Inbox Zero Self-Hosting Cost: What Free Really Costs

Inbox Zero is open source and MIT-licensed, so the license is genuinely free. The bill is not. Once you count a VPS, a Postgres database, TLS, monitoring, and the LLM API calls the assistant makes on every message, the self-hosted path has a real monthly cost that most 'free forever' write-ups quietly skip.

This page prices the self-hosted path honestly against the hosted plan at getinboxzero.com and against a managed alternative. Every number is a worked example you can adjust: change the mailbox volume, swap the model, pick a bigger VPS, and the math updates in your head. No competitor prices are quoted for other tools because packaging shape changes often enough that a cached number is stale within a quarter — verify current numbers on the vendor's own page the day you commit.

We build one of the options in this comparison — AI Emaily — and we are not self-hostable. That means on the specific 'runs on your server' axis we lose to the open-source Inbox Zero app outright, and we say so plainly rather than sneaking a hosted product into a self-hosting post as if it belonged there. Where we appear below is as the honest hosted answer for the reader whose real requirement is 'no vendor retains our mail' rather than literally 'code and data on our hardware.'

The short answer#

Assume a single-mailbox deployment of the open-source Inbox Zero app on a small VPS, using a bring-your-own-key model at commodity rates. As of August 2026 you should budget in the following ranges. Verify current numbers on the vendor and provider pages before you commit — this category moves.

Criteria that actually matter#

Every 'is self-hosting worth it' post lists infrastructure. The rows that decide the answer are the ones most write-ups skip: the LLM API bill, the recurring hours, and whether the project is still actively maintained six months from now.

We are pricing the honest total cost of ownership, not the cheapest possible line item. A $5 VPS with a saturated CPU is a false economy the first time your IMAP sync stalls; a free Postgres instance with no backups is a real liability the first time it corrupts. Budget to the middle column.

• Compute — the app plus Postgres plus a reverse proxy on a VPS that will not thrash
• LLM API spend — the meter that never stops, priced per active mailbox not per user
• Storage — Postgres for state, object storage for backups, both grow with mail volume
• Ops time — updates, TLS renewal, dependency bumps, credential rotation, incident response
• Deliverability edge cases — the app does not send mail itself, but Gmail OAuth quotas and Google Cloud project setup still cost time
• Project health — last commit date, issue-tracker responsiveness, security-advisory feed

Scoring table: self-host vs hosted plan vs managed alternative#

Three paths, priced on the same axes. Ranges reflect real variance across VPS providers, model choices and team sizes; the hosted plans have their own tiers that you should read on the vendor's page rather than trusting a number cached here.

A concept picture of the two paths#

Before the worked example, it helps to hold the shape of the choice in one image. You are weighing a bill you pay in dollars against a bill you pay in hours, and both bills recur every month.

Worked example: one power user, one modest team#

Two profiles. The numbers use a small VPS, Postgres co-located on the same box, S3-compatible backups, Let's Encrypt for TLS, and OpenRouter-priced or direct-provider LLM calls at commodity rates. Round up rather than down; running the box at 95% capacity is how a Sunday-morning incident happens.

Where the LLM bill actually comes from#

The bill is not one call per email. An AI triage-and-drafting app makes multiple model calls per message — classify, summarize, decide if a reply is needed, generate the draft, sometimes re-check against user rules — and each call spends both input tokens (the message plus context) and output tokens (the answer).

For a mailbox that receives 100 substantive messages a day and drafts replies on a quarter of them, expect somewhere in the low tens of dollars a month at mid-tier model prices. Double the volume or move to a top-tier model on every call and it climbs fast. Halve the volume and use a small model for triage, top-tier only for drafts, and it drops fast.

That variance is the whole reason 'what does it cost' has no single number. The rate card is public; your mailbox volume and the model you pick set the multiplier.

1. 1

   Estimate volume honestly

   Count substantive incoming messages, not the newsletters you archive without reading. Multiply by working days.

2. 2

   Split by task

   Triage every message. Draft only a fraction. Summarize on demand. Assign a model tier to each task rather than one model to all.

3. 3

   Look up per-million-token prices

   OpenRouter's model list is the fastest way to compare across providers on one page. Direct provider pricing sometimes beats it for a specific model.

4. 4

   Multiply and add safety margin

   Estimate tokens per call, multiply by calls per day, multiply by 22 working days, add 30% for context growth and retries.

5. 5

   Set a hard monthly cap on the API account

   Every provider supports usage caps or alerts. Set them before your first deployment, not after your first surprise bill.

Red flags: line items that quietly blow the budget#

None of these are theoretical. Each has burned real teams running open-source AI apps in the last twelve months. Watch for them before they arrive.

• Runaway retries — a rule that fires every minute against every message multiplies your API bill by a factor you did not plan for
• Larger model creep — 'let's try the big one for a week' becomes the default and quadruples the monthly spend
• Backups that were never tested — the first restore attempt is not the time to discover the pg_dump was truncated
• Dependency drift — a Postgres major-version bump or a Node LTS transition costs half a day and a maintenance window
• Gmail OAuth reverification — Google's app-verification cycle adds real hours, especially if you request restricted scopes
• Abandoned upstream — an open-source project with no commits for six months becomes a security liability the moment its dependencies age
• Model provider rate hike — public rate cards change; a mid-tier model rebranded and repriced is a real event you plan around

What we'd pick and why (honest)#

The choice usually collapses to one question first: is 'runs on our hardware' a hard requirement, or a proxy for 'no vendor retains our mail'? Answer that honestly before you shortlist anything, because the two produce different picks and getting it wrong wastes a quarter of ops time.

If it is a hard requirement — a signed customer contract, a regulator, an internal policy that explicitly forbids third-party processors — the honest pick is the open-source Inbox Zero app itself, self-hosted. You will pay the VPS, the Postgres, the LLM API bill and the ops hours documented above. In exchange you get code and model keys on hardware you own, mail that stays at Gmail via OAuth, and no vendor middleman. Inbox Zero the app has better Gmail-specific bulk-unsubscribe than most alternatives on this list; on that dimension it is the tool to beat, and we are not trying to.

If self-hosting is a proxy for 'no vendor retains our mail long-term' — you want no training on your mail, zero retention with model providers, and the option to route model calls through your own account — a managed assistant with those terms and BYOK support is usually the better trade. You spend the ops hours on the actual business instead of on patching a mail server, and the privacy guarantees you actually cared about are still in place.

That is where AI Emaily fits. We are a hosted AI email assistant across Gmail, Outlook and IMAP, with Copilot mode that drafts and files but asks before sending, gated Autopilot for actions you approved in advance, and an undo-able audit log so no automated action disappears silently. We do not train on user mail, we run through providers under zero-retention terms, and BYOK for OpenAI or Anthropic is supported if you would rather route model calls through your own account. Packaging is a 7-day full-access trial (card required, $0 if you cancel before day 7), then paid plans — no permanent free tier. We build AI Emaily. If your requirement is literally 'code runs on our hardware,' none of that helps and the open-source Inbox Zero app is the right answer instead.

If you are somewhere in between — you like the idea of self-hosting but the ops hours in the worked example above make you flinch — the hosted plan at getinboxzero.com is designed for exactly that reader. Same product, someone else runs it. Compare the current pricing on their page against the fully-loaded cost we sketched above and pick the cheaper column for your volume.

One last check before you deploy#

Whichever direction you go, treat the vendor's or project's own live page as the source of truth on the day you commit. Every number here — VPS pricing, LLM per-token rates, hosted-plan packaging shape — was written in August 2026, and any one of them can move between then and the day you read this. Open the pricing pages linked below, plug your real volume into the worked example, and let today's numbers win.

For the self-hosted path, the Inbox Zero GitHub repo has the deployment docs, and OpenRouter's model list is the fastest way to compare per-token pricing across providers on one page. For the hosted-but-zero-retention shape, our pricing page and the AI email assistant feature page linked below are the primary source for what we do and don't retain.