Lirae
The enterprise chat where people and AI agents work in the same rooms and act on internal systems, in a governed and verifiable manner.
- Client
- Bitapp Product
- Year
- 2026
- Technologies
- Rust
- Matrix
- MCP
- open-weights LLM
- Amazon Bedrock
- PostgreSQL
- pgvector
- xAPI
- React

Lirae is the enterprise chat where the team and its AI agents work in the same rooms. The agents don’t just respond: they take action on internal tools (trackers, code platform, cloud, files) on behalf of the person who requests it, through a control layer that decides what is allowed, respects each person’s actual permissions, and records everything in a tamper-proof log.
It runs on an infrastructure that you own, from start to finish: the data, and even the metadata, remain within the company.
Key points
- 01
People and agents in the same chat
Based on Matrix, open and end-to-end encrypted; you talk to an agent as you would to a colleague, in natural language. - 02
It acts with your permissions
The agent authenticates to the system as the person making the request; the system itself makes the decision – Lirae never copies or extends permissions. - 03
Human confirmation for risky actions
Reads are performed immediately; sensitive writes and deletions await approval in the chat by an authorized person. - 04
Tamper-proof log
Every action, including rejections, is recorded in an append-only cryptographic chain: who made the request, what was done, and with what outcome. - 05
Sovereign AI
The models run in your cloud account or on your hardware, in European regions; the content of the rooms never leaves your premises. - 06
Improves with use
Each agent gains experience on your systems and consults the curated documentation you provide.
Agents who are experts in each system
Instead of one bot per person, there is one specialized agent per system, which anyone can consult: one for Redmine (projects, issues, hours), one for GitLab (projects, issues, merge requests, pipelines), and a cross-functional agent that brings them together in a single conversation, for example, by reading an issue on the tracker and creating the corresponding one on the code platform.
Each person links their access from a private chat with the agent only once, using a personal key or their Google or Microsoft login; the agent deletes the pasted keys from the history.
Governed by design
The language model is considered untrusted: it proposes actions, and the control plane decides. Each request passes through a pipeline that verifies that the tool exists and is permitted for the agent, validates the arguments, resolves the identity of the requester and, if the action is risky, suspends it pending confirmation. If something is uncertain (an unknown request, an unconnected login, a confirmation that does not arrive), Lirae stops instead of making guesses.
The administration console
Operators manage the installation from a web console: they explore the audit log and verify its integrity, configure the agents and the tools they can use, manage users and linked logins, maintain the agents' memory, and check pending confirmations. Every change made by an operator is also recorded in the log.
Who is it for?
For organizations that want AI agents working on their systems but need to be able to trust what they do and hold them accountable: regulated companies, public administrations, and any team required to maintain control over their data. Lirae is open source and self-hosted, either on-premises or in your cloud.


