Contract Q&A is a Chat page, triage lives in Dify, internal APIs sit in a script. AngusAI is one path: four agent types, knowledge, then a real publish switch. API Keys are not wired to chat yet — call with a user token or PAT.
What is AngusAI?
The Build product in Xiaocan Cloud’s AngusKit: a private agent platform. You configure four agent types, try them in Playground, then publish as Web, embed, or an OpenAI-compatible API, including local models such as Ollama.
When should I use each of the four agent types?
General chat is for multi-turn Q&A with few extras. RAG is for answers that must cite a knowledge base. Workflow is for multi-step, scheduled, or approval paths and must bind a published flow. Code interpreter runs code. After changing type, re-select attached resources.
How do we choose versus Dify or Coze?
Coze is shorter for cloud demos. Dify is more familiar for orchestration. Compare AngusAI when you need on-prem, local models, a publish path, and the same login as the suite. It is not ChatGPT’s website and is not Cursor.
How do we publish as Web, embed, or an OpenAI-compatible API?
Change the agent from offline to published, then create and bind an app. Entry points are console chat, a share page, an embed, or the compatible API. The API model field must match the agent code. Use a user token or PAT; keys on the API-key page are not wired to chat yet.
Are local models / Ollama supported?
Yes. Providers include Ollama and “Other” OpenAI-compatible endpoints such as vLLM. The Base URL must be reachable from the AngusAI server, not localhost on a laptop.
How do we attach knowledge bases and tools to an agent?
Select knowledge bases, tools, or API sets on the agent’s related resources. Put outbound credentials in the secret center; objects store references only. RAG without a bound corpus, or workflow without a published flow, fails at chat time.
What can guardrails block?
The Guardrails sidebar can apply keyword, regex, PII, length, and topic rules on input and output. That is policy, not a prompt instruction. Docs state guardrails do not apply to Playground chats—only to live API and app sessions.
How does Playground differ from production publish?
Playground is a console trial and does not appear in the chat dropdown. New agents start offline; shipping requires published status and a bound app. A green Playground is not traffic on the share page or compatible API, and guardrails do not run there.
When should we pick workflow versus conversational agents?
Use general chat for multi-turn Q&A with few tools. Use workflow for fixed multi-step, scheduled, or approval work, bound to a published flow. To call internal systems, set the default reasoning mode to TOOL_CALL and include it in the allow list.
How should secrets and API keys be managed?
Store outbound secrets in the secret center. Sidebar API keys are not used for chat auth. Call chat APIs with a user token or PAT. Do not put secrets in the repo or in chat.
What should we check before shipping an agent?
Confirm where model calls go, whether answers cite the corpus, that tool secrets sit in the secret center, that guardrails apply on the live entry, and that publish/rollback work. Playground is not regression; keep AngusTester reports for citation, refusal to invent, and guardrails.
Where should we start in the docs?
Start with the AngusAI introduction and quick start, then Agents and AI apps. Models and Ollama are in model management; guardrails have their own chapter.