Chatbot & Workflow Integrations
Setup guides for connecting Transient with AI clients and automation platforms. Each integration is a thin adapter over the same MCP or REST API contract.
Which surface to use
Claude Desktop
MCP (stdio)
Full tool surface. Native MCP support — add the TI server to your config and all tools are immediately available.
ChatGPT
REST API · MCP (Desktop)
Custom GPT actions via REST API. ChatGPT Desktop supports MCP on Business plan and in developer mode — enable it in Settings → Developer, add the TI server config, and use ti_create_upload_handoff for local files.
Perplexity
REST API
Point a custom tool definition at /v1/answer with your API key.
Gemini (Google AI Studio)
REST API
Use a Function Calling tool definition pointing at /v1/answer. Works in both Gemini API and AI Studio custom agents.
Grok (xAI)
REST API
Pass /v1/answer as a tool in the xAI API tools array. Grok will invoke it when evidence retrieval is needed.
Setup guides
Step-by-step setup instructions and code examples for each integration category.
AI Clients
Claude, ChatGPT, Perplexity, Gemini, Grok
MCP and REST API setup for major AI clients. Includes Desktop MCP config, Custom GPT Actions, and function calling examples.
View guideAutomation Platforms
Make, Zapier, n8n, Power Automate
Trigger-based document pipelines. HTTP module setup, code step examples, and response routing for each platform.
View guideProductivity Tools
Notion, Google Drive, Slack
Connect TI to where your documents live. Fetch, analyse, and route cited results back into your existing tools.
View guideCommon guardrails for all clients
Call ti_workflow_policy once at session start (MCP)
This loads the canonical TI connector rules into the model context, ensuring deterministic upload→ask behaviour across restarts.
Local paths only block in sandboxed environments
Claude Desktop and local MCP servers can read files from disk directly. In sandboxed clients (ChatGPT Desktop, Claude.ai web), paths like /mnt/data/file.pdf are not accessible — extract the text first or use ti_create_upload_handoff to generate a browser upload URL.
Enforce citation discipline in your system prompt
Explicitly instruct the model to only present claims traceable to the returned citations array. Prevents probabilistic completion from overriding evidence-first output.
Handle empty citations explicitly
If citations is empty, the correct response is 'No evidence found.' Do not ask the model to synthesise from its own priors when TI returns nothing.