Docs/Guides/Publish & Share

    How to Use AVCodex with Claude Code

    Last updated · MAR 2026·Read as Markdown

    The AVCodex MCP Server brings the platform into Claude Code. Instead of switching between your editor and the AVCodex dashboard, you can create agents, manage knowledge sources, configure custom actions, and analyze conversations from where you already work. Programmers writing SIMPL# Pro, Q-SYS Lua, or AMX NetLinx in Claude Code can stand up a tier-1 support agent for the room they just commissioned without breaking flow.

    This guide walks through what you need to get productive with AVCodex MCP in Claude Code.

    Here's what happens when you use AVCodex MCP with Claude Code:

    1. You speak naturally. Ask Claude to create an agent, add a Biamp manual, or check analytics.
    2. Claude calls the right tools. The MCP server translates your request into API calls.
    3. AVCodex does the work. Agents are created, files are indexed, settings are updated.
    4. You see the results. Claude confirms what happened and suggests next steps.

    No clicking through menus. No copy-pasting IDs. Just describe what you want.

    Prerequisites#

    Before you begin:

    • Claude Code installed and running.
    • An AVCodex account (Free tier works, but some features require Studio or Studio Pro).
    • Your AVCodex API key from the dashboard.

    Connecting the MCP server#

    Follow our setup guide to connect the AVCodex MCP server to Claude Code:

    AVCodex MCP Server Setup Guide

    The guide covers:

    • Getting your API key from the AVCodex dashboard.
    • Configuring Claude Code (and other MCP clients like Cursor).
    • Verifying the connection.

    Once connected, Claude Code gains access to 84 tools for managing AVCodex agents.

    1. Agent management#

    Create, update, and configure agents without leaving the terminal.

    What you can do:

    • Create new agents with custom system prompts.
    • Update settings, models, and branding.
    • Search across all your agents.
    • Delete agents you no longer need.

    Example prompts:

    code
    "Create a Crestron tier-1 support agent for our managed-services portfolio"
"Change my Q-SYS programming agent to use Claude Sonnet 4"
"List all agents in my workspace"
"Update the system prompt for agent 12345 to be more direct (techs are usually on a phone)"

    2. Knowledge sources#

    Add documents, websites, and videos to give your agents AV expertise.

    Supported source types:

    • Documents. PDF, DOCX, TXT, MD, PPTX, XLSX, CSV, JSON, HTML.
    • Web content. Any URL, with optional full-site crawling.
    • Video. YouTube videos (transcripts are extracted automatically).
    • Social. Instagram, TikTok, Facebook posts.

    Example prompts:

    code
    "Add the Biamp Tesira Programming Guide PDF to my design agent's knowledge base"
"Crawl the entire docs.crestron.com site for my support agent"
"Upload these three commissioning checklists to agent 456"
"What knowledge sources does my Q-SYS agent have?"

    How document processing works:

    1. Upload. You provide a file or URL.
    2. Extraction. LlamaParse extracts text from complex documents (the table of inputs and outputs in a Crestron processor spec, the device tables in a Biamp Tesira programming guide).
    3. Chunking. Content is split into semantic segments.
    4. Embedding. Each chunk gets a vector representation.
    5. Indexing. Vectors are stored for fast retrieval.
    Pro Tip: Use extractionInstructions when uploading documents to guide the parser. For example: "Focus on the device addressing tables and ignore the cover-page legal disclaimer."

    3. Custom actions (API integrations)#

    Connect agents to external services so they can take real action.

    What custom actions do:

    • Query your PSA, project tracker, or asset database.
    • Open service tickets in Autotask, ConnectWise, or Halo.
    • Send notifications via Slack, Teams, or email.
    • Fetch real-time data from any API (control system status, room availability, headend health).

    Example prompts:

    code
    "Create an action that searches our D-Tools product catalog by SKU"
"Add a HubSpot integration to look up the AV contact at a client site"
"Test the 'create service ticket' action with sample data"
"List all custom actions on my support agent"

    Parameter types:

    Value source Description
    STATIC Fixed value you define
    AI_GENERATED AI determines from conversation
    SYSTEM_VARIABLE Runtime values like {{system.userId}}
    TOOL_OUTPUT Value from another action's result
    VARIABLE Reference to an agent variable

    4. Conversation analytics#

    Understand how techs and clients interact with your agents.

    Metrics:

    • Total conversations and messages.
    • Unique users over time.
    • Messages by source (web, API, Slack, WhatsApp, email).
    • Top queries and common topics.
    • Model usage breakdown.

    Example prompts:

    code
    "Show analytics for my Crestron support agent over the last 30 days"
"What are the top 20 questions field techs ask my support agent?"
"Export all conversations from last week"
"Which of my agents gets the most traffic?"

    5. Evaluations (evals)#

    Test your agents against expected responses to catch regressions.

    Note: Evals require a Studio Pro plan subscription.

    How evals work:

    1. Create a baseline. Start from an existing chat session.
    2. Define expectations. Edit responses to show what's correct.
    3. Run the eval. Claude re-answers each question.
    4. Compare results. See pass/fail with detailed scoring.

    Evaluation methods:

    Method What it measures
    factuality Are the facts accurate?
    faithfulness Does it stick to provided context?
    answer_relevancy Is the answer on-topic?
    context_relevancy Is the retrieved context useful?

    Example prompts:

    code
    "Create an eval for my Crestron support agent based on session abc123"
"Run all evals for agent 456"
"Show me the results of the last eval run"

    6. Message tags (intent detection)#

    Detect intents and trigger actions.

    Use cases:

    • Flag messages that need a programmer (escalation off tier-1).
    • Detect when a remote fix isn't possible (truck-roll-needed).
    • Identify upsell opportunities (a client asking about adding rooms).
    • Route conversations by topic (Crestron vs. Biamp vs. Q-SYS).

    Example prompts:

    code
    "Create a tag for 'truck-roll-needed' with 50 training phrases"
"Generate utterances for detecting techs stuck on Cresnet bus issues"
"List all tags on my support agent"
"Show me messages that matched the 'projector lamp' tag"
    Pro Tip: Use generate_utterances to create diverse training phrases. More utterances means better detection accuracy.

    7. Deployment channels#

    Your agents can live anywhere your techs and clients are.

    Supported channels:

    • Web embed. Add a chat widget to a client portal or service-management site.
    • Slack. Deploy as a Slack bot in your service ops channel.
    • WhatsApp. Connect via WhatsApp Business for international techs.
    • Email. Auto-respond to incoming AV-help inboxes.
    • Voice. Phone calls via Twilio/LiveKit (a hotline for field techs).
    • API. Build custom integrations into your existing tooling.

    Example prompts:

    code
    "What Slack workspaces is my agent connected to?"
"Show voice configuration for agent 789"
"Get email deployment settings for my AV help desk"

    8. AI-powered helpers#

    Let Claude help you build better agents.

    Available AI tools:

    Tool What it does
    generate_system_prompt Create a complete prompt from a brief description
    improve_system_prompt Get suggestions to enhance an existing prompt
    generate_suggestions Create conversation starters based on agent capabilities
    recommend_model Get model recommendations for your use case
    generate_utterances Create training phrases for intent detection

    Example prompts:

    code
    "Generate a system prompt for an AV programming assistant focused on Q-SYS Lua"
"Improve my support agent's prompt to be more concise (techs on phones)"
"Suggest conversation starters for my onboarding agent for new programmers"
"What model should I use for a Crestron SIMPL# Pro coding assistant?"

    Walk through standing up a complete support agent for an integrator.

    Step 1: Create the agent#

    code
    "Create an agent called 'Acme AV Support' in my main workspace.
It should be a tier-1 control system support agent. Friendly but
direct. Field techs are usually on a phone in a rack room. Crestron
and Biamp gear primary. Escalate to a programmer if it gets to module-
level code."

    Claude creates the agent and returns the ID.

    Step 2: Add knowledge#

    code
    "Add docs.crestron.com and support.biamp.com to the support agent.
Crawl all linked pages."

    Claude starts the crawl and monitors progress.

    Step 3: Configure actions#

    code
    "Add a custom action that opens a service ticket in our Autotask
instance when a tech needs to escalate."

    Claude sets up the API integration with the proper parameters.

    Step 4: Set up intent detection#

    code
    "Create a tag to detect when a fix needs a tech on site (truck roll).
Generate 100 training utterances."

    Claude creates the tag and populates it with diverse phrases.

    Step 5: Test the agent#

    code
    "Chat with my support agent and ask 'TSW-1070 shows offline. What do
I check first?'"

    Claude tests the agent and shows you the response.

    Step 6: Create evals#

    code
    "Create an eval based on that conversation. The answer should walk
through Cresnet bus power, then IP via Toolbox, then check the
program for the device ID."

    Claude sets up automated testing for QA.

    Note: This step requires a Studio Pro plan subscription.

    Step 7: Check analytics#

    code
    "Show me how the support agent is performing this week."

    Claude pulls metrics on conversations, top queries, and user engagement.

    For knowledge sources#

    • Be specific with extraction instructions. Tell the parser what to focus on (device tables, signal flow, error code references).
    • Refresh URLs periodically. Manufacturer firmware notes update.
    • Monitor processing status. Confirm documents completed successfully.
    • Use hybrid search. Combine keyword and semantic matching for best results.

    For custom actions#

    • Test before deploying. Use test_custom_action to verify API calls.
    • Use variables for secrets. Store API keys in agent variables, not action definitions.
    • Chain actions carefully. Use TOOL_OUTPUT to pass data between actions.
    • Add clear descriptions. Help the AI know when to use each action.

    For system prompts#

    • Start with generation. Let generate_system_prompt create a foundation.
    • Iterate with improvements. Use improve_system_prompt for refinements.
    • Be specific about tone. "Talk like an AV peer, not a help desk script".
    • Include examples. Show the AI what good responses look like (a real Cresnet troubleshooting walkthrough, a real DSP signal-flow explanation).

    For evals#

    • Test critical paths. Focus on the most common service requests.
    • Run after changes. Confirm updates don't break existing behavior.
    • Review failures carefully. Failed evals reveal prompt or knowledge gaps.
    • Use terminal format. Pass format: "terminal" for readable output.

    Most-used commands#

    Task Example prompt
    Create an agent "Create a Crestron support agent for Acme AV"
    Add documents "Upload this Biamp PDF to agent 123"
    Add a website "Crawl docs.crestron.com for my agent"
    Update the model "Switch agent 456 to GPT-4o"
    Check analytics "Show last 30 days of analytics for agent 789"
    Test the agent "Chat with my support agent: 'TSW-1070 offline'"
    Create an eval "Create an eval from session abc123"
    Run evals "Run all evals for agent 456"

    Tool categories#

    Category Tools
    Agent management create_app, update_app, get_app, list_apps, search_apps, delete_app, chat_with_app
    Knowledge add_document_source, add_url_source, add_text_source, list_knowledge_sources, get_knowledge_source, refresh_knowledge_source, delete_knowledge_source, get_crawl_progress
    Custom actions create_custom_action, update_custom_action, test_custom_action, list_custom_actions, get_custom_action, delete_custom_action
    Analytics get_app_analytics, search_conversations, get_conversation, get_top_queries, export_conversations, get_model_usage, get_response_time_analytics, get_user_engagement
    Evals create_eval, run_eval, list_evals, get_eval, update_eval, delete_eval, list_eval_results, get_eval_result
    Tags create_tag, generate_utterances, list_tags, get_tag, update_tag, delete_tag, list_tag_instances
    AI helpers generate_system_prompt, improve_system_prompt, recommend_model, generate_suggestions, list_available_models

    "Agent not found"#

    Confirm the agent ID. Use list_apps to see all available agents.

    "Knowledge source still processing"#

    Document processing takes time. Use get_knowledge_source to check status, or get_crawl_progress for URL sources.

    "Custom action failed"#

    Use test_custom_action with parameterOverrides to debug. Check that your API endpoint is reachable and credentials are correct.

    "Eval results seem wrong"#

    Review the expected responses in your eval. The AI compares against these, so they need to be accurate examples.

    • How Knowledge Sources Work. Deep dive into RAG and embeddings.
    • Custom Actions Overview. Advanced API integration patterns.
    • Analytics and Chat History. Understanding agent performance.
    • AVCodex MCP Server Setup. Connection and configuration details.

    *AVCodex · Your AV expertise. Amplified by AI.*

    Was this helpful?
    Edit this page →