The agent creator, end to end
Mrs. NetShow is an AI agent.
From nothing to a live agent: create it in the Studio or over MCP, give it a look, a voice and a brain, publish it and put it on a site.
This page follows one agent from nothing to a live, talking agent on a website: how it is created, what it saves, how it gets a look, a voice and a brain, how it is published, and where strangers and other agents meet it. Every NetShow agent says it is an AI, from its first greeting.
Two ways in
-
In the Studio (no code). An owner opens
GET /studioand follows the steps. Whereai.features.studio_three_step_startis on,GET /studio/startturns it into three steps on one page: name it, give it a face and a voice, go live. Explore without signing in; signing in is needed to save. -
From your own tools (MCP). A signed-in developer calls the
agent_createtool on the MCP server. See MCP Developer Guide for the token.
Step 1: create the agent
In the Studio, the creator form posts to POST /dashboard/agents/store (signed in, same-site form). It needs a name and a description, and can carry a personality, a voice, starting capabilities, a template and a brain choice. A faster path, GET /studio/hire/{role}, starts from a ready role template.
From MCP, agent_create needs a name and a system_prompt. An agent made this way is always private; publish it from its settings page.
{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "agent_create",
"arguments": {
"name": "Front Desk",
"description": "Greets visitors and books appointments.",
"system_prompt": "You are Front Desk, an AI assistant for a dental office. Say you are an AI. Help visitors book a visit.",
"voice_enabled": true
}
}
}
curl --max-time 30 \
https://app.netshow.ai/api/mcp \
-H 'Accept: application/json' \
-H 'Authorization: Bearer <YOUR_TOKEN>' \
-H 'Content-Type: application/json' \
--data '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"agent_create","arguments":{"name":"Front Desk","description":"Greets visitors and books appointments.","system_prompt":"You are Front Desk, an AI assistant. Say you are an AI.","voice_enabled":true}}}'
Step 2: give it a look
Every default look is a character, never a photo face: a live ASCII face when tiny, a sketch, a toon, a simple orb with a little mouth, or a 2D or 3D character. The look changes with the space it has. A human-like face is an opt-in option, not the default. The owner picks the look in the Studio; the kit's look names are in the live widget guide.
Step 3: give it a voice and a brain
- Voice. Voices come through the NetShow proxy: marin for Mrs. NetShow, cedar for Mr. NetShow, or another voice the owner picks. The order NetShow tries and the captions fallback are in the live widget guide. Browser voices and recorded clips are never used.
-
Brain. The agent answers on a model lane from the model registry (for example
chat.default), not on a fixed model name, so a model upgrade is one registry change. The owner can also send work to a harness; see The harness router. - Knowledge. The owner adds facts the agent should know; see Knowledge & Memory.
The owner edits everything later on one page, GET /dashboard/agents/{id}/unified.
Step 4: try it
GET /dashboard/agents/{agentId}/test-drive lets the owner talk to the agent before anyone else can. Strangers can meet a published Studio agent at GET /studio/agents/{slug}.
Step 5: publish it
A new agent is private. Publishing flips one thing, its public switch, through POST /dashboard/agents/{agentId}/publicity; it never changes the agent's address or puts it in the marketplace. Where the three-step start is on, its Go live button does the same through POST /studio/start/go-live/{agentId}.
Step 6: where it lives
Once public, the agent has doors:
-
GET /agent/{slug}: its public page, where it greets first. -
GET /agent/{slug}/actions.json: what it can do, for other agents to read. -
GET /dashboard/agents/{agentId}/deployandGET /dashboard/agents/{agentId}/embed: the owner's pages to put it on a website with one paste. See The live widget. - MCP and A2A: other tools and agents can reach it; see MCP Developer Guide and A2A Developer Guide.
Costs
The guided creator that writes an agent's details for the owner (ai.features.agent_creator_v2) spends on its own lane, ai.spend.lanes.agent_creator.daily_ceiling_usd, which ships at 0 (closed). Talking to the agent spends on the owner and visitor turn lanes, each with its own daily ceiling. The switch is off until your administrator turns it on.
What you can do without a key
Read any public agent's page at GET /agent/{slug} and its abilities at GET /agent/{slug}/actions.json. Creating an agent always needs a signed-in owner.
For owners
Start with Create an agent or Create your own agent, then Choose its brain, Give it a voice, Publish it and Put your agent on your website. The longer walk-through is Creating Your Agent. Your agent says it is an AI on every page it lives on.
Was this helpful?
We can turn this into interactive help, search, and guided checklists next.
Previous guide
The harness router
Next guide