🛠️ 9 Drops: Multi-Agent Plugins, Skills, Web Agents

Token prices are going up. TechCrunch flagged it last night and the timing is not subtle: the big labs are pre-IPO, costs are climbing, and the operators in the middle are the ones who absorb the delta.

I will say this plainly: the window where you can build on hosted APIs at current rates is closing. Not this quarter, maybe not this year. But the trajectory is clear, and every operator I respect is doing the same thing in response: routing more work to open-weight models, trimming context aggressively, and building pipelines that can swap providers without a rewrite.

This week's shortlist reflects that instinct. The two most-starred repos trending right now are a multi-harness agent plugin marketplace that runs across Claude, Codex, Cursor, and Gemini from a single harness, and a skills repo for operators who want reusable, portable Claude Code abilities that are not locked to one session. Both are hedges against vendor lock-in dressed up as productivity tools.

Nine drops total today. Two are skills, seven are repos. Concrete, cloneable, most of them running this week.

---

[Skill] wshobson/agents, Multi-harness agentic plugin marketplace for Claude Code, Codex CLI, Cursor, OpenCode, and Gemini CLI. One install, one plugin architecture, five runtimes. If you are hedging against any single AI provider and want your agent plugins to survive a toolchain swap, this is where to start. 36,482 stars and climbing.

[Skill] mattpocock/skills, A real engineer's .claude directory, open-sourced. Skills for production workflows, straight from someone who ships. 120,309 stars. Worth cloning and reading before building your own.

[Repo] mvanhorn/last30days-skill, Agent skill that pulls from Reddit, X, YouTube, HN, Polymarket, and the web, then synthesizes a grounded 30-day summary on any topic. Useful for competitive research, market pulse checks, and pre-build landscape scans. 30,586 stars and trending today.

[Repo] Leonxlnx/taste-skill, Gives your AI good taste. Specifically, it stops the model from generating generic, pattern-matched slop by injecting aesthetic constraint into the generation loop. 36,432 stars. Run it once and you will not go back. Trending this week.

[Repo] zarazhangrui/frontend-slides, Creates polished web slides using a coding agent's frontend skills. Not a presentation tool, a demonstration of what a well-directed Claude Code agent can do on pure frontend output. 20,641 stars.

[Repo] nottelabs/notte, Framework for building web agents and deploying serverless browser automation on reliable infra. If you need a production-grade headless browser layer without managing your own Playwright fleet, this is the operator path. 1,973 stars.

[Repo] camel-ai/camel, The oldest multi-agent framework still in active production development. Role-playing agent architecture, scaling law research baked in, and 17,131 stars of community signal. If you are designing cooperative agent systems, CAMEL gives you a tested foundation.

[Repo] mudler/LocalAI, Run any model locally: LLMs, vision, voice, image, video. No GPU required. The practical hedge for operators who want API-independence without Ollama's surface constraints. 46,732 stars.

[Repo] chroma-core/chroma, The lightweight open-source vector DB operators reach for first when building RAG. Easy local setup, Python-native, and no infrastructure overhead for early-stage pipelines. Start here before graduating to Milvus.

---

[MCP] antonbabenko/deliberation, One MCP server that routes your query to Codex, Gemini, Grok, and 400+ OpenRouter models for second opinions or arbiter-mediated consensus. Works across Claude Code, Codex, Cursor, Kiro, and OpenCode. The non-obvious use: pipe your architecture decision through deliberation before committing, then surface the disagreement, not the agreement. Disagreement is the signal.

[MCP] DomDemetz/claude-soul, Self-correcting learning engine for Claude Code with persistent identity, behavioral pattern tracking, and cross-session memory. Not session context, actual behavioral drift correction across sessions. Wire it in if you run long or complex projects where Claude's defaults keep reverting on you.

---

Token prices are rising and operators are the first to absorb it. TechCrunch names it the "Tokenpocalypse": as the big labs head toward public markets, pricing pressure shifts onto API users. Build your pipelines to swap providers, keep context windows lean, and route classification and routing tasks to smaller open-weight models now.

Weaviate shipped Engram, a conversational memory layer that updates without blocking inference. The bottleneck it solves is real: most agent memory architectures write-on-commit, which adds latency proportional to conversation length. Engram decouples the write path. Operators building long-context agent workflows should watch how it integrates with Chroma and Milvus backends. (Lokmattimes)

Vercel Sandbox now supports persistent drives in private beta. Mount a drive at a configurable path, detach it, reattach it to another sandbox. For operators who run ephemeral compute environments and hit the stateless wall, this is the structural fix. (Vercel Blog)

Anthropic's internal safety team acknowledged Claude is approaching self-improvement capability. The public statement calls for a coordinated slowdown across labs. Practically speaking, it tells you how close the next capability jump is. The operators who have their harnesses and eval loops in place will be ready when it lands. (Dallas Express)

---

This week: MCP servers. Wire real external tools into Claude so it acts on them, not just talks about them.

The non-obvious part: most operators set up one MCP server and stop. The pattern worth stealing is scope separation. Project-scoped servers (.mcp.json at the repo root) travel with the repo and are visible to anyone who clones it. User-scoped servers (~/.claude.json under the mcpServers key) run across every project on your machine and are invisible to collaborators. Put your GitHub or Supabase server at user scope. Put project-specific tools at project scope.

1. Add a user-scoped server: open ~/.claude.json, add your server under the mcpServers key with a command and args field. Example: "github": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"] }. 2. Add a project-scoped server: create .mcp.json at the repo root with the same shape. Commit it. Everyone who clones the repo gets the server. 3. Confirm the connection inside your session: run /mcp. You will see a list of connected servers and their available tools. If a server shows as disconnected, the command path is wrong before anything else.

You will know it worked when Claude stops asking you to paste a file's contents and just reads it directly from the connected source.

Claude Code docs: MCP

---

The move: route your cheap, high-volume tasks to a local model and keep the frontier model for the hard calls.

The Tokenpocalypse only bites if every call hits a hosted frontier API at full rate. Most pipelines have a fat tail of cheap work (classification, routing, extraction, reformatting) that a small open-weight model handles fine, and that is the work you want off your metered bill before pricing climbs.

1. Stand up LocalAI with a small open-weight model (run it on your own box, or rent a GPU on Runpod if you need more headroom). It exposes an OpenAI-compatible endpoint, so pointing your client at it is a base-URL swap, not a rewrite. 2. Tag each task by difficulty: send classification, routing, and extraction to the LocalAI endpoint, and keep reasoning and final output on the frontier model. 3. Log the token split for a week so you can see which calls actually needed the expensive model.

You will know it worked when your hosted-API bill drops and the output quality on the calls that matter does not move. Most operators find 40 to 60 percent of their calls were never worth frontier pricing.

This Friday's kit is Ad Factory: a brand-agnostic video ad production pipeline that ships finished 9:16 cinematic ads at a fraction of what an agency charges. The core technique is a bokeh trick that makes AI-generated footage stop reading as AI, which is the single tell killing most operators' video ad attempts right now. If you have been watching AI video get good enough to use in paid placements and wondering why it still looks fake, Friday's kit is the answer. Grab operator access before Friday.

---

Nine drops, two Stack picks, four signals. The thread running through all of it: the operators who hedge their stack now, before the pricing shift lands, will have pipelines that cost half as much to run when it does. That is not a prediction. It is a design decision you can make today.

See you Tuesday.

This issue contains affiliate links; we may earn a commission.