CEREBRO

Coding agents

Claude Code v2.1.219: Opus 5 (claude-opus-5) is now default Opus, 1M context, fast mode at $10/$50 per Mtok, plus a sandbox.network.strictAllowlist setting to silently deny non-allowlisted hosts for sandboxed commands.
Auto mode default for Claude Code Pro/Max/Team: Anthropic now defaults new sessions to auto model selection, a strong confidence signal in auto-routing over manual model picking.
Claude Code v2.1.218: /code-review now runs as a background subagent so review work stops filling the main conversation.
bb: agentic IDE that orchestrates multiple coding agents (including itself) across desktop/web/CLI/HTTP, with steerable live threads.
herdr: always-on background runtime for coding agent terminals, sessions survive lid-close/network-drop/restart and reattach from any terminal or SSH.
oh-my-pi: Rust-core terminal coding agent (fork of Pi) with hash-anchored edits, 60+ providers, LSP/DAP tool integration, ~80k LOC.
prime-agent: open-source agent built on a "Recursive Language Model" abstraction, treating context as variables and subagents as function calls inside a persistent REPL.
orca: ADE for running Codex/Claude Code/OpenCode/Pi side-by-side in separate worktrees with unified usage/rate-limit tracking.

Agent Plugins Spec v1.0.0: vendor-neutral packaging standard for Agent Skills and MCP servers as distributable plugins, a step toward portable agent extension ecosystems.
claude-skills (alirezarezvani): 362 skills/plugins/agents spanning 13 coding tools (Claude Code, Codex, Gemini CLI, Cursor, etc.), largest open skills library seen so far.
agent-skills (addyosmani): production-grade skills packaged around a define→plan→build→verify→review→ship pipeline for consistent agent workflow quality gates.
swarm-forge: tmux-based multi-agent orchestration platform with a shared "constitution" and role-based branches, Uncle Bob's take on disciplined agent swarms.
archify: agent skill that turns a repo/system description into self-contained interactive architecture/sequence/flow diagrams across five diagram types.

EMILIA Protocol: open "consequence firewall" that verifies exact authority before an agent can touch money, code, infra, or regulated state, independently verifiable evidence trail for unattended agent work.
MarketNow MCP interceptor: real-time MCP proxy blocking .env reads and dangerous commands across a 9,248-server MCP marketplace, notable both as tooling and as a sign of how big the untrusted-MCP-server problem has gotten.
opencodex: local proxy translating Codex's Responses API to any backend LLM (Claude, Gemini, Grok, DeepSeek, Ollama), lets Codex CLI/App/SDK and Claude Code run against arbitrary models.

GLM-5.2 is the step change for open agents: Nathan Lambert flags GLM-5.2 as crossing a real open-agent capability threshold, worth tracking against Claude/GPT for terminal-agent-grade tasks.

Toolport: single MCP endpoint aggregating multiple tools for AI agents, one more entrant in the MCP-aggregator-gateway space.
TencentDB Agent Memory: team-level memory hub turning chats/docs/code into four governed, shareable memory assets (Chat Memory, Skill, LLM-Wiki, Code-Graph) usable across agents/frameworks.
Obscura: Rust headless browser (real V8, CDP) built specifically for AI agent web scraping/automation, positioned as a lighter/stealthier alternative to Puppeteer-class tools.
Hexis: git-backed skills/tools/context store for AI agents, versioning agent capabilities like code.
SnapState: persistent state layer for AI agent workflows, targets the same statefulness gap herdr/bb are solving from the runtime side.
Firecrawl MCP (Extract): Firecrawl's new MCP server for agent-ready web context extraction, direct competitor to Obscura/scraping-for-agents plays.

OmniRoute: MIT AI gateway aggregating 290+ providers/500+ models with quota-aware fallback and prompt-compression claiming 15-95% token savings, works with Claude Code/Codex/Cursor/OpenCode/Cline/Copilot.
Comp AI CRM: open-source CRM built agentic-first, agents do the data-finding/writing instead of a chatbot bolted onto forms.