AI tools
AI tool blocks are the parts of an AI feature that sit around the answer: the web pages it read, the tools it called, the context it used, the memory it kept and what the reply cost. Orvain's AI tools family holds fourteen, from a web search answer with numbered sources, tool calls that show what each app was sent and returned, a code review taken hunk by hunk and a document read into fields, through image generation and a builder canvas, to a context window, a memory panel, model usage with its receipt, answer feedback, run controls, a smart recommendation and a voice agent on a live call.
What these have in common is that each one makes a hidden step inspectable. A model that called Stripe twice, read six sites or forgot a preference is doing something the person on the other side has a right to see and correct. The blocks show the step at the size it deserves, a receipt that can open into detail, and the control that changes it in place, so checking the work never means leaving it.
14 blocks
Using AI tools
Reach for it when
- An assistant that acts through other apps, where each call and its result should be checkable.
- Answers built from the web or from documents, where the sources decide whether the answer can be trusted.
- Generation people refine rather than accept whole: a draft, an image, a page, a code change.
- Usage that costs money or room, such as tokens, context or a budget, and that people choose how to spend.
Reach for something else when
- A plain conversation with no tools or sources. The assistant thread in the AI Chat family is lighter and says everything there is.
- A multi-step agent run with approvals and checkpoints. The Agents family is built around the run rather than one reply.
- A form someone fills in by hand. A form layout is faster than asking a model to read what the person could have typed.
Where it shows up
Refund assistant
A support assistant that reads a ticket, finds the double charge in Stripe and refunds it. Each tool call is a row with its receipt, the refund waits for an OK, and a busy Stripe is retried in place rather than failing the answer.
Invoice intake
Bills dropped into an inbox are read into the fields an accounting app needs. Each value is marked on its page, a date the page gives twice is asked about instead of guessed, and the bill is added with a receipt and Undo.
Questions
How should an AI answer show the tools it used?
As a short list of calls above the answer, one row each, with the app, what the call did and a receipt such as an ID or amount. Each row opens to what was sent and what came back, as fields or raw JSON. A call that changes something, like a refund, should stop and ask before it runs, and a retried call should say how many tries it took.
How do you show how full a model's context window is?
As a share of the window with the parts that fill it: the system prompt, tools, memory, files and the turns of the conversation, each with its size. People can then remove a file or summarise older turns and see the room come back. A single percentage with no breakdown tells them they are running out but not what to give up.
Premium components,
free in your inbox.
Join the Orvain Dispatch and get production-ready components, free to keep: React and Vue code plus the Figma frame, yours to ship.
Orvain MCPAI tools
AI tool blocks are the parts of an AI feature that sit around the answer: the web pages it read, the tools it called, the context it used, the memory it kept and what the reply cost. Orvain's AI tools family holds fourteen, from a web search answer with numbered sources, tool calls that show what each app was sent and returned, a code review taken hunk by hunk and a document read into fields, through image generation and a builder canvas, to a context window, a memory panel, model usage with its receipt, answer feedback, run controls, a smart recommendation and a voice agent on a live call.
What these have in common is that each one makes a hidden step inspectable. A model that called Stripe twice, read six sites or forgot a preference is doing something the person on the other side has a right to see and correct. The blocks show the step at the size it deserves, a receipt that can open into detail, and the control that changes it in place, so checking the work never means leaving it.
14 blocks
Using AI tools
Reach for it when
- An assistant that acts through other apps, where each call and its result should be checkable.
- Answers built from the web or from documents, where the sources decide whether the answer can be trusted.
- Generation people refine rather than accept whole: a draft, an image, a page, a code change.
- Usage that costs money or room, such as tokens, context or a budget, and that people choose how to spend.
Reach for something else when
- A plain conversation with no tools or sources. The assistant thread in the AI Chat family is lighter and says everything there is.
- A multi-step agent run with approvals and checkpoints. The Agents family is built around the run rather than one reply.
- A form someone fills in by hand. A form layout is faster than asking a model to read what the person could have typed.
Where it shows up
Refund assistant
A support assistant that reads a ticket, finds the double charge in Stripe and refunds it. Each tool call is a row with its receipt, the refund waits for an OK, and a busy Stripe is retried in place rather than failing the answer.
Invoice intake
Bills dropped into an inbox are read into the fields an accounting app needs. Each value is marked on its page, a date the page gives twice is asked about instead of guessed, and the bill is added with a receipt and Undo.
Questions
How should an AI answer show the tools it used?
As a short list of calls above the answer, one row each, with the app, what the call did and a receipt such as an ID or amount. Each row opens to what was sent and what came back, as fields or raw JSON. A call that changes something, like a refund, should stop and ask before it runs, and a retried call should say how many tries it took.
How do you show how full a model's context window is?
As a share of the window with the parts that fill it: the system prompt, tools, memory, files and the turns of the conversation, each with its size. People can then remove a file or summarise older turns and see the room come back. A single percentage with no breakdown tells them they are running out but not what to give up.
Related blocks
All blocksPremium components, free in your inbox.
Join the Orvain Dispatch and get production-ready components, free to keep: React and Vue code plus the Figma frame, yours to ship.
