Manufact Review

MCP is still young infrastructure. Betting on it means betting Anthropic's protocol becomes the standard agents call tools through — that's not settled, but Manufact isn't waiting around. The mcp-use SDK is open source, in both TypeScript and Python, and the Cloud Inspector lets you debug a deployed server across GPT, Claude, and Gemini mid-session. That's a real capability, not a wrapper.

Pricing is usage-based and transparent: $5 free credits, $25/month Hobby with $30 in credits, $250/month Startup with 3M requests. No Dockerfiles, no YAML — push to GitHub and it deploys.

The tradeoff: you're building on a protocol spec that could shift, and switching off Manufact Cloud later means re-hosting servers elsewhere. Fine for a pilot, riskier for mission-critical infrastructure today.

The architecture decision here is actually sound: mcp-use is open source, so I'm not fully hostage to Manufact Cloud if I want to self-host the SDK output elsewhere. That matters for a VP signing off on a 3-year bet — the exit ramp exists even if it's not comfortable. Cloud Inspector and the auto-deploy-from-GitHub flow are the kind of operational maturity I want to see before I let a team ship production tool servers.

What gives me pause is the credit-based pricing model layered onto usage I can't fully predict yet — 300k requests/month on Hobby sounds fine until a widget goes viral inside ChatGPT and burns through pay-as-you-go fast. Cross-client eval suites across ChatGPT, Claude, and Cursor are the right instinct for regression testing against a moving spec.

Organizations with RBAC at the Startup tier (10 members, $250/month) is table stakes, not differentiation. I'd run this for a pilot team building Apps Store submissions before I'd trust it as core infra.

$0.01 per credit. Free gives $5/month, Hobby $25 for $30 in credits plus overage, Startup $250 for $300 plus overage. That's a usage meter, not a seat count. Team of 50 doesn't map cleanly to this pricing — it maps to request volume, tool-call volume, session replay volume. Harder to budget, easier to pay for only what you use.

Annual billing saves relative to monthly: Hobby at $300/year versus $25×12=$300 — no discount there actually, same number. Startup is $3,000/year versus $250×12=$3,000. Flat either way, so annual just locks the commitment, doesn't buy a discount. Worth asking before signing yearly.

Enterprise starts "custom, from $1,000/month" per their pricing page — a real floor, not just "contact sales" fog. That's useful. No invoicing terms, payment net-days, or cancellation window listed for the paid tiers. Auto-renewal behavior on annual plans isn't stated either. Fine for a $25 card swipe. Worth a contract read before $3,000/year.

Analytics covering traffic, tool-call volume, latency, session replay, and error rates is the right instrumentation set for a protocol with almost no observability tooling of its own. That's the part I'd actually care about at 2am. Alerts are mentioned but I couldn't find detail on routing, escalation policies, or whether they integrate with PagerDuty or Opsgenie — for an SRE that's the gap between 'we have metrics' and 'we have an oncall workflow.'

Edge-runtime support plus multi-region on the Startup tier ($250/month) suggests they've thought about blast radius, but unlimited requests only show up on Enterprise, meaning capacity planning under real load means a sales call.

The mcp-use SDK being open source is good for postmortems — I can read the actual RPC handling instead of trusting a black box when debugging a cross-model failure between GPT and Claude mid-incident.

The thing I keep coming back to is the MCP Inspector working without a live LLM. That's the kind of feature only someone who's been burned by flaky model calls during debugging would bother shipping. Same instinct behind the 'vibe coding' scaffold — it's a nice-to-have on the surface, but it tells you the team is thinking about the first ten minutes, not just the demo.

First hour is npx create-mcp-use-app and you're looking at a widget in a local inspector. That's a real 'wow, okay' moment, not homework. Where it gets murkier is month three: this is a brand-new protocol ecosystem, and discoverability past the starter templates (Chart Builder, Maps Explorer) depends on docs I can't fully audit from here.

Embedded chat UI as a shareable artifact is a nice daily-use touch — lets you show a non-technical teammate the server without a client install. Free tier's $5 credits versus the jump to Startup is the kind of thing you'll feel on a Tuesday, not launch day.

MCP is still a young spec. Betting a hosting business on it is the same move Heroku made on buildpacks before Docker ate the category — worked out fine for Heroku, less fine for a dozen clones nobody remembers. Manufact's answer is to open-source the SDK and sell the convenience layer on top. Smart hedge.

The credit-based pricing is where I'd slow down. $25 a month buys $30 in credits at Hobby, then pay-as-you-go kicks in with no visible per-unit rate in what I could find. That's a meter you can't read until you're already paying it.

Marketing calls itself 'the MCP cloud' — fine, grounded, not 'the future of agents' or similar superlative. Submission checklist to the ChatGPT Apps Store and Claude Connectors is a concrete, checkable claim, not vapor. Enterprise floor starts at $1,000/month, custom above that — normal, not hidden.