lobu

Purpose

Lobu is an open-source control plane and runtime for organization-level AI agents, with the core goal of providing shared company context and a "brain." It uniformly records tool calls, data changes, device signals, API events, and conversations as events, and associates them with the entities and identities they belong to (people, projects, customers, households, companies), thereby giving agents an organizational state that can be queried in real time and reused persistently, rather than temporary records rebuilt separately by each session.

Main Capabilities

• Event-sourced context layer: Once connected, connectors from polling, Webhooks, APIs, and agent writes flow into the same append-only event log.
• Entity and knowledge association: Events can be indexed as searchable knowledge and associated with typed entities such as Company, Project, Customer, Incident, or custom schemas; corrections supersede old facts while preserving provenance.
• Multi-agent sharing: Authorized MCP clients read and contribute to the same organizational state according to their respective authorizations; ChatGPT, Claude, Codex, and custom agents can all use it.
• Persistent experts: Lobu experts have independent identities, instructions, tools, persistent conversations, and access policies. They can be invoked by other agents through the web app, Slack, or MCP, with task results returned to the original conversation.
• Governance and security: Identity, source permissions, approvals, credential proxying, provenance tracking, and auditing are all handled server-side; connector actions follow the organization's Auto, Ask, and Block policies, and guardrails can inspect inputs, outputs, and tool calls.
• Background automation: Automations are versionable background responsibilities that can be triggered manually, on a schedule, by connector events, or by the persistent output of other Automations.
• Multi-channel support: Experts can serve Slack, Telegram, WhatsApp, Discord, Teams, Google Chat, the web app, and the REST API.
• Optional execution: Built-in just-bash lightweight execution; when stronger isolation or more compute is needed, remote sandboxes such as Vercel Sandbox can be connected.

How to Use

There are three usage paths:

1. Add shared context to an existing agent: Point any MCP client at Lobu, complete OAuth, and connect the sources you want to share, without needing the Lobu agent runtime or lobu.config.ts. Run lobu connect to detect the client, install supported MCP and skill packages, or obtain the native handoff method required by the host. The same MCP endpoint works with Claude Code, Codex, OpenCode, Antigravity, ChatGPT, Claude Desktop, Cursor, and custom MCP clients.
2. Run persistent Lobu experts: Create experts for long-term responsibilities, converse in the Lobu web app or Slack, and other agents can invoke them via MCP. Local run example:lobu run will start the local stack with embedded Postgres and open the web UI at :8787, applying the project's lobu.config.ts automatically only in embedded mode. When using external Postgres, you need to set DATABASE_URL and ensure pgvector is available.lobu init(including lobu init -y) will generate ENCRYPTION_KEY in the project .env; manual setup requires first providing a 32-byte base64 value (openssl rand -base64 32), and it must remain the same key as the persistent database, otherwise encrypted keys will be unreadable.
3. Build with the CLI and TypeScript SDK: Mint tokens via lobu token create ; MCP tools and typed SDK operations share the server-side tool registry and access rules;lobu memory run and lobu memory exec Distributed via the MCP endpoint. External agents can pass scripts to Lobu's query_sdk and run_sdk MCP tools. Scripts must export an async function by default, with the execution context as the first parameter and the ClientSDK as the second parameter. Top-level await or return will cause a compilation failure.

Applicable scenarios

Suitable for technical teams already using Claude Code, Codex, ChatGPT, or custom agents that want these agents to share company context or delegate to persistent experts. Typical entry points are: one team, one to two connection sources, one recurring responsibility. It can be used for persistent responsibilities such as customer research, support triage, release coordination, incident follow-up, or internal domain experts.

Limitations and notes

• Local embedded execution (just-bash) and embedded execution mode are policy and convenience boundaries, not a virtual machine for running malicious code; when stronger isolation is needed, a remote sandbox provider should be used.
• Shared context and MCP delegation do not require Lobu to execute code for the calling agent; whether to execute and which sandbox executes it is a deployment choice.
• When upgrading from an older Mac installation (before the ai.lobu.mac app identifier), the DMG must be installed manually; in-app updates cannot migrate across bundle identifiers. /Applications/Owletto.app must also be deleted if it still exists. As of 20.0.1, settings and device identity can be retained across the rename, but login items and privacy authorizations cannot migrate with it.
• The canary tag only moves when a maintainer triggers Publish Packages from main; before depending on it, check the workflow run result; for stable releases use @latest. Package managers may cache moving tags, so pinning a specific version is recommended.
• Specific billing, pricing, platform coverage, and performance data are not yet provided.

Security and privacy

Permissions and auditing remain in the Lobu gateway. The Lobu MCP server and credential broker layer handle provider and connector credentials, OAuth and token refresh, and third-party API proxying; Workers only receive scoped placeholders or short-lived derived access, and OAuth tokens or persistently stored credentials are never distributed. Connector actions follow organizational approval policies, and connection data remains organization-scoped.