Building a web service no longer means you must provision a separate server or database first. Beyond website deployment, Cloudflare offers server-side execution, data storage, AI inference, object storage, and automation.
The interesting part is that some of these can be tried on a small scale with a free plan. The catch is that each service measures usage differently, and some features do not have a free allowance.
This guide describes the purposes and published free limits of selected services as of October 2026. These are historical figures from the official pricing pages, not a guarantee of what an account receives today. Check the latest pricing and your own account settings before using them.
1. Server code and data processing
| Service | Free allowance cited in October 2026 | Potential use |
|---|---|---|
| Workers | 100,000 requests per day | APIs and lightweight server logic |
| D1 | 5 million rows read and 100,000 rows written daily; 5 GB overall storage | Public and lightweight SQL data |
| Durable Objects | 100,000 requests and 13,000 GB-seconds per day | Connection and sequential state |
| KV | 100,000 reads and 1,000 writes per day; 1 GB storage | Settings and simple caching |
| Queues | 10,000 standard operations per day | Asynchronous jobs and retries |
Workers — execute code without managing a server
Workers runs code in response to requests without requiring you to maintain a traditional server. It suits lightweight APIs, server-side input validation, and processing public information.
The free plan also had a 10 ms CPU time limit per request. Long-running computation or processing large files calls for a different architecture.
D1 — SQL without a separate database server
D1 is a serverless relational database. Charges and limits are based largely on rows read and written, rather than only the number of queries. The published free tier also included 5 GB of account-level storage and a 500 MB limit per database.
It can be useful when experimenting with public market information or small amounts of service metadata.
Durable Objects — state that must remain consistent
Durable Objects can serialize operations and maintain state associated with a particular object. They are worth considering for device connection or message-delivery state where consistency matters. The free plan supported SQLite-backed Durable Objects.
KV and Queues — read-heavy settings and asynchronous work
KV is well suited to frequently read values that change relatively infrequently. It does not provide immediate global consistency after every update, so it should not be used indiscriminately for real-time control state.
Queues separates receiving a request from processing the work, allowing retries or deferred execution. The free allowance is not equal to a message count: writes, reads, and deletes can each consume operations. The free-tier maximum message retention period was 24 hours.
2. AI and semantic search
| Service | Free allowance cited in October 2026 | Possible experiment |
|---|---|---|
| Workers AI | 10,000 Neurons per day | Classifying public content and creating embeddings |
| AI Gateway | Core request analytics and caching available without extra charge | Managing model requests |
| Vectorize | 30 million queried dimensions and 5 million stored dimensions per month | Semantic search |
Workers AI — inference without operating GPU hardware
Workers AI provides serverless model inference. It is worth testing for small tasks such as classifying articles or generating embeddings.
10,000 Neurons per day does not mean 10,000 calls. Consumption varies by model and input. Some models or use cases may require a paid plan or billing setup.
AI Gateway — a management layer, not a model
AI Gateway helps manage requests to models: usage visibility, caching, and rate limiting are potential applications.
Even if Gateway's own features are free, the external AI provider may still charge for inference. Review logging and retention settings before routing prompts or responses that contain sensitive data.
Vectorize — search by meaning rather than exact keywords
Vectorize supports vector-similarity search. It might make sense once a public article collection becomes large enough to justify finding documents with related meanings. Its limits are calculated using dimensions queried and stored, not just request counts.
Embedding generation has separate usage costs. For a small content archive, ordinary search may already be sufficient.
3. File storage and browser automation
| Service | Free allowance cited in October 2026 | Potential use |
|---|---|---|
| R2 Standard | 10 GB-month storage; 1 million Class A and 10 million Class B operations per month | Public files and test assets |
| Browser Run | 10 minutes of browser time daily; 3 concurrent sessions | Rendered-page inspections |
| Turnstile | Up to 20 free widgets | Limiting automated form submissions |
R2 — object storage for files
R2 can store images, documents, and other objects. The storage and request allowances above applied to Standard Storage, not necessarily other storage classes. Network egress was free, but should be considered separately from other billing categories.
There is no reason to upload private shopping screenshots to R2 merely because storage is available.
Browser Run — inspect rendered pages
Browser Run can operate a browser in the cloud to inspect pages and rendered output. The free usage window was short, making small-scale visual checks a more suitable initial experiment than extensive browser automation.
When automatically accessing third-party sites, respect their terms of use and access policies.
Turnstile — an additional control for public inputs
Turnstile can help distinguish automated requests without relying on traditional puzzle-style challenges. It is worth considering for contact forms and other public input surfaces.
It does not replace authorization, authentication, or server-side validation.
4. Which ONSEMO projects might benefit first?
| Project | Experiment to consider | Question to ask |
|---|---|---|
| BLOG | Classifying public articles and semantic search | Is the archive large enough to need better search? |
| INVEST | Organizing public market data in D1 | Can transparent, rule-based processing handle the task? |
| CART | In-browser OCR | Can personal screenshots remain on the device? |
| CONNECT | Existing state management and optional Queues | Does this genuinely improve message delivery reliability? |
| WEB TOOLS | Browser Run for visual checks | Is small-scale automated verification useful? |
| HOME | Local-first features and optional connections | Are device and video data still protected? |
These are ideas under consideration, not a committed roadmap or completed features.
In particular, CART prioritizes local OCR and encrypted storage. CONNECT is not designed to send end-to-end encrypted message contents to server-side AI. INVEST can begin with public data collection and transparent rules without presuming an additional LLM.
5. Questions before trying a free service
Free availability is only part of the decision.
- Does the feature work on the plan I currently have?
- Are free limits measured per project or shared across the entire account?
- If a limit is exceeded, will requests fail or will I incur charges?
- Will data be sent to Cloudflare or an external model provider?
- Does the extra technology solve a real problem, or only increase operational complexity?
Being able to adopt a technology and having a good reason to adopt it are different things. Free features offer a useful way to test ideas on a small scale before making long-term architectural commitments.
I intend to make these choices only when a real need arises. Unlike an account of a development environment I actually built, the services in this article are candidates for future experiments.
---
Based on official public documentation as it stood on October 15, 2026. Free limits, available models, and billing requirements may change; check the linked official pages before implementing anything.