What it does
Attach one or more git repositories, a folder or a public https address, and each gets an expert that builds a searchable knowledge base, writes an overview of itself and answers questions with citations to exact files and lines. A question is routed to the matching experts and their answers are combined.
A knowledge bot names a set of repositories as one bot with its own memory (notes that go stale when the code they cite changes), its own instructions and an automatic code graph: imports, definitions, HTTP routes, queues, packages and how repositories connect.
Search first, then answer
Search runs by keyword and by meaning, using a small local embedding model. The relevant code is found first, then the answer is written and checked: the citations must exist and the file names must appear in the sources. With no model on the server you still get the cited matches.
Answers come from a small local model unless you choose another provider. A checked answer can still be wrong or muddled, so open the sources for anything important.
Ask from where you work
Use a bot in the dashboard, from the bot command, or connect your own AI tool over MCP so Claude Code and other MCP apps use the same bots. A teammate can also ask a shared bot from Slack, Teams or Google Chat. See the chat page.
Access follows people
A shared bot only shows what each person may see: repository access is granted per person and group, with roles from viewer to admin. A private repository attached by https address needs a read-only access token (stored encrypted, write-only, never shown again) or a connected git account. Otherwise clone it yourself and attach the folder. The bot never asks for anyone’s personal git password.
Where your code goes
Your repositories, the index and the notes stay on the computer or server you run Triazz on. What leaves is the specific text sent to the AI model you chose and the calls to the git hosts you connect. With a local Ollama model, no text goes to a model provider. See the Privacy page and the Security & Trust Centre.
Honest limits
- The relevance cut-off was calibrated on one embedding model with a small sample.
- A shared server is designed for roughly 50 to 100 people; larger organisations run one server per team.
- Triazz needs Node 22+ (and Ollama for local models), and is set up with us because the source is not public.
- Email, team chat, tasks and browser QA are part of the product too. See the email page.
Want to try it with your team?
Write to us and we will talk it through.