One knowledge base for your team and its AI agents
Everything your team writes and says, indexed into one base and answered under the permissions you already have: in the app, and over MCP to Claude Code, Cursor or any other client.
What it knows
Issues, code reviews, documents, chat, call transcripts and your git repositories. Nothing to write or maintain by hand: the base fills itself as the team works.
How it answers
Ask in plain words. The answer comes from your project, links to where it was found, and respects your permissions: you only see what you could open yourself.
Who can ask
Your team, from any screen in the app. Your coding agents, through the same connection they use to work with issues and code. One base, the same answers for people and agents.
Code: searched in place, not copied
Source code is part of the search, but it stays in your repositories: the assistant and your agents search it there over MCP, with ripgrep and symbol lookup, and nothing is sent to the embedding provider. CI logs and deploys work the same way: read on request, not indexed.
How a question is answered
- 1
Two passes at once
A keyword pass and a vector pass run on the same question.
- 2
One ranking
The two rankings are merged into one list.
- 3
Expanded along links
The issue, the review that closed it and the commits that named it come together.
- 4
Checked against permissions
Every fragment is checked for the person asking; an agent inherits the permissions of who started it.
- 5
With its sources
Every answer names where it came from, so you can open the source.
Questions about the knowledge base
- How is it different from a wiki?
- Nobody writes it. It fills up from work that happens anyway: issues, reviews, documents, chat and calls. When the work changes, the base changes with it.
- Who can see what?
- The same people who can open the source. An answer is built only from fragments the person asking is allowed to read, and an agent gets the permissions of the person who started it.
- What goes to a model provider?
- The text of the indexed sources goes to an OpenAI embedding model to build the index. Answers are generated with Anthropic models, or with your own key. Source code is searched in place and is not sent to the embedding provider. Call transcripts are indexed only after participants consent.
- Can our agents use it outside the app?
- Yes. Connect the MCP server once, and Claude Code, Cursor or any MCP client gets the same search and the same tools for issues, reviews, branches and deploys, signed in as you.