Admiral

A company knowledge base that grows automatically as the team works, shared by people and coding agents across AI tools.

Admiral is a company knowledge base that automatically collects the team's work sessions with coding agents. Conversations, actions, and decisions are accessible through a web dashboard and MCP, even when switching computers or AI tools. It currently integrates with Claude Code, Codex, and GitHub Copilot.

It runs on your own server and receives data from developers' machines through Lookout, the program that manages collection and sharing rules.

Sessions

The dashboard brings together sessions from different tools, organized by project, team, and developer. Each session can be opened to read the conversation, examine tool calls, and check models and usage.

The collection updates as the team works and can also include transcripts already stored on connected machines.

Search and agent access

Search finds text in prompts, responses, and tool output, showing matching lines with their surrounding context. It supports regular expressions, whole words, and case-sensitive matching. Results can be filtered by project, developer, team, and date range, or used as the starting point for another search.

Agents can access the same content through MCP. From a Codex session, for example, an agent can search for a solution a colleague discussed with Claude Code and read the relevant passages. Access follows the permissions of the developer using the agent.

Project ledger

Each project has a ledger where agents record changes, decisions, and checks. Entries describe what was done, the reasons for the choice, and the outcome, with a reference to the session when available.

A new entry can replace an earlier one without deleting it. The dashboard shows current decisions and their history, with text search, filtering by component, and Markdown export.

Company conventions

Conventions define the rules the team's agents must follow, either across the company or for a single project. Each rule can include a reason and tags, such as backend or security, to help find the rules relevant to the work at hand. Agents receive general guidelines through MCP when they start; they read specific rules as needed, regardless of the AI tool being used. Rules are managed in the dashboard, and agents can propose new ones, which take effect after a person approves them.

Usage

The dashboard shows costs, sessions, requests, tokens, and rejected actions for the selected period. The daily chart and tables can be grouped by project, model, team, or developer.

Aggregated figures link to individual sessions, connecting usage to the work performed.

Lookout

Machines connect using a command generated in the dashboard, which installs Lookout and configures the supported tools. Lookout receives OpenTelemetry data, stores it locally, and forwards it to the server; it also syncs transcripts.

The desktop interface and command line provide controls for sessions, connections, and usage. Multiple Admiral instances and other OpenTelemetry collectors can be connected, each with an independent queue.

Sharing and permissions

Lookout's rules determine what is sent to each instance, with settings for individual project folders. Private sessions share only costs and token counts. A delay before sending gives the developer time to make a session private; if its content has already been shared, Lookout requests its removal from the server.

A single Admiral installation can host multiple companies with separate data. Roles control access to features and content, including the option to limit developers to reading only their own sessions.

Installation

The server is a single Rust executable that includes the API, telemetry collection, and web dashboard. It uses SQLite or PostgreSQL; the dashboard is written in React and TypeScript. Lookout is also written in Rust, with an egui-based desktop interface.