From apps to agents

Until now, the AI in “build an app with AI” happened at build time. Your assistant wrote the code, then stepped away, and the app that shipped was ordinary software — a form, a tracker, a dashboard. Useful, but not smart. That’s changed. Apps you build on Croft can now call an LLM themselves, while they run. The intelligence doesn’t leave when the assistant does — it lives inside the app.

Your apps can think now

You connect an AI model the same way you connect any other tool. A connector holds the key, your app calls the model through Croft’s gateway, and the key never touches the app’s code. The models you can reach: Anthropic, OpenAI, Google Gemini, Mistral, xAI (Grok), DeepSeek, Perplexity, and OpenRouter — one key, many models. You bring your own account; Croft never resells tokens or marks them up.

What that unlocks inside an app: classify an incoming message, summarise a document, pull fields out of a form, draft a reply, score a lead, answer a question over your own data. The app stops being a static form and starts making judgements.

Three ingredients, and you’ve got an agent

An “agent” isn’t magic — it’s software that can perceive, decide, and act on its own. Croft now gives you all three pieces:

1. An LLM to reason with — the new AI connectors. The app can think about what it’s seeing.
2. A schedule to run on its own — automations. “Every morning at 7.” “Every 15 minutes.” Nobody has to open the app.
3. A public webhook to react to the world — an endpoint the outside can POST to. A form is submitted, a payment lands, another system fires an event — and your app wakes up and handles it.

Put them together and you’ve built an agent: it runs when it should, looks at what came in, reasons about it with a model, and takes an action — writes to its database, calls a connector, sends an email, pings you.

What people build

• A triage agent. A webhook catches every inbound support email or form. An LLM reads it, judges the urgency, drafts a first reply, and routes it to the right person — flagging the ones it’s unsure about for a human.
• A morning briefing. On a schedule, the app pulls yesterday’s numbers through your connectors, asks a model to summarise what changed and why it matters, and emails the team before standup.
• A lead scorer. A form submission hits a webhook; the agent enriches it, scores fit with an LLM, drops the hot ones into your CRM connector, and messages the owner.
• A watcher. Every few minutes it checks a signal — a price, a status page, a queue — and only interrupts you, with a model-written explanation, when something actually needs you.

Agents you can put in a business

Here’s the part the standalone agent frameworks skip. An agent that can read your data and act on the world is exactly the thing you don’t want running unsupervised on someone’s laptop. On Croft, an agent inherits every guardrail the platform already has:

• The model key is held by the platform, not pasted into code — and a sensitive connector needs an admin’s approval, with an audit trail.
• It runs on your own private server, behind enterprise sign-on, with access granted app by app.
• Every change is versioned and reversible, and its data is backed up. An agent that goes sideways is a one-click rollback, not an incident.
• You bring your own model — swap Anthropic for Gemini, or a cheaper DeepSeek, without rewriting the app, and never pay a token markup.

That’s the whole Croft thesis, extended one step further: the speed of building with AI, with the guardrails a business needs — now for agents, not just apps.

Start with a prompt

Ask your assistant for one:

“Build me an app with a public webhook that takes a support email, uses Claude to classify it as billing, bug, or sales, drafts a reply, and emails the right person — plus a page where I can see everything it’s triaged.”

It writes it, deploys it to your croft, and wires up the connector. You’ve got an agent by lunch — private, logged, and yours.

See connectors →  ·  See automations →

More from the blog