agentik-host — MECE Market Analysis

Date: 2026-05-28 08:16 UTC Product: agentik-host — MCP-native file hosting for AI agents. Two tools (upload_file, get_file_url), signed-JWT access tokens, free tier via tweet-share, paid tiers $12 / $39. Backblaze B2 underneath (currently 4 MB Vercel-Route-Handler ceiling, direct-to-B2 PUT planned). Method: Three parallel Exa-driven research streams; synthesis below. Every number traces to a primary-source URL. Constraint honoured: No marketing-surface comparisons. Strict focus on economic moats, value-chain position, switching cost.

TL;DR (3 paragraphs)

The market exists, the wedge is real, the window is 12 months. TAM is bounded by ~55M paid AI seats globally (not the 900M ChatGPT WAU pool); SAM is ~700K–1.1M MCP-tool-using, English-speaking, file-touching individuals today, growing with the protocol's underlying ~50× annual SDK-download curve. SOM in year one is realistically 200–450 paid users / $40K–$100K ARR (base), with a $500K-ARR bull case tied to whether tweet-share virality matches R2's beta-launch trajectory.

The three incumbents that matter — Tigris, Vercel Blob, Cloudflare R2 — all share one structural flaw: their pricing model is mis-shaped to agentic call patterns. Tigris bills per-request, Vercel bills per-byte-through-CDN with a $0.15/GB egress markup over the R2 they originally resold, R2 bills $4.50/M Class A ops. An LLM agent's natural rhythm — list-peek-fetch-fetch-again-write-small — punishes the customer on each model. None of the three can drop their fee structure without breaking the bundling thesis that lets storage subsidize the rest of their platform. Tigris is the closest threat (MCP-native, S3-compat, $25M Series A funding the right roadmap) but is structurally blocked from regulated workloads (no Object Lock, no HIPAA BAA, no FedRAMP) until late-2026 at earliest.

Power-user sentiment is unanimously aligned with agentik-host's existing design choices on the two highest-frequency pains — (A) MCP has no binary upload primitive in the protocol; everyone is building duct-tape staging servers, and (E) OAuth on MCP works in theory and breaks in practice, so "token-in-URL, no signup" is the consensus winning pattern. agentik-host's upload_file + tweet-share-token IS the answer the cohort is independently arriving at. The biggest gap is that we are not explicitly marketing the screenshot-from-MCP → permanent URL workflow, which is by quote-volume the single most-recurrent agent file scenario in the wild.

1. TAM / SAM / SOM with verifiable proxies

1.1 TAM — three independent proxies, triangulated

Proxy A — MCP SDK downloads (protocol-surface signal). Anthropic + Linux Foundation, Dec 2025: 97M monthly SDK downloads (Python + TypeScript) and 10,000+ active public MCP servers. (anthropic.com, agentmarketcap.ai). Downloads are CI-heavy but the 17-month curve (2M → 22M → 45M → 97M) is the cleanest "is the protocol mainstream" signal.

Proxy B — Aggregated MCP-capable client populations (disclosed):

• Claude Code: ~$2.5B ARR Feb 2026 → 1M–10M paid seats at $20–200/mo (venturebeat.com)
• Cursor: 1M DAU, 2M total users, 1M paying, 50K enterprise teams, $2B ARR (sacra.com/c/cursor)
• Windsurf: 1M+ active, 4,000+ enterprises (windsurf.com)
• ChatGPT (MCP since Apr 2025): 900M WAU, 50M paying (backlinko.com/chatgpt-stats)

De-duped: ~900M MCP-reachable humans, ~55M paying AI seats.

Proxy C — AI-tooling developer pool. Stack Overflow 2025, 49K respondents: 84% use or plan to use AI tools, 51% daily. GitHub's ~180M dev population × 84% = ~150M devs in the AI-tooling pool (survey.stackoverflow.co/2025).

TAM convergence: the file-hosting-relevant TAM is bounded by paying-AI seats, not WAU. ~55M paid AI seats is the cap; the 900M free-tier WAU is upside if conversion ramps. At a midpoint $18/mo blended ARPU on even 0.5% of paying seats, the headline math is real but capped at low-mid-eight-figure ARR within 24 months — this is a wedge product, not a $1B SAM.

1.2 SAM — narrower filters

Stack the filters from the protocol-using subset:

SAM today ≈ 700K–1.1M addressable individuals. Growing with the underlying protocol curve: even a 5× annual decay on the 50× SDK-download growth rate puts SAM at 3–5M within 12 months.

1.3 SOM — 12-month realistic capture

Adoption-curve anchors:

SOM scenarios (12-month):

1.4 Confidence + adjacent markets

Sturdiest: Cursor 1M DAU / $2B ARR (Series D filing). MCP 97M downloads / 10K servers (Anthropic statement). SO 2025's 84% AI-tooling adoption. Solid: Claude Code WAU (derived from ARR, not disclosed). Windsurf 1M (self-report). Mostly vibes: "% of agent sessions touching files" (no public number; 20–35% inferred). Adoption-curve mapping by analogy.

Three adjacent markets to map post-SOM:

1. n8n / Zapier MCP-flow operators — non-coders who need file handoffs between automations. Proxy: n8n's 200K+ workflow-builders.
2. Lovable / Bolt.new / Replit AI-generated-app hosting — apps that need user-uploaded files but builders don't wire S3. Proxy: 92% of US devs use AI coding tools daily (vibecoding.app).
3. Agentic-research tools (Manus, Devin, ChatGPT Agent) producing PDFs/datasets/images that need durable URLs.

2. Top 3 incumbents — structural vulnerabilities (MECE: tech debt, business model, regulatory)

Selected for threat surface (could realistically eat our SOM in 12 months), not brand awareness. Discarded mcp.gd as a peer (single-operator, ~15-day listing age) in favor of R2 (Cloudflare's millions-of-devs distribution).

2.1 Tigris Data — the bull-case incumbent

Value-chain position. Sells globally-distributed S3-compatible object storage direct (self-serve usage-based) and OEM-bundled into "developer-first clouds" (Fly.io, Beam). $25M Series A (Spark + a16z, Oct 2025) explicitly funds exiting big-cloud onto Tigris-owned hardware. They pay Fly compute today but are moving off it.

Switching cost a customer pays leaving Tigris. Low. Intentionally S3-API-compatible; bucket re-pointed with an endpoint URL change. This is their wedge AND their structural weakness — no proprietary lock-in surface.

Wedge against Tigris: ship Object Lock + KMS day one; market explicitly to regulated agent workloads (healthcare-coder agents, finance copilots). Price agent ops (PUT/LIST) at half Tigris's per-request rate and commit to free egress with no "extraordinary bandwidth" escape clause.

2.2 Vercel Blob — the bundled-distribution incumbent

Value-chain position. Storage line item on the Next.js hosting invoice. Originally R2-backed (TechCrunch May 2023) — now "S3-backed" per GA blog (May 2025). They are reselling upstream object storage with a CDN markup, not running their own bytes.

Switching cost a customer pays leaving Vercel Blob. High. @vercel/blob SDK only authenticates via VERCEL_OIDC_TOKEN (Vercel-only) or a long-lived BLOB_READ_WRITE_TOKEN. No S3 API surface. Per PutPut: "You cannot use it on AWS, Cloudflare, or a VPS."

Wedge against Vercel Blob: "Works anywhere an HTTP client runs" — agents on Modal/Fly/local Claude Code can hit us with one token, no platform tax. Flat-rate or per-GB egress at half Vercel's $0.05–$0.06/GB stack. No per-second op rate limits below several-hundred/sec (sized for fleet-of-agents, not human dashboard users). Direct-from-storage public URLs that don't go through our edge — no shared-fate with our CDN.

2.3 Cloudflare R2 — the millions-of-devs-distribution incumbent

Value-chain position. Cloudflare's S3-compatible object store, sold to existing Workers/DNS/CDN customers on the same invoice. Cloudflare runs the bytes themselves. Critically: 16 first-party remote MCP servers, including https://r2.mcp.cloudflare.com/mcp with OAuth.

Switching cost. Low for the bytes (S3 API). High for the ecosystem — Workers bindings, D1 joins, Durable Objects, Workers AI — once a customer is invested in the broader Cloudflare graph, R2 is the storage tab they reach for first.

Wedge against R2: Burst throughput much higher than 100 PUT/s per logical store (we shard internally — invisible to the agent). Per-prefix scoped MCP tokens for least-privilege agent access. Bulk-list/prefix-walk operations priced flat or near-free to break the LIST-tax. Don't try to win Workers integrators — pick our SOM and stay in it.

2.4 Cross-incumbent patterns (the structural opening)

1. None of the three is purpose-built for agentic call patterns. Tigris monetizes per-request, Vercel monetizes per-byte-through-CDN, R2 monetizes per-Class-A-op. An LLM agent's natural rhythm — list, peek, fetch, fetch-again, write-small, write-small — punishes the customer on each model.
2. All three couple storage to a non-storage surface they cannot decouple. Tigris to Fly/Beam OEM contracts; Vercel Blob to the Vercel platform; R2 to the Workers fabric. None can offer "object storage portable across every agent runtime" without breaking the bundling story.
3. All three publish rate-limit or throughput ceilings we can publicly out-quote. Vercel: 120 ops/s on Pro. R2: 100 PUT/s per bucket. Tigris: per-request priced. agentik-host can publish "10× this for the agent tier" as a marketing line that's also a real architectural decision.
4. Compliance is structurally underweighted in all three. SOC2 Type II is the universal floor; HIPAA BAA + FedRAMP are spotty and entanglement-bound to the rest of the platform.

2.5 Honest counterpoints (where they beat us today)

• Durability claim parity. All three cite 11 9s. We can't match the narrative for at least a year of operational runway; audited 11 9s requires SOC2 Type II + multi-DC posture from day one.
• Free-tier economics. R2: 10 GB-mo + 10M Class B ops free. Tigris: 5 GB + 110K ops. Vercel Hobby: 5 GB + 100 GB transfer. We need at least equivalent, and unit economics need to survive.
• Distribution. Vercel ships Blob to every Next.js project. Cloudflare ships R2 binding to every Worker. We earn each install.
• Tigris's $25M Series A funds the exact roadmap we'd want to ship. The 12-month window is real but it is a window — Tigris will close A1 + some of A3 gaps in 2026.

3. Power-user pain — 5 unaddressed, with ≥3 independent voices each

3.1 Pain A — "MCP has no real file-upload primitive"

The most-cited pain in the dataset. MCP is JSON-RPC; there's no binary data channel. Everyone is building duct-tape staging servers, base64 hacks, or sidecar HTTP services.

• "MCP (Model Context Protocol) is JSON-RPC — it has no binary data channel … tools that accept a sourcePath parameter can only read files from the server's local filesystem. The client's files don't exist there." — danielrosehill, 2026-04-12 (repo)
• "The protocol itself has no file transfer mechanism … every MCP server that wraps an API requiring file uploads has this same limitation." — kenimo49, 2026-03-06 (dev.to)
• "We need an official way to be able to stream data to clients (ideally as binary), and not use JSON RPC as a wrapper." — SamMorrowDrums, MCP spec issue #527 (GitHub)
• "MCP today passes everything through JSON-RPC (base64 for binary) — doesn't scale for large payloads." — bhanquier, SEP-2433 PR

Why incumbents can't easily fix. Both Tigris and Vercel Blob route through their own SDKs that are also JSON-RPC-incompatible with the MCP transport (they use REST/PUT against custom endpoints). R2's MCP server is account-level OAuth — not the file-upload primitive Pain A is asking for.

agentik-host coverage. ✅ PRIMARY WEDGE. upload_file + get_file_url is the staging-server pattern, made invisible.

3.2 Pain B — "Vercel Blob is R2 + a $0.15/GB tax I didn't see coming"

• "Vercel is wrapping R2 and is charging $2 per million reads … they're also charging $0.15/GB for egress … That's quite the cost increase." — top HN comment, 2023-05-01 (news.ycombinator.com/item?id=35774730)
• "~2 TB transferred in a single day. $129, almost entirely from Vercel's 'Fast Data Transfer.'" — Josh Collinsworth, The Vibes Tax, 2026-02-17
• "My Vercel Blob storage got blocked after I accidentally downgraded from Pro to Hobby … 6000 images, system stopped working … support ticket opened, no response for 40+ hours." — Vercel Community, 2026-03-13

agentik-host coverage. ⚠️ PARTIAL. Flat $12/$39 tiers eliminate egress-surprise — but we are not making this legible in copy. Should be in the "What we offer" bullets.

3.3 Pain C — "Presigned URLs silently inherit role TTL"

The most universally lived-in S3 pain in the dataset. Everyone designs for 7-day URLs and gets 12-hour / 1-hour / 15-minute.

• "Wait fifteen minutes after running this example … your pre-signed URL [expires in] a measly fifteen minutes." — James Sherwood-Jones (jsherz.com)
• "What's the point of having a 7-day expiry on the pre-signed URL [if it expires in 12 hours]?" — AWS re:Post, 2024-08-05
• "You cannot safely build any system which relies on presigned URLs lasting longer than 10 minutes when generated on an EC2 instance." — DrewBarclay, botocore #2604

agentik-host coverage. ✅ Stateless JWT — never expires. Lean in: "The URL your agent gets back from upload_file doesn't expire in 15 minutes."

3.4 Pain D — "R2 rate limits + the r2.dev → custom-domain trap"

• "Managed public bucket access through an r2.dev subdomain is not intended for production usage and has a variable rate limit applied." — official R2 docs
• "So i need to buy domain, then setup cloudflare zone, then rent cloudflare worker, only after all that hidden cost i can finally make my bucket accessible from internet. I was guessing that free egress is too good to be true." — CF community, 2024-01-21
• "When you encounter an HTTP 5XX error, it is usually a sign that your Cloudflare R2 bucket has been overwhelmed by too many concurrent requests." — Cloudflare's own R2 troubleshooting docs (structural single-DO-per-bucket limit, documented)

agentik-host coverage. ✅ Public URLs work from day one, no domain purchase, no Worker rent.

3.5 Pain E — "OAuth on MCP is broken in practice; token-in-URL wins"

• "OAuth works in theory but breaks in practice. Token refresh is unreliable across Claude, ChatGPT, and VS Code." — Apigene, 50-thread analysis, 2026-04-12
• "The single biggest adoption barrier for MCP servers is connection setup … Token-in-URL is the sweet spot." — Graham Rowe, Phase Transitions, 2026-03-09
• "OAuth assumes a browser. It assumes redirect URIs … An MCP server running inside Claude Code has none of these things … I tried half a dozen approaches. Each one worked in isolation and broke when integrated with the MCP transport layer." — Task Board author, 2026
• "Zero signup. Just point your MCP client to the endpoint." — chen_yuan, dev.to, 2026-05-22
• "NO AUTHENTICATION or API KEY required? No. Not at all." — file.kiwi MCP server README

agentik-host coverage. ✅ PRIMARY WEDGE #2. The tweet-share free tier IS literally "no signup, no API key, paste this URL." Must be marketed AS a structural answer to Pain E, not as a giveaway.

3.6 Mismatch list — recurrent pains we don't yet address

3.7 Sentiment-by-cohort summary

• Vercel Blob power users: (1) Surprise egress bills via Fast Data / Fast Origin Transfer. (2) "It's just R2 with a markup, plus lock-in."
• R2 power users: (1) r2.dev throttling forces custom-domain purchase for prod. (2) Hot-object 5XXs + single-DO-per-bucket throughput ceiling.
• Tigris power users: (1) Reliability — 500 errors at Fly console; DNS/ISP edge cases. (2) Permission edge cases. Tigris owners are visibly responsive; this gap may close.
• MCP-native cohort (mcp.gd, Fastio, diskd, file.kiwi): (1) "No signup, no key" framing wins. (2) Tool surface either too thin (single-purpose) or too sprawling (Fastio's 251 tools). agentik-host's 2-tool surface is dead-center.

4. Synthesis — where SAM × incumbent vulnerability × unaddressed pain converges

This is where MECE earns its keep. Mapping the three streams onto a single matrix:

The 12-month strategic answer: double down on the token-in-URL + two-tool + permanent-URL wedge (already shipped), make Pain B / Pain C / Pain E legible in copy, and ship the screenshot-share workflow + direct-to-B2 PUT to close the highest-volume mismatch. Defer compliance (HIPAA BAA, FedRAMP) and bulk-tenant features until after $100K ARR + first enterprise inbound.

5. Economic moats — what we can hold and what we can't

Real (defensible):

1. Token-as-tenant cryptographic identity. Our JWT-as-tenant model is genuinely structural — the stateless property means our free-tier acquisition cost is zero (no DB row per signup). No incumbent can copy this without rewriting their auth stack.
2. Two-tool MCP surface as a context-budget play. Once an agent is using upload_file and get_file_url, the cost-to-switch is not the code (it's trivial); it's the 15 seconds of context budget the user has already amortized into their system prompts. As context budgets stay tight, two-tool surfaces compound.
3. Free tier with zero ongoing cost (no email-marketing-funnel obligation). Tigris, R2, Vercel all carry CRM overhead. Tweet-share doesn't.

Not real (don't bet on these):

1. Storage cost. B2 underneath; Tigris/R2 will match or beat us on per-GB economics.
2. Durability narrative. 11 9s is table stakes; we can't out-narrate AWS.
3. Distribution. We have none. Vercel/Cloudflare each have millions of devs already paying them. We have to earn each install.

Switching costs in our favor (real):

1. Each tenant has a bucketId baked into a JWT they've copy-pasted into N MCP clients. Asking them to migrate = re-pasting into N config files. That's painful — not the bytes.
2. Agents' generated artifacts have URLs already shared with downstream consumers (other agents, Slack, embedded in docs). Migrating means rewriting outbound URLs. Network-effect-flavored switching cost.

Switching costs against us:

1. We have no lock-in surface against an agent that hasn't sent us much data yet. The first-100MB customer can leave us trivially.
2. Our presigned-GET URL is currently the B2 endpoint directly — bypassing us. If we proxy through agentik-host.vercel.app/api/files/... instead, we gain a switching cost (the URL has our hostname) but lose Vercel-bandwidth economics.

6. Recommended 12-month strategic moves (ranked by leverage)

1. Make Pain E legibility the v0.5 narrative. "OAuth on MCP doesn't work. Token-in-URL wins. We made it the whole product." This is a TOFU-content angle and a landing-page H2.
2. Ship the screenshot-share workflow as a named primitive. "Your agent generated a screenshot. Now what?" — one-line MCP tool, one-line API, permanent URL. This is the highest-cite-volume agent-file scenario; nobody has explicitly named it.
3. Direct-to-B2 PUT for free tier (the deferred v0.2 phase-2). Closes the embarrassing 4 MB Vercel-Route-Handler ceiling, makes the "5 GB on Pro" claim real, lets the 50–150 MB screenshot/audio case work.
4. Publish throughput SLA per tier. Pro: 500 ops/s sustained, 2000 burst. Team: 2000 sustained, 10000 burst. Concrete numbers we can architect to (multi-pool internally), invisible to the agent. Public out-quotes Vercel 120 ops/s.
5. Lab/audit log export on Team tier. Closes the Fastio "files persist indefinitely + audit" claim. Required to unblock first enterprise inbound.
6. Acquire the r2.dev-fatigued cohort. Concrete CTA on landing: "R2 told you to buy a domain just to get a public URL? Skip that." Direct comparison page, named (PutPut already proved this works against Vercel).
7. Stop saying "Backblaze." (Already done in v0.4.)

Explicitly deferred (12+ months):

• HIPAA BAA / FedRAMP Moderate posture
• Multi-region prefix-isolation (Tigris's global routing thesis)
• Custom-domain support
• Resumable uploads (TUS) beyond a basic chunked-PUT pattern
• Anything that asks for a credit card before first upload

Citations

(All linked inline above; consolidated for verifiability.)

Stream 1 anchors: anthropic.com · venturebeat.com Anthropic ARR · sacra.com/c/cursor · windsurf.com · backlinko.com/chatgpt-stats · survey.stackoverflow.co/2025 · bloomberry.com 1400 MCP servers · culta.ai devtools benchmarks · blog.cloudflare.com/r2-ga · siliconangle.com Tigris Series A

Stream 2 anchors: tigrisdata.com/pricing · tigrisdata.com/docs/.../s3-compatibility · trust.tigrisdata.com · Vercel Blob 503 outage thread · Fast Origin Transfer thread · Vercel Blob pricing · putput.io compare · R2 open beta · R2 pricing

Stream 3 anchors: danielrosehill mcp-file-staging-service · MCP issue 527 · josh.ing The Vibes Tax · HN 35774730 · jsherz.com presigned URL gotcha · AWS re:Post 7-day expiry · Cloudflare community r2.dev rant · apigene.ai MCP auth analysis · phasetransitionsai MCP patterns · claude-code#23028 screenshot pain · claude-code#37418 session killer

Report generated by 3 parallel Exa-driven research streams + synthesis. Each headline number traces to a primary source. Constraints (no marketing-surface comparisons; strict focus on economic moats, value-chain position, switching cost) honoured throughout.