Eight platforms scored on 12 measurable metrics, every value cited to public documentation. Full rubric below; corrections welcome.
Results
| Rank | Platform | Score /100 | Category |
|---|---|---|---|
| 1 | Manufact (formerly mcp-use) | 95 | MCP deployment platform |
| 2 | Cloudflare Workers | 50 | Edge compute platform |
| 3 | Smithery (Arcade) | 44 | Registry + connection layer |
| 4 | MintMCP | 42 | Enterprise MCP gateway |
| 5 | Vercel | 33 | Web app platform |
| 6 | MCPHosting.io | 26 | MCP hosting service |
| 7 | Stainless | 23 | MCP server generator |
| 8 | Zapier MCP | 17 | Hosted integration gateway |
Two category caveats before anything else. Zapier MCP doesn't host user code at all, and Stainless generates servers you host elsewhere; both score low on deployment-lifecycle metrics by design, and their scores should be read as "fit for the MCP deployment job," not product quality. The category column exists so you can discount accordingly.
Why this benchmark exists
Every "best MCP platform" article we found, including our own, is prose. Vendor pages claim "deploy in minutes" and "full observability" with nothing to compare against. This page instead pins 8 platforms against 12 concrete metrics, each with a published rubric and a source link, so the numbers can be checked and re-run by anyone.
Methodology
- Evidence base: public documentation, pricing pages, and official blogs, fetched during the week of August 3–10, 2026. No private demos, no vendor briefings.
- Documentation rule: a capability counts only if it is documented publicly. "Not documented" scores as absent, even if the feature may exist. This keeps the benchmark reproducible.
- Scoring: each metric is scored 0–5 against the rubric published in its section. All 12 metrics are weighted equally; the total (max 60) is normalized to 100.
- Conflicts: where a vendor's marketing page and docs disagree, docs win, and the discrepancy is noted.
- Corrections: if you're a vendor and we've missed documented functionality, email [email protected] with a link. We'll re-score and log the change in the changelog below.
- Disclosure: this benchmark is maintained by Manufact, which is also ranked in it. The metrics reflect what we believe matters for teams deploying their own MCP servers and apps. Every value for every platform, including ours, is cited so the bias is auditable. Platforms optimized for other jobs (gateways, generators) score low on axes they don't target.
For the prose version of this comparison, see the 8 best MCP deployment platforms in 2026.
The 12 metrics
1. MCP SDK / language support
Languages with first-class, documented support for building and deploying a server. Rubric: 0 = none, 2 = one language, 3 = two, +1 for an any-language escape hatch (Docker/URL), cap 5.
| Platform | Value | Score |
|---|---|---|
| Manufact | TS + Python (mcp-use SDKs), FastMCP preset, any via Dockerfile (docs) | 4 |
| MintMCP | TS + Python FastMCP, any via custom command/Docker (docs) | 4 |
| Smithery | JS module upload; any language via URL publish (docs) | 3 |
| MCPHosting.io | Python FastMCP + Node.js (site) | 3 |
| Cloudflare | TS (Agents SDK) (docs) | 2 |
| Vercel | JS/TS (mcp-handler) (repo) | 2 |
| Stainless | TS only, generated (docs) | 2 |
| Zapier | No server SDK; consumption only (docs) | 0 |
2. Programmatic testing of deployed servers
Can you invoke a deployed server's tools via API/CLI and run automated test suites? Rubric: 0 = none documented, 2 = health checks/logs only, 4 = API/CLI tool invocation, 5 = plus test suites with CI gating.
| Platform | Value | Score |
|---|---|---|
| Manufact | REST API, CLI client tools call, eval suites with CI gating (testing, cross-client) | 5 |
| Smithery | CLI smithery tool call, REST, typed SDKs (CLI docs) | 4 |
| MintMCP | Health checks pre-traffic, Admin MCP status, logs (hosted CLI) | 2 |
| Cloudflare | Interactive only (Inspector, Playground) (test guide) | 0 |
| Vercel | Local Inspector workflow only (docs) | 0 |
| Stainless | Local Inspector only (docs) | 0 |
| Zapier | Manual ("ask your AI"); test calls billed (usage) | 0 |
| MCPHosting.io | Not documented | 0 |
3. Embedded MCP Inspector
An inspector inside the platform's own dashboard for exercising tools, resources, and prompts against the deployed server. Rubric: 0 = none, 2 = external/adjacent playground, 5 = embedded full inspector.
| Platform | Value | Score |
|---|---|---|
| Manufact | Cloud Inspector embedded in dashboard; also maintains the open-source MCP Inspector (inspector, analytics traffic toggle) | 5 |
| Cloudflare | AI Playground acts as a remote MCP client; separate property, not in the Workers dashboard (blog) | 2 |
| Smithery | Server page "try" flow; playground undocumented (build docs) | 2 |
| Vercel / MintMCP / Stainless / Zapier / MCPHosting.io | Not documented | 0 |
4. Deployment methods
Three paths checked: GitHub-connected deploys, CLI deploys, and platform-as-MCP (an agent can drive deploys through an MCP interface). Rubric: about 1.7 points per path.
| Platform | GitHub | CLI | Via MCP | Score |
|---|---|---|---|---|
| Manufact | Yes (GitHub App) | Yes (npx mcp-use deploy) | Yes (official Manufact MCP server with deploy tools) | 5 |
| Cloudflare | Yes (Workers Builds) | Yes (Wrangler) | Not in MCP docs | 3 |
| Vercel | Yes | Yes | No | 3 |
| MintMCP | Not documented | Yes | Yes (Admin MCP) | 3 |
| Smithery | Not documented | Yes (publish) | No | 2 |
| Stainless | Release flow via GitHub | Build CLI, no hosted deploy | No | 2 |
| MCPHosting.io | Yes | Not documented | No | 2 |
| Zapier | N/A (dashboard only) | No | No | 0 |
5. Publishing checks for MCP stores
Automated validation against ChatGPT Apps Store / Claude Connectors / registry requirements. Value = documented check count. Rubric: 0 = none, 1 = private-store approval only, 3 = registry-level checks, 5 = store-requirement checks with fix guidance.
| Platform | Value | Score |
|---|---|---|
| Manufact | 6 check categories mapped to ChatGPT + Claude store requirements, per-check fix guidance, autofix, submission pack (publishing checks, docs) | 5 |
| Smithery | Metadata scan, OAuth 401 conformance, server card fallback, vendor verification checklist (publish docs) | 3 |
| MintMCP | Private-store admin approval + pre-release health checks (quickstart) | 1 |
| Cloudflare / Vercel / Stainless / Zapier / MCPHosting.io | None documented | 0 |
6. AI models in the cloud testing/chat interface
Hosted chat for testing the deployed server, and how many models/providers it can swap. Rubric: 0 = none, 2 = hosted chat, +2 model swap across providers, +1 documented full model list. Manufact is capped at 4 because its exact chat model list is undocumented beyond the three named families.
| Platform | Value | Score |
|---|---|---|
| Manufact | Dashboard Chat + model swap across GPT, Claude, Gemini (3 provider families) (site, quickstart) | 4 |
| Cloudflare | AI Playground with the Workers AI open-model catalog (Llama, Gemma families; no GPT/Claude) (models) | 3 |
| All others | No hosted testing chat documented | 0 |
7. Analytics completeness
Count of documented MCP-relevant dimensions/filters (tool calls, errors, latency percentiles, clients, sessions, geography, tokens, replay, traffic filters). Rubric: 0 = none, 1 = generic request metrics, 2–3 = a few MCP dimensions, 4 = rich but adjacent focus, 5 = 8+ MCP-native dimensions.
| Platform | Documented dimensions | Score |
|---|---|---|
| Manufact | 9+: tool calls, sessions, error rate + issue triage, p50/p95/p99 per tool, client/version/protocol, geography, AI-clustered goals, tokens, session replay, internal-traffic toggle (analytics, sessions) | 5 |
| MintMCP | 7: users, tokens, cost, model mix, harness/client filter, audit log, SIEM export; gateway-usage oriented (usage) | 4 |
| Smithery | Tool calls, usage patterns, runtime logs (build) | 2 |
| MCPHosting.io | Request/response pairs, tool usage, user behavior (claimed, no docs) (site) | 2 |
| Zapier | Activity log + audit log (security) | 1 |
| Cloudflare | Generic Workers metrics, no MCP dimensions | 1 |
| Vercel | Generic function observability, no MCP dimensions | 1 |
| Stainless | X-Stainless-MCP header for self-serve segmentation (configure) | 1 |
8. Developer experience (0–5, rubric-derived)
Five sub-criteria, one point each: complete public docs; zero-config first deploy; local dev + templates; self-serve pricing; consistency (no deprecated primary paths, no docs/marketing conflicts).
| Platform | Score | Deductions |
|---|---|---|
| Manufact | 4 | Docs and pricing page disagree on plan limits; test-suite clients limited to Inspector + ChatGPT today |
| Cloudflare | 4 | McpAgent deprecated mid-migration to SDK v2; templates in transition |
| Vercel | 4 | MCP surface is thin (one handler package); auth server BYO |
| Smithery | 3.5 | Post-acquisition docs gaps (playground, GitHub deploys undocumented) |
| MintMCP | 3 | No self-serve pricing; quote-based onboarding |
| Stainless | 3 | MCP generation explicitly experimental; no hosting |
| Zapier | 3 | Easy start, but Beta label and per-call task billing |
| MCPHosting.io | 1.5 | No docs site, no CLI/API reference, no paid pricing published |
9. Agent readiness
Can an AI agent operate the platform itself end to end? One point each: official platform MCP server; llms.txt / agent-readable docs; full REST API; agent-drivable deploys; agent-drivable testing.
| Platform | Value | Score |
|---|---|---|
| Manufact | Official MCP server (deploy, events, tool breakdowns), llms.txt docs, REST API, CLI (llms.txt) | 5 |
| Smithery | REST API, llms.txt, typed SDKs, registry API (llms.txt) | 4 |
| MintMCP | Admin MCP for deploys/config/status (docs) | 4 |
| Cloudflare | Official MCP servers for the Cloudflare platform (blog), full API; MCP-driven deploys not documented | 4 |
| Zapier | llms.txt, agent install manifests (Claude Code, Cursor, Copilot, Kiro) (llms.txt) | 2 |
| Vercel | Full REST API and CLI; no platform MCP server in MCP docs | 2 |
| Stainless | Build CLI/API | 2 |
| MCPHosting.io | Not documented | 0 |
10. Time to deploy (vendor-claimed)
Docs-based benchmark, so values are vendor claims, marked as such. Rubric: 5 = under 2 minutes claimed, 4 = under 5, 3 = "minutes"/standard git deploy, 2 = multi-step release pipeline, 1 = no claim and no documented path.
| Platform | Claim | Score |
|---|---|---|
| Manufact | "under 60s from git push to live" (site) | 5 |
| Cloudflare | "less than two minutes" (blog) | 5 |
| MintMCP | Quickstart "~5 minutes", server start "under 1 minute" (quickstart) | 4 |
| Zapier | "under 5 minutes" setup (site) | 4 |
| MCPHosting.io | "minutes" (site) | 3 |
| Vercel | No claim; standard git build | 3 |
| Smithery | "minutes" (connection, not hosting) | 3 |
| Stainless | Spec to SDK to release pipeline to external hosting | 2 |
11. Preview deployments on PR
A live server URL per branch/pull request. Rubric: 5 = yes documented, 4 = claimed on marketing page only, 2 = preview builds without running deployments, 0 = none.
| Platform | Value | Score |
|---|---|---|
| Manufact | Per-branch URL <slug>--br-<branch>.run.mcp-use.com/mcp, Hobby+ (deployments) | 5 |
| Vercel | Preview deployment on every push, all plans (plans) | 5 |
| Cloudflare | Branch preview URLs with PR comments (docs) | 5 |
| MCPHosting.io | "Open a PR, get a branch deployment" (homepage only) (site) | 4 |
| Stainless | Preview SDK/MCP builds on spec PRs, no running server (preview builds) | 2 |
| Smithery / MintMCP / Zapier | Not documented / N/A | 0 |
12. End-to-end testing
Automated tests that exercise the deployed server inside real AI clients, including browser-rendered widget verification. Rubric: 0 = none, 1 = recommended without tooling, 3 = single-client automation, 5 = multi-client E2E with LLM-judged assertions and widget rendering.
| Platform | Value | Score |
|---|---|---|
| Manufact | E2E publish checks run live in ChatGPT and Claude, verifying tool calls and widget rendering ($2/run); cross-client eval suites with LLM-judge pass/fail per client and model ($1/run) (publish checks, testing) | 5 |
| Cloudflare | Best-practices doc recommends evals; no tooling shipped (docs) | 1 |
| All others | Not documented | 0 |
Score matrix
| Metric | Manufact | Cloudflare | Smithery | MintMCP | Vercel | MCPHosting | Stainless | Zapier |
|---|---|---|---|---|---|---|---|---|
| 1. SDK/languages | 4 | 2 | 3 | 4 | 2 | 3 | 2 | 0 |
| 2. Programmatic testing | 5 | 0 | 4 | 2 | 0 | 0 | 0 | 0 |
| 3. Embedded Inspector | 5 | 2 | 2 | 0 | 0 | 0 | 0 | 0 |
| 4. Deploy methods | 5 | 3 | 2 | 3 | 3 | 2 | 2 | 0 |
| 5. Publishing checks | 5 | 0 | 3 | 1 | 0 | 0 | 0 | 0 |
| 6. Models in chat | 4 | 3 | 0 | 0 | 0 | 0 | 0 | 0 |
| 7. Analytics | 5 | 1 | 2 | 4 | 1 | 2 | 1 | 1 |
| 8. Developer experience | 4 | 4 | 3.5 | 3 | 4 | 1.5 | 3 | 3 |
| 9. Agent readiness | 5 | 4 | 4 | 4 | 2 | 0 | 2 | 2 |
| 10. Time to deploy | 5 | 5 | 3 | 4 | 3 | 3 | 2 | 4 |
| 11. PR previews | 5 | 5 | 0 | 0 | 5 | 4 | 2 | 0 |
| 12. E2E testing | 5 | 1 | 0 | 0 | 0 | 0 | 0 | 0 |
| Total /60 | 57 | 30 | 26.5 | 25 | 20 | 15.5 | 14 | 10 |
| Normalized /100 | 95 | 50 | 44 | 42 | 33 | 26 | 23 | 17 |
Reading the results
The gap between first and second place is mostly three metrics where the field scores zero: programmatic testing, publishing checks, and E2E testing. Manufact, Smithery (for testing), and MintMCP (for analytics) are the only platforms with documented answers to what happens after deployment.
Cloudflare's second place is earned on infrastructure fundamentals: deploy speed, previews, OAuth depth, and an agent-operable platform. Its zeros are all lifecycle metrics. A team pairing Cloudflare hosting with external testing and analytics tooling would close much of the gap, at integration cost.
Smithery over MintMCP is close and reasonable to dispute. Smithery wins on programmatic tool invocation and registry checks; MintMCP wins on analytics and language support. They target different buyers.
Zapier's 17 and Stainless's 23 say little about those products' quality at their actual jobs (integration gateway; SDK generation). They're included because teams shortlist them for MCP work, and the scores quantify the category mismatch.
Limitations
This is a documentation benchmark, not a load test. Time-to-deploy values are vendor claims, no latency or uptime was measured, and pricing was not normalized into a metric. A hands-on round (deploying the same reference server to every platform with a stopwatch and a test harness) is the planned follow-up. Documentation quality also correlates with score by construction: a platform with undocumented features will underscore. That is intentional.
Changelog
- 2026-08-10: initial publication. Evidence collected August 3–10, 2026.
Maintained by Manufact. Manufact is ranked in this benchmark; the methodology, rubrics, and every underlying citation are published above so you can re-score it yourself. Vendor corrections: [email protected].
