Locara

AI Agent Marketplaces — 2026 Survey

What this is: A comparative read of the curated marketplaces for AI agents, skills, and tools as of mid-2026. Eight competing platforms, all positioning on “find a quality agent/skill, install it, run it” — with different curation philosophies, monetization models, and distribution surfaces. Why it matters: The discovery and curation layer Locara is building doesn’t exist in a vacuum anymore. Eight marketplaces are already competing on this surface. Locara’s positioning depends on knowing which spaces are saturated, which are wedge-able, and which patterns to copy. Most relevant to Locara: This is the competitive landscape Locara’s app/skill catalog enters into. Locara’s differentiation cannot be “first curated AI marketplace” — that ship has sailed. It must be specifically “first local + sandboxed + manifest-enforced curated AI marketplace.”

The eight (as of Q2 2026)

  1. GPT Store (OpenAI) — Launched January 2024. Custom GPTs created via the GPT Builder UI, browseable in a curated directory, monetized via OpenAI’s revenue program. Closed-platform: GPTs only run inside ChatGPT.
  2. Claude Skills + Anthropic Skills repo — Skills launched October 2025; the open-standard format published December 2025. Anthropic ships first-party “managed skills” (docx, pptx, xlsx, pdf, scheduling). Third-party skills live in scattered GitHub repos.
  3. MCP Hubs — Community-run indexes of Model Context Protocol servers. Three largest in 2026: mcp.so, Smithery, PulseMCP. Each has its own submission process, ranking algorithm, and curation standards. None has “won” yet.
  4. Hugging Face Spaces + Models Hub — The default registry for open-weights models and demos. Spaces are interactive demos (Gradio/Streamlit); Models is the registry. Predates the others; anchors the open-weights world.
  5. Replit Agent Market — Launched 2025. Agents that run inside Replit’s cloud workspace, with first-party hosting and revenue share. Tied tightly to Replit’s runtime.
  6. LangChain Hub — Curated chains, prompts, and agents in the LangChain ecosystem. Smaller scale, narrower scope, primarily a developer reference.
  7. Vercel Agent Gallery — Curated showcase of agents built on the Vercel AI SDK. Curation-by-invitation; biased toward agents demonstrating Vercel’s framework patterns. Distribution-by-template more than installable apps.
  8. Cloudflare AI Marketplace — Cloudflare’s bundle of AI models, agents, and tools deployable on Workers. Edge-first; less consumer-facing than the others.

Plus the long tail: Hugging Chat Assistants, Poe Bots (Quora), Microsoft Copilot Studio agents, GitHub Copilot Extensions, AI Agents Directory (1,300+ listings), Awesome-MCP-servers lists.

Recurring patterns across the eight

  • Curated, not open-floodgates. Every successful marketplace is curated — either editorial (Vercel), automated-with-human (OpenAI’s GPT Store), or community-PR (MCP Hubs). The lesson from npm/Pinokio that frictionless publishing breeds spam is internalized.
  • Discovery is the bottleneck, not publishing. The 2026 consensus: shipping an agent is solved (any developer with a model and a prompt can publish); getting it discovered in a marketplace with thousands of competing listings is where most agency-built agents fail to get traction. The Steam / App Store long-tail problem replays here.
  • Closed-platform marketplaces have stronger curation but smaller ceilings. GPT Store and Replit Agent Market are tightly bound to their host platforms. They have more control but don’t reach users outside.
  • Open-protocol marketplaces (MCP Hubs, HF Hub) compete on curation rather than platform control. They’re better positioned for cross-vendor adoption but have weaker monetization.
  • Monetization is mostly unsolved. OpenAI’s GPT revenue share is unclear/inconsistent. Replit shares revenue. MCP Hubs are mostly free. HF charges for compute, not for curation. The Shopify-or-Steam-style “0% under $1M / 15% above” model has not yet arrived for AI agents.
  • Trust signals are weak. Most listings rely on author reputation, GitHub stars, or prose descriptions. Real verification (capability declarations, sandboxing, signed builds, audited behaviors) is essentially absent.

What the marketplaces have actually figured out

  • OpenAI’s GPT Builder UX — anyone can create a GPT with no code; the builder is itself an agent. Lowering the bar to publish raised total volume but tanked average quality.
  • MCP Hubs’ submission discipline — Smithery and mcp.so review submissions, require manifests, and rank by structured signals. Closer to Homebrew than to npm.
  • Vercel Agent Gallery’s editorial bias — by curating only agents that demonstrate the AI SDK’s patterns, Vercel turned the gallery into educational content for the framework. Marketing-as-curation.
  • Hugging Face’s “Spaces as living demo” — every model has a demo any user can try without local setup. Flattens the trial barrier.
  • Anthropic’s open-standard publication — preempts the “lock-in” critique and invites cross-vendor adoption.

What the marketplaces have failed at

  • Sandboxing. None of the eight enforces capability declarations or runtime isolation as a precondition of listing. Trust is reputation-based.
  • Quality signals. Most marketplaces show install counts and prose descriptions. Verifiable capabilities, audit trails, and behavior tests are absent.
  • Cross-marketplace portability. A skill from Anthropic doesn’t run on OpenAI; an MCP server runs anywhere but quality varies wildly.
  • Long-tail discovery. Every marketplace has a 1% capturing 90% of installs problem (echoing App Store / Steam).
  • Privacy positioning. None of the eight is structurally local-first. All run in someone’s cloud (OpenAI’s, Anthropic’s, Replit’s, Vercel’s, Cloudflare’s).

Specific learnings for Locara

  1. The wedge is “local + sandboxed + manifest-enforced,” not “curated AI.” Curation is table stakes — eight competitors have it. Locara’s defensible position is the structural one: apps run on the user’s machine, in a wasm/native sandbox, with capability declarations the runtime can enforce. None of the eight does this.
  2. Don’t try to be the canonical MCP Hub. mcp.so, Smithery, and PulseMCP are entrenched. Be the curated, sandboxed subset of MCP servers — “the MCP servers that have been verified to run safely under Locara’s capability model.” Smaller catalog, higher trust.
  3. Adopt the Skills folder format directly. Don’t invent a parallel format. A Locara “skill” is just an Agent Skill folder with an additional locara.json capability manifest — leveraging the existing standard while adding the missing safety layer.
  4. Editorial curation as marketing (Vercel pattern) is worth copying. A Locara catalog of 30 high-quality apps that demonstrate the framework’s value is better than 3000 mediocre ones. Use the catalog as a teaching tool for app authors.
  5. Don’t compete on count; compete on what’s promised. GPT Store and HF Hub already have the volume. Locara’s catalog should ship with verifiable claims: “this app touches only your ~/Documents/receipts folder, never the network, uses model X, runs entirely on your device.” The verification is the product.
  6. Avoid the host-platform trap. GPT Store and Replit Agent Market are strong inside their platforms but have ceilings. Locara apps should be installable as .locapp files independent of any registry — the registry is the curated channel, not a lock-in surface.
  7. Plan monetization before the catalog scales. None of the eight has a clean dev-rev story. Locara has time to design one (Shopify’s 0%-under-$1M / 15%-above is the cleanest precedent) and ship it before the catalog grows past 100 apps. Don’t repeat the Raycast-extension-authors-paid-nothing-for-years path.
  8. Don’t try to win every category. The marketplace surface fragments by use case: GPT Store for chat-first GPTs, MCP Hubs for tool servers, HF for model demos. Locara’s lane is end-user-installable AI applications that run locally — a category none of the eight cleanly serves.
  9. Quality signals as a differentiator. Verified-publisher identity, signed builds, declared capabilities, public security review — all things the eight existing marketplaces have left undone. Locara can lead on this without building anything novel; it’s just discipline.
  10. Long-tail discovery is unsolved everywhere. Don’t pretend Locara has a magic answer. Curated catalog (10–100 apps) for v1; design for scale later. The bigger marketplaces are not on a path to solving this either.

References