AI agent with tools
An assistant that doesn't just chat: it looks things up, does the math and suggests your next step.
Live. Uses the AIM API.
What your client sees
Tell it what you need
Pick a goal, the way a client would write it. None of them says the price: the agent looks it up.
Under the hood
What it does, in technical terms
Tell it what you need and watch, step by step, how an AI agent plans, looks things up on the site with real tools, works out the estimate with the quote calculator's rules and suggests your next step.
- Technologies
- Workers AI
- Gemma
- Function calling
- Hono
- Zod
- SSE
- Backed by
Architecture
1.Your browser
It sends the goal and the currency, and draws each event validated with Zod: the plan, each tool with its arguments, its result and the answer.
2.The agent on Cloudflare (Hono)
A loop with caps on steps, tools, tokens and time. Anything that doesn't match the schema is refused as one more step, without breaking the run.
3.Gemma 4 on Workers AI
Native function calling, temperature 0.2 and 350 output tokens per call.
4.Read-only tools
search_site, get_service, estimate_quote and suggest_next_step, over the site's content. None of them writes or goes out to the network.
By the numbers
- calls to the model per run, at most
- 5
- tool calls per run, at most
- 6
- maximum time per run
- 45 s
- neurons in a typical run
- 36–76
The code
Decisions
- DEC-033: An AI agent with tools in the Lab (opens in a new tab)
- DEC-018: A free model for the AI Assistant (opens in a new tab)
Recorded in docs/DECISIONES.md (in Spanish), with their context and the alternatives ruled out.
Pick a goal and watch how it works.
Your reference estimate
The range shows up here once the agent does the math, with where each amount came from.
It's an AI and it can make mistakes: confirm prices and timelines with Alejandro.
Your goal is processed on Cloudflare Workers AI, with Google's open Gemma model, to plan and write the answer. This site doesn't store it; even so, don't write personal data. Read the privacy policy.
Would you rather talk to Alejandro? Message him on WhatsApp (opens in a new tab) or use the contact form.
Give the agent a goal
Write what you need the way you'd tell a person. The agent decides what to look up, uses the site's tools and shows you each step as it works.
Tools it can use
- Search the site: services, projects, Lab demos, posts and FAQ.
- Look up a service: what it includes, its reference timeline, its starting price and its extras.
- Work out the estimate: with the same code as the quote calculator, in the currency you chose on the site.
- Suggest the next step: the quote calculator with the service chosen, the form or WhatsApp.
Up to 5 steps and 6 tool calls per run. The tools only read this site's content.
It's an AI and it can make mistakes. Before you decide, confirm prices and timelines with Alejandro.
Your goal is processed on Cloudflare Workers AI, with Google's open Gemma model, to plan and write the answer. This site doesn't store it; even so, don't write personal data. Read the privacy policy.
Would you rather talk to Alejandro? Message him on WhatsApp (opens in a new tab) or use the contact form.
Want it for your business?
This is part of what I build in the chatbots service. See what's included or get a reference estimate for your case.
Curious how it is built? Look under the hood
How it works
It shows how I build AI agents that use real tools within clear limits: the model decides what to look up, and the code runs it, validates it and sets the bounds.
The open Gemma 4 model runs on Workers AI and calls the tools with native function calling. The tools run on AIM’s API on Cloudflare and only read this site’s content: they search the services, projects, Lab, blog and FAQ, show a service’s details, work out the estimate with the same code as the quote calculator and build the link for the next step. None of them writes anything or calls another site.
Each step arrives via Server-Sent Events: what the agent plans, each tool with its arguments, what it returned and the answer as it’s being written. A run has at most 5 steps, 6 tool calls and a time limit. If the model asks for a tool that doesn’t exist or passes arguments that don’t match its schema, the call is refused and you see it as one more step; the agent doesn’t crash.
Prices come only from the estimate tool. If the answer mentions an amount that no tool returned, the page warns you. Your goal travels as data, never as instructions, and the site doesn’t store it.
It shares the session and the limits with the AI Assistant and the AI code reviewer: you pass the security check once, and each run counts as two uses of the daily quota because it calls the model several times.