Ava
Ava is the AI teammate inside your InfraVoice workspace. Tell it what you need in plain words — "create an appointment booking agent", "attach this tool, here's the curl", "put our number on Katy", "how many calls did my agents take this week?" — and it asks only what it can't infer, shows you a plan, builds on your approval, and puts every change in front of you before it goes live.
Everything on this page happens in the dashboard at ultravoice-dev.us.inc: Ava is the first screen after sign-in and sits at the top of the left rail. Driving it over HTTP is covered in the Ava API.
Nothing runs on its own. Ava proposes first — a plan you approve, then a list of changes you accept one by one. Undo is one click and every version is kept.
What you can ask
| Ask for | Example | What happens |
|---|---|---|
| Build an agent | "Create an appointment booking agent for Northside Dental" | A short interview, then a plan: steps, transitions, tools, the number. Approve it and the flow is written to the canvas, published and ready to call. |
| Edit an agent | "Add a callback branch when nobody picks up" | The change is proposed against the live flow as a reviewable list of edits — nothing moves until you accept. |
| Attach a tool from a curl | Paste the curl command | Method, URL, parameters and auth are read from the curl and turned into a tool on the agent. The token is handed to the platform encrypted and never appears in the chat. |
| Create or attach knowledge | "Create a knowledge base from our FAQ and attach it to Katy" | The knowledge base is created, filled and bound to the agent. |
| Bind a phone number | "Put +1 (415) 555-0182 on the booking agent" | The number is routed to the agent for inbound and outbound calls. |
| Staff a call center | "Set up a Support Desk with Ava and Kai as AI first responders and Maria for tier 2" | Project, AI agents, human agent and routing are set up together. |
| Ask about the workspace | "How many calls did my agents take this week?" | A plain answer. No plan, no changes. |
Type @ in the composer to point Ava at a specific agent or call.
How a turn works
| Stage | What you see | What you do |
|---|---|---|
| 1. Triage | Ava reads your message against the workspace and picks a path: build, edit, operate or answer. | Nothing — it takes a moment. |
| 2. Interview | One question at a time, with options, only for what it cannot infer. | Pick an option or type your own answer. |
| 3. Plan gate | The plan: steps, transitions, tools, knowledge and the values you still owe. The flow is drawn on the canvas next to the chat. | Approve to build, Decline to stop, or say what to change. You can also edit the flow on the canvas first. |
| 4. Build | The canvas fills in live as steps are written; tools and the number are wired; the flow is validated. | Watch. Anything the validator had to fix is listed under Fixed automatically. |
| 5. Review changes | Every proposed change, grouped by steps, transitions, tools and global settings, with before and after. | Accept or reject each change, or all of them. Anything you do not accept is not applied. |
| 6. Live | A short receipt of what was applied, readiness (providers, phone number, flow valid) and follow-up suggestions. | Try the agent, keep going with a follow-up, or Undo. |
Interview
Ava never asks for what it can read from the workspace. When it does ask, each question comes with options and a free-text choice. You can answer several questions at once by typing the answers together — it matches them to the open questions. Answering a question a second time edits your earlier answer.
Plan gate
The plan is one document: the flow design, the agent settings, tools, variables and knowledge, plus the open items you still need to supply. Three ways to respond:
- Approve — builds exactly what the plan says.
- Decline — stops the turn. The chat stays open; say what you would like instead.
- Say what to change — any free text at the plan gate is treated as a revision request, never as approval. Ava rewrites the plan and shows it again.
You can also Download plan.md to keep or share the plan as Markdown.
Review changes
Changes are shown in groups: steps, transitions, tools and global settings such as the off-topic handler. Each one shows what it was and what it becomes. Accept them one by one or all at once. A change you leave untouched counts as rejected, so a partial review is always safe.
The canvas
The plan opens in the Flow Builder beside the chat. Edits you make on the canvas before approving are kept as a plan draft and merged into the build when you approve; editing after the build works the same way as any other edit and comes back as a changeset.
Every build passes through a validation layer that mirrors the Flow Builder's own checks, so the canvas never opens with validation errors. What it fixes on its own:
| Problem | Fix |
|---|---|
| A handler step (off-topic, objections, callbacks) that nothing leads to | Promoted to a global handler with a trigger condition |
| An ordinary step that nothing leads to | Wired into the nearest branch that still has a free exit |
| A step with no way out | Closed with an end step |
| A transition to a step that no longer exists | Dropped |
| A question or message step with no text | Filled from the step's label |
| No start step | The welcome step becomes the start |
Fixes are listed in the review so you can see exactly what changed. If something cannot be fixed safely, Ava asks instead of guessing.
Undo and versions
Every applied changeset creates a new version of the agent. Undo takes the agent back to the previous version in one click; the versions list lets you pick any earlier one. If a build applies only partially — for example one tool write fails — the receipt says what landed and what did not, and undo is still available.
Follow-ups
After every reply Ava adds a few follow-up suggestions marked with ↳ — the next sensible steps given what just happened, such as attaching the knowledge base you just created to an agent, or binding a number to the agent you just built. Click one to send it.
Limits and credits
- Turns. Each chat has 30 turns. Sending a message spends one; answering questions, approving plans and reviewing changes do not. A workspace that runs Ava on its own model key (AI Providers → credential type
voika) is not charged per turn. - Chats. A workspace keeps its 10 most recent chats; opening an eleventh removes the least recently used one.
- One turn at a time. A chat runs a single turn at once. A new message while a turn is running is refused until it finishes.
- Message length. Up to 8,000 characters — long enough to paste a full curl command or a page of FAQ text.
- Secrets. A token pasted inside a curl is kept out of the transcript, the plan and the review, and is passed to the platform encrypted when the tool is created.
Tips. Name the business, the caller's goal and the outcome — "book, reschedule or cancel a cleaning" beats "handle appointments". Paste real curls for tools rather than describing them. Use
@to target an agent. At the plan gate, say what to change in one sentence instead of declining and starting over.
Drive it over HTTP
Everything above is also available as an API: open a chat, send a command, read the turn as a stream, answer the questions, approve the plan and the changeset, undo. See the Ava API.