Infrastructure decision · July 2026

Do you need AWS? No. Not yet.

TachiBot is a local MCP server — users run it themselves via npx with their own keys. Months 1–2 of the plan need zero infrastructure. A cloud version only enters at Month 3, as the hosted-tier skeleton — and even then, AWS is the wrong first platform.

01 /

What actually needs a cloud

and what doesn't

Needs nothing: the open-source server (npx, local), directory listings, Sponsors/Polar, productized services, content. That's avenues 1, 2, 4, 5, 6 — the entire 90-day revenue plan runs on a landing page you already have.

Needs a cloud: only "TachiBot Cloud" (avenue 3) — the managed remote MCP server that holds keys and meters margin. And its Month-3 goal is a skeleton, not a launch: one small container, one database, a handful of test users.

Don't buy enterprise infrastructure for a proof-of-loop. A $10/mo box proves the same thing.

02 /

Platform pick

solo-dev reality

A Recommended

Cloudflare Workers

The purpose-built home for remote MCP in 2026: first-class remote-MCP support (McpAgent, built-in OAuth flows, Durable Objects for session state), generous free tier, zero server ops. Best match for "MCP server as a product".

$0–5 /mo · skeleton scale

B Solid alt

Fly.io / Railway / Render

If you'd rather keep the existing Node/FastMCP codebase untouched: one Docker container + managed Postgres, deploy from the repo in an afternoon. Slightly more ops than Workers, total code reuse.

$5–25 /mo · container + db

C Later

AWS (or GCP/Azure)

Justified only by: an enterprise client contractually requiring it, a compliance path (SOC 2), or real scale. Until then it buys you IAM, VPCs and ops burden — and nothing the skeleton needs.

defer · revisit at first enterprise deal

03 /

Hosted-skeleton steps

what you actually have to do

Step 1 · transport

Make TachiBot speak remote MCP

Switch stdio → streamable HTTP transport for a hosted mode (FastMCP supports httpStream) behind a flag, e.g. TACHIBOT_TRANSPORT=http.the only real code change — everything else is packaging

Step 2 · identity & keys

Auth + key custody

Per-user API tokens to reach your endpoint (OAuth later; Workers gives it near-free).who is calling
Server-side provider keys — your OpenRouter org key serves all users; users stop needing any keys. This is the whole product.kills BYOK friction — the reason hosted has margin

Step 3 · metering & safety

Count everything, cap everything

Real token capture per user per call from provider responses into SQLite/Postgres.same work as the usability cost-tracking feature — double payoff
Wire cost-monitor caps into the live path + per-user rate limits.an unmetered proxy with your org key is a wallet with no lid

Step 4 · money & proof

Bill and validate

Gate access behind the existing Polar sub; price = model pass-through + 10–30% margin (or flat tier with fair-use cap).reuses avenue-2 billing rail — no new Stripe work
Invite 3–5 test users, watch a month of real margin data, then decide on a public beta.the Month-3 decision gate from the 90-day plan

04 /

Budget reality

skeleton phase

$0–25/mo

total infra for the skeleton (Workers free tier or one small container + db)

1 flag

code delta: http transport behind an env flag — the repo stays one codebase

10–30%

margin layered on OpenRouter pass-through — the number the skeleton must prove

AWS services required before an enterprise client or compliance demand appears