Studio Guide

Studio is in development

The Studio visual builder is being developed for Astromesh Cloud v0.1.0. The 5-step wizard configuration format (WizardConfig) is fully functional via the API — Studio provides a visual interface for the same configuration.

Astromesh Studio is the visual agent builder at studio.astromesh.io. It provides a 5-step wizard for creating and deploying agents without writing YAML or code. Under the hood, it generates the same WizardConfig that you can also POST directly to the Cloud API.

The 5-Step Wizard

1. Step 1: Identity & Model

   Configure the agent’s name and which LLM it uses.

   Fields:

   • Agent name — URL-safe identifier (used in API paths and runtime ID)
   • Display name — Human-readable label shown in Studio
   • Primary model — The LLM to use for responses
   • Fallback model — Used if the primary model fails or is unavailable

   Available models:

   Provider	Models
   OpenAI	gpt-4o, gpt-4o-mini, gpt-4-turbo
   Anthropic	claude-3-5-sonnet, claude-3-haiku
   Groq	llama3-8b-8192, mixtral-8x7b-32768
   Ollama	llama3, mistral, phi3 (platform-hosted, no key required)

   To use OpenAI or Anthropic models, you need to add a provider key first (see Authentication).

2. Step 2: Prompt & Persona

   Write the system prompt that defines your agent’s behavior and personality.

   Fields:

   • System prompt — The main instruction set for the agent. Supports Jinja2 templating.
   • Persona — Short label used in Studio (e.g. “Support Agent”, “Data Analyst”)

   Jinja2 variables available in prompts:

   Variable	Description
   {{ session_id }}	Current session identifier
   {{ context }}	Context dict passed in the run request
   {{ context.user_id }}	Example nested context access
   {{ today }}	Current date

   Example system prompt:

   You are a helpful support agent for Acme Corp.
   You are speaking with user {{ context.user_id | default('a customer') }}.
   Always be concise, friendly, and professional.
   If you cannot answer a question, direct the user to support@acme.com.

3. Step 3: Tools

   Select which tools the agent can call during a conversation.

   Available tools:

   Tool	Description	Status
   calculator	Arithmetic and math expressions	Available
   web_search	Search the web for current information	Available
   weather	Get current weather by location	Available
   datetime	Get current date/time in any timezone	Available
   code_interpreter	Execute Python code snippets	Coming soon
   file_reader	Read uploaded files	Coming soon
   sql_query	Query a connected database	Coming soon
   http_request	Call external REST APIs	Coming soon

   Tools are managed by the platform’s ToolRegistry. The agent automatically decides when to call a tool based on the user’s query — you don’t need to configure invocation logic.

   MCP Tools: Custom tools via Model Context Protocol (MCP) are supported in the core runtime. Studio support for MCP tool configuration is coming in a future release.

4. Step 4: Memory

   Configure how the agent remembers information across turns in a conversation.

   Memory type:

   Type	Description
   conversational	Stores recent chat history (recommended for most agents)
   semantic	Vector embeddings for long-term knowledge retrieval
   episodic	Event log of past interactions

   Memory strategy (for conversational memory):

   Strategy	Description
   sliding_window	Keep the last N messages. Fast, predictable cost.
   summary	Summarize older messages to stay within token limits
   token_budget	Keep messages until a token budget is reached

   Window size: Number of messages to retain (for sliding_window). Default: 20.

   Memory is scoped to session_id. Reusing the same session_id across calls gives the agent continuity. Using a new session_id starts fresh.

5. Step 5: Guardrails & Limits

   Add safety filters and resource constraints.

   Input guardrails (applied to user messages before the model sees them):

   Guardrail	Description
   pii_filter	Detect and redact personal identifiable information
   prompt_injection	Block prompt injection attempts
   topic_filter	Restrict to allowed topics (requires configuration)

   Output guardrails (applied to model responses before returning to the user):

   Guardrail	Description
   content_safety	Filter harmful or inappropriate content
   hallucination_check	Flag responses with low confidence
   format_validator	Ensure output matches expected format

   Limits:

   • Max tokens — Maximum response length. Default: 1000.
   • Max iterations — Maximum tool call loops per request. Default: 5.
   • Timeout — Request timeout in seconds. Default: 30.

Deploying from Studio

After completing the 5-step wizard, click Deploy to transition the agent from draft to deployed. Studio calls POST /orgs/{'{slug}'}/agents/{'{name}'}/deploy on your behalf.

Once deployed:

• The agent gets a public endpoint you can copy from Studio
• You can run test queries in the Studio playground (uses POST .../test)
• Usage metrics appear on the org dashboard

Editing a Deployed Agent

To edit a deployed agent in Studio:

1. Open the agent
2. Click Pause — this deregisters the agent from the runtime
3. Make your changes in the wizard
4. Click Deploy to bring it back online

Changes to the system prompt, tools, or model take effect immediately on the next deploy.

API Equivalent

Everything Studio does maps directly to Cloud API calls. If you prefer working in code or CI:

# Create (equivalent to completing the wizard)
curl -X POST "https://api.astromesh.io/api/v1/orgs/$ORG/agents" \
  -H "Authorization: Bearer $TOKEN" \
  -d '{ ... WizardConfig ... }'

# Deploy (equivalent to clicking Deploy)
curl -X POST "https://api.astromesh.io/api/v1/orgs/$ORG/agents/my-bot/deploy" \
  -H "Authorization: Bearer $TOKEN"

# Test in playground (equivalent to Studio playground)
curl -X POST "https://api.astromesh.io/api/v1/orgs/$ORG/agents/my-bot/test" \
  -H "Authorization: Bearer $TOKEN" \
  -d '{"query": "Hello", "session_id": "test-1"}'

See the API Reference for complete documentation.