in separate workspaces. Its public GitHub repository showed 18,000-plus stars for Anthropic’s own Claude Code and described Claude Squad as a way to run several agents at once. (github.com 1) (github.com 2) Another strand is traffic control. A February 18, 2026 walkthrough from Sebastian Maniak showed how developers can route both Model Context Protocol tool calls and Claude Code model requests through AgentGateway instead of sending them straight to Anthropic. (maniak.io) That setup adds concrete controls that individual command-line tools usually lack: JSON Web Token authentication, rate limiting, OpenTelemetry traces, centralized secrets, and audit logs. AgentGateway’s Anthropic documentation says it can match incoming model names, inject the Anthropic API key, and forward requests to the provider’s messages endpoint. (maniak.io) (agentgateway.dev) A third layer is memory. Pgvector, an extension for PostgreSQL, lets developers store embeddings — numerical fingerprints of text — inside a standard database and run nearest-neighbor search to pull back similar records. (docs.spring.io) That matters for agent workflows because it gives teams a way to bolt recall onto ordinary application databases instead of standing up a separate vector store. Spring AI’s reference shows pgvector tables, HNSW indexes, and similarity search built directly on PostgreSQL. (docs.spring.io) The debugging side is getting its own user interface. Streamlit, the Python framework widely used for lightweight data apps, documents a path from local scripts to shareable web dashboards, which makes it a natural shell for prompt inspectors, trace viewers, and agent debugging tools. (docs.streamlit.io) What is emerging is less a single product launch than a stack around the stack: Anthropic provides the coding agent, and the community is adding session managers, security gateways, recall layers, and browser-based diagnostics. The result is that working with large language model agents is starting to look more like ordinary software infrastructure, with logs, routing, storage, and developer tooling wrapped around the model call. (github.com) (agentgateway.dev)