Connect your AI assistant through MCP and you can manage your voice agents by talking to it: read a prompt, change one section, check what is really running, and go back to an earlier version.
Every change is previewed first and saved only after you say yes.
Your assistant acts as you, with your permissions, and a saved change affects real calls. Read the preview before you approve it. Try changes on a test agent before you change an agent that is taking calls.
🔄 How a change works#
- Read. Your assistant reads only what it needs: one prompt section or one group of settings.
- Preview. It checks the change and shows you the exact differences. Nothing is saved.
- Approve. You say yes. Your account is recorded as the approver.
- Save. AI Sync saves the change, then reads it back from the live agent.
- Verify. The assistant tells you what the agent is running and whether calls are confirmed to use it.
- Restore. If you do not like the result, go back to any earlier revision.
💬 What to ask#
| You want to | Ask your assistant |
|---|---|
| Find an agent | "Find my agent named Grace" |
| Read the whole prompt | "Show me the complete active prompt for Grace" |
| Read one section | "Show me Grace's greeting phrases" |
| See settings | "What are Grace's interruption settings?" |
| Change one section | "Change Grace's call flow to this, and show me the preview first" |
| Change one sentence | "In Grace's prompt, replace this sentence with that one" |
| Change a setting | "Preview setting interruption sensitivity to 0.5 on Grace" |
| Save without going live | "Save that as a draft, don't put it live yet" |
| Check what is running | "Verify Grace's configuration" |
| See history | "Show Grace's revision history" and "Compare the last two revisions" |
| Go back | "Preview restoring Grace to the revision from yesterday" |
| Trace a bad call | "Which revision was Grace on for this call?" |
| Test | "Start a browser test session with Grace" |
✏️ What gets changed, and what does not#
- Only what you ask for changes. Anything left out of a change stays as it is.
- Emptying something has to be explicit. Your assistant must name the section or field to clear. A blank value is refused, so nothing is emptied by accident.
- One agent at a time. A change never spreads to other agents. To change several, preview each one.
- The automatic text in a prompt stays. AI Sync adds its own blocks to a prompt, such as the call direction and caller memory. Section edits leave them alone.
A prompt written as one block of text counts as one section. To change part of it, ask your assistant to replace an exact passage. The passage must appear exactly once, or the preview is refused and your assistant will ask for a longer one.
📄 Drafts#
A draft is stored only. Callers do not hear a draft. Putting a draft live is a separate step with its own preview and approval.
If the agent changes after you save a draft, the draft can no longer be activated, because that could undo the newer change. Preview the edit again on the current version.
✅ How to read the result#
"Saved" does not always mean callers have it. Read these words carefully:
| Result | What it means | What to do |
|---|---|---|
| Saved, read back, and the platform now holds this version | V2: the live agent has this configuration. | Run a test call to hear it. |
| Pending confirmation | V1: saved to the latest version, but a test call is still needed to confirm callers get it. | Test before relying on it. |
| Forwarded | V1 agent replaced by a V2 copy. Callers reach the V2 agent, so they will not get this change. | Edit the V2 agent named in the result. |
| Draft | Stored only. Not live. | Activate it when ready. |
| Not saved | The save failed. The result says whether the previous configuration was confirmed intact. | Read the reason, fix it, preview again. |
| Agent changed after this preview | Someone else edited the agent, so nothing was saved. | Preview again on the current version. |
| Changed outside this flow | Live no longer matches the last revision. Someone edited another way, such as in the agent screen. | Read the differences. |
A failed save keeps your previous configuration whenever it can. If it cannot confirm that, the message tells you to check the agent before the next call.
🕘 History and restore#
Every approved change creates a revision that cannot be edited or deleted. It holds:
- The complete prompt and settings
- The revision it came from, and the exact differences
- Who asked, who approved, and why
- The time in UTC, plus validation and verification results
Restoring adds a new revision that points to the old one, so history is never erased. Prompt changes also appear in the agent screen's existing prompt history.
A restore does not change phone number assignments, billing or folder placement.
Changes made elsewhere#
Changes made in the agent screen, through older tools, or to numbers, senders, pronunciation, knowledge bases, API functions or the widget do not go through this flow. If they change the agent's configuration, the next preview records them as an observed revision, so you can still restore from that point. Who made an observed change is not known.
Tracing a call#
Ask which revision a past call ran on. The answer comes from timestamps: it is the newest tracked revision before that call. An edit made outside this flow after that revision would not show.
🔒 Who can do what#
- You only ever see agents your account and role can access. Someone else's agent gets the same "not found" as one that does not exist.
- Your package decides whether you can use V1, V2 or both.
- Keys, tokens and webhook headers show as
[redacted]. They are stored encrypted so a restore can bring them back. - Prompts and call transcripts are treated as information, never as instructions to your assistant.
AI Sync records who approved a change, but it cannot prove a person said yes in the chat. Make sure your assistant shows you each preview and waits for your answer. Automatic approvals are not available.
🧪 Test sessions#
A test session is for a browser. It never dials a phone number, so it cannot reach a customer. Your assistant gets the connection details, which a browser client uses to talk to the agent; a chat alone cannot hold the conversation. To hear an agent yourself, use its test call.
📚 More#
- Voice agent settings reference: every setting, with units, ranges and V1/V2 differences
- Agent configuration API: the routes behind these tools
- Connect MCP