English · 简体中文
An AI operations workspace for SREs
Give Pilot a goal, hear from each specialist, and keep resources and execution records in one workspace.
macOS interface preview · v0.3.33 · Fictional demo data · UI shown in Chinese
- Pilot brings specialists into the conversation. Each specialist responds under its own name in the same chat. Pilot tracks missing replies and combines the responses when you ask for a summary.
- Resources stay in context. Switch between specialists and project resources in the sidebar. Reference resources, Skills, and MCP tools when describing the work.
- Execution stays visible. Use native Lake specialists or Codex through ACP, follow the actual executor and task records, and review tool calls and approvals. Historical failed or unknown operations remain available.
Open a specialist to see its task records, then inspect a task for execution details. The fixed Pilot entry returns to the main conversation; the Specialists / Resources tabs keep each list's position and expanded folders.
Specialist task history · Fictional demo data
Enter /plan to toggle planning, or /plan <goal> to start a planning request. Lake investigates with read-only tools, asks about material choices, and presents a reviewable plan. Revise it in the conversation or choose Execute plan to continue under the existing session permissions. Mode and plans persist per conversation.
- Download the current release for macOS Apple Silicon, requiring macOS 13.5 or later.
- Connect a model in settings, or connect through Magpie Agent. For Codex specialists, sign in to Codex locally and complete its setup in Agent settings.
- Add project resources, Skills, and MCP integrations. Type
@to reference them in a request to Pilot. - Ask Pilot to assign work or bring specialists into the conversation. Each specialist's sidebar entry opens its task records.
Pilot is the built-in coordinator in the sidebar. Its permission settings govern specialist operations; saved project restrictions and read-only Plan scope still apply. Usage and current scope (Chinese) →
The current stable release, v0.3.33, provides macOS packages only. Some earlier functionality remains in progress; see the release notes and compatibility status for the current limits.
Lake Memory (experimental): separately opt in to automatic accumulation and scoped preference use. Current instructions take precedence; consolidation runs in the background. Ordinary conversations and Pilot coordinator replies can use expression preferences; facts and operational experience still need verification. Manual strategy assets retain compatibility management. Online correction, candidate extraction, and new comparison experiments are retired; existing content and historical records are retained. Ordinary requests no longer write experimental prompt manifests. See the Lake Memory guide (Chinese).
Local development
Create a working branch from develop, open a PR into develop, and validate changes in Lake Dev. Production changes reach master through a develop → master release PR. Before editing, read the branch and release workflow and contributing guide. AI agents must also follow the root AGENTS.md.
You need Go 1.26.6 (see go.mod), Node.js 24, and pnpm. Before your first macOS build, run scripts/setup_lake_signing.sh. Then build and launch the isolated Lake Dev app:
pnpm --dir web-agent install --frozen-lockfile
scripts/build_lake_desktop.sh --dev --install
open "$HOME/Applications/Lake Dev.app"Dev uses ~/.lake-dev, separate from production data. The built application embeds the frontend and does not require Node.js at runtime. See Contributing for CLI and standalone web development.
Using Lake? We welcome real-world use cases and suggestions.
Usage and development · macOS · Windows development · Contributing · Apache-2.0