Every edge function is a
complete backend.

Aerostack edge functions are TypeScript serverless functions that run on Cloudflare's global edge network. What makes them fullstack is that every binding you'd normally wire up separately — a relational database (D1/SQLite), a key-value cache (KV), object storage (R2), queues, vector search, Durable Objects, and Cloudflare's AI runtime — is injected at runtime with no connection strings, no environment wrangling, and no separate provisioning step. You write a TypeScript handler, call `aerostack deploy`, and in roughly 30 seconds the function is live worldwide. There's no server to manage, no warm-up penalty, and no cold start. Most edge platforms give you compute; Aerostack gives you compute plus the full data layer already attached.

The gaps are structural, not just cosmetic. AWS Lambda runs in a handful of regions; Aerostack edge functions run at Cloudflare's 300+ global locations, so a request is handled close to the user without any routing configuration on your part. Lambda has a well-documented cold-start problem that gets worse with VPC attachments; Aerostack functions have zero cold start by design. The deeper difference is the data layer: with Lambda you configure a VPC, manage connection pools, add a separate cache layer, and wire credentials for each service. With Aerostack, database, KV cache, object storage, queues, and AI are all native bindings — available as typed objects in your handler with nothing to configure. For teams building AI-adjacent or latency-sensitive workloads, that eliminates an entire class of infrastructure toil.

Every function has access to the full Cloudflare primitive stack as native bindings injected at runtime: D1 for relational SQL (SQLite-compatible), KV for low-latency key-value reads, R2 for object storage with no egress fees, queues for async fan-out and background jobs, vector search for semantic retrieval and embedding lookups, the Cloudflare AI runtime for on-edge inference, and Durable Objects for strongly-consistent coordinated state. None of these require connection strings or credentials in your code — they're available as typed objects in the function context from the moment the handler starts. You pick which bindings a function uses when you set it up, and only those are attached, keeping the surface area minimal.

Any edge function can be exposed as an MCP tool, which means any AI agent connected to your Aerostack workspace can call it by name — no HTTP URL, no authentication plumbing, no prompt-engineering to describe the endpoint. The agent sees the function as a named capability with a typed schema, calls it like any other tool, and gets a structured result back. This is what makes edge functions a first-class part of the agentic layer: your business logic, database queries, or external API wrappers become agent-callable tools without any extra integration work. Functions can also back skills — reusable, parameterized tools shared across agents — and can be called from within workflow nodes, so the same function serves both interactive agent sessions and automated pipelines.

Deploy the function, then link it to an MCP server in your workspace — once linked, the function shows up as a named tool with a typed schema that any connected AI agent can call directly. There is no separate HTTP integration to build and no endpoint description to prompt-engineer: the agent sees the function as a first-class capability and invokes it like any other tool. This is the part incumbent edge-function platforms like Supabase and Netlify do not do. They hand you an HTTP endpoint; Aerostack turns the same TypeScript logic into an agent-callable tool. The same function can also back a reusable skill and run inside a workflow node, so one piece of business logic serves direct API calls, agent sessions, and automated pipelines.

Yes. There are two ways to deploy. The first deploys the function on its own and returns a public Cloudflare Worker URL in roughly 30 seconds — no auth, ideal for prototypes and internal tools. The second deploys the function with a gateway in one click: it auto-generates a consumer API key, applies rate limiting, sets up a starter usage plan, and turns on request analytics and logs. You do not write any auth or rate-limit middleware yourself — the gateway handles key validation, throttling, and usage tracking in front of the function. You can start with a raw function and add a gateway later when you are ready to share it more widely or put guardrails around it.

The free tier includes access to all edge function capabilities — compute, bindings, and the AI runtime — plus 500,000 AI inference tokens per month. That covers a meaningful volume of production workloads: semantic search, on-edge classification, generation tasks, or any combination, all without a credit card. The compute itself runs on Cloudflare Workers, which has its own generous free execution tier. When you bring your own model key (OpenAI, Anthropic, or any provider you configure), the platform markup drops to zero and the token budget is effectively your provider account's limit. There's no cold-start tax on the free tier — function latency is identical regardless of plan.

Write your TypeScript handler, configure your bindings in the function settings, and run `aerostack deploy` from the project directory. The CLI builds, validates, and propagates the function across Cloudflare's network — the whole process takes roughly 30 seconds from command to globally live. There's no region selection, no multi-region deployment config, and no DNS setup; Cloudflare's anycast routing ensures every request hits the nearest available location automatically. Subsequent deploys follow the same path: change the code, run the command, and the updated version is live worldwide in the same window. Rollbacks work the same way — redeploy the previous version and the change propagates in about 30 seconds.

Yes — edge functions are one of the primary building blocks for the rest of the Aerostack platform. As an MCP tool, a function is callable from any MCP tool node inside a workflow, so a multi-step agentic automation can read from a database, call an external API, or run an inference step just by calling the function by name at the relevant node. Skills are parameterized, reusable tool definitions that also back edge functions — you build the logic once as a function and surface it as a named skill available across every agent and bot in the workspace. This means the same TypeScript business logic that handles a direct API call is also the engine behind a skill a conversational agent invokes and a step inside a published workflow.

An edge function is the right choice when the logic is self-contained, deterministic, and fast: a database read, a transform, a webhook handler, a short AI inference call. It is not the right choice when the task involves multi-step reasoning where the path depends on model output, or when you need a human-approval pause, conditional branching across many tools, or state held across minutes or hours. Those are workflow problems, not function problems. Similarly, if you're integrating a third-party SaaS and the integration involves multiple API surfaces with complex auth, an MCP server keeps that complexity in one place rather than spreading it across several functions. The rule of thumb: if you can describe the logic in a single sentence with no 'it depends', use a function; if the logic has branches the model decides, use a workflow.

Cloudflare edge functions — also called Cloudflare Workers — are serverless TypeScript/JavaScript functions that execute at Cloudflare's network of 300+ global locations rather than in a fixed region. This means the function runs physically close to whoever makes the request, cutting round-trip time without any routing configuration on your part. Aerostack builds on top of Cloudflare Workers and extends them with a managed primitive layer: instead of provisioning a D1 database, a KV namespace, an R2 bucket, and a Vectorize index separately and wiring them all up, Aerostack injects all 8 bindings — D1, KV, R2, queues, AI runtime, vector search, Durable Objects, and Hyperdrive — into your function at runtime, automatically. You declare which bindings a function uses, call aerostack deploy, and the function is live at every Cloudflare edge location in roughly 30 seconds. The result is a Cloudflare edge function with a complete data layer already attached — not a bare worker you still need to configure.

A regular serverless function handles HTTP requests, runs business logic, and optionally calls external APIs. An AI function does all of that but also has AI inference, embeddings, and vector search built in as native bindings — not external HTTP calls to a provider. With Aerostack edge functions, env.AI gives you on-edge inference via the Cloudflare AI runtime, env.VECTORIZE gives you semantic search over your own embeddings, and env.DB lets you store and retrieve structured data alongside vector results — all in one TypeScript handler, all without managing API keys for a separate AI provider. The practical difference: a regular serverless function calling OpenAI adds 200–800ms of API latency per inference call. An Aerostack AI function runs inference at the same edge location as the function itself, with no external round-trip. That matters for real-time use cases like streaming chat, semantic search, and on-demand classification.