Smithery

Arcade.dev announced on 5 August 2026 that it had acquired Smithery, with co-founder Anirudh Kamath joining Arcade's team (https://www.arcade.dev/blog/smithery-joins-arcade/). Smithery still runs at smithery.ai with the same product, and the announcement frames the deal as combining Smithery's developer-facing registry with Arcade's enterprise security/governance focus — Arcade's own pitch is being 'the actions runtime for enterprise AI agents.' The GitHub org behind the CLI has already moved to arcadeai-labs (the old smithery-ai/cli URL now redirects there; https://api.github.com/repos/arcadeai-labs/smithery-cli). What this means for a buyer today: pricing and the free tier are unchanged, but anyone evaluating Smithery for the long term should treat 'independent, developer-first MCP registry' as no longer accurate — it is now a business unit inside a security/governance vendor, and future direction (support terms, enterprise-only features, pricing changes) is Arcade's call.

Smithery is two things bundled together. First, a public registry of MCP servers — 15,482 listed as of 18 September 2026 per its own API (https://registry.smithery.ai/servers), of which 191 carry Smithery's 'verified' badge. Second, a hosted connection layer: instead of an agent developer wiring up OAuth, token refresh, and credential storage for every third-party service separately, Smithery's Connect API and CLI (smithery mcp add <server>) handle the OAuth dance and store credentials so a connection made once is reusable across Claude, GPT, and other MCP-capable clients (https://smithery.ai/docs/use/connect). It also lets developers publish their own MCP servers for distribution with usage analytics.

Three tiers, publicly listed at https://smithery.ai/pricing (confirmed via a rendered fetch on 18 September 2026, since the pricing numbers are not present in the static HTML — they load client-side): Hobby is free with 50K RPCs/month and 3 namespaces; Pay as you Go is $10/month covering 100K RPCs with 100 namespaces, then $0.10 per additional 1K RPCs; Custom is contact-sales, adding custom rate limits, an uptime SLA, and Slack support. This is usage-metered pricing (by RPC — calls through the Connect layer), not a flat per-seat model, so cost scales with how much an agent actually calls tools rather than headcount.

The open-source component — the smithery CLI, published to npm (https://www.npmjs.com/package/smithery, latest 1.2.0 as of 31 May 2026 per the registry: https://registry.npmjs.org/smithery/1.2.0) and mirrored on GitHub at arcadeai-labs/smithery-cli — is licensed AGPL-3.0 (confirmed against the repo's actual LICENSE file: https://raw.githubusercontent.com/smithery-ai/cli/main/LICENSE), not the more common MIT or Apache-2.0 seen on most CLI tooling. AGPL's network-use clause means that if you modify the CLI and run the modified version as a network-facing service for others, you must offer the modified source to those users — a real constraint for anyone planning to fork it into their own hosted product, though it imposes no obligation on ordinary local use.

The CLI repo (835 stars, 98 forks as of 18 September 2026: https://api.github.com/repos/arcadeai-labs/smithery-cli) last had a code push on 31 May 2026, shipping v1.2.0 — nearly four months before this check, though the repo itself still shows recent GitHub-side activity (last API updated_at 17 September 2026) and 44 open issues, suggesting active use even without new commits. This predates the Arcade acquisition (5 August 2026) by over two months, so the CLI's release cadence hasn't visibly changed since the deal — but there's also no confirmed roadmap for it post-acquisition.

Smithery's homepage links to 'Browse 21,857+ MCPs' (https://smithery.ai/), but the registry's own public API returns a total count of 15,482 servers as of 18 September 2026 (https://registry.smithery.ai/servers?pageSize=1). The two numbers may be counting different things — total ever published versus currently listed and active, for instance — but Smithery doesn't explain the discrepancy anywhere public, and the API is the more conservative, checkable figure.