lobu

Lobu

Use this skill when setting up a brand-new Lobu agent from scratch (follow "Onboarding a New Project" below) or when running, validating, evaluating, or connecting an existing Lobu project, or operating Lobu memory from a coding agent.

The AGENTS.md generated into a new project is the source of truth for the config API (the define* helpers, connectors, auth, Automations, memory) — this skill does not duplicate it. For an existing project, jump to "Core Model" + the relevant reference section below.

Onboarding a New Project

Use this playbook when a new user asks you to build them a Lobu agent from scratch. Set it up end to end: ground yourself in what the user already has, interview, discover, scaffold, prove the complete local path, then offer the deployment choice. Do not call the setup complete until every path the user selected has actually passed end to end. If an external provider, authorization, deployment, or rate limit blocks the proof, name the exact blocker and leave the setup as blocked rather than claiming success.

1. Ground yourself in what the user already has

Before interviewing, look at what already exists so your questions and suggestions are personal rather than a blank slate. This is discovery, not a design decision — do not let it skip the interview.

• Local: run lobu context list to see existing contexts; check the current directory and home for existing Lobu projects (lobu.config.ts), configured MCP servers, and CLI auth state. Note what the user already set up.
• Cloud: if the user has an active session (do not create one just to look), call client.organizations.list(), client.agents.list() for each org, client.catalog.listInstalled({ kinds: ["connectors"] }), and search_memory for prior knowledge. This reveals existing agents, connectors, and memory you can build on.
• Tell the user what you are looking at so they can correct you, and let what you find shape the options you offer. Never treat this discovery as an identity signal: orgs, emails, or domains never prove a company or product, and never mention a competitor or third party unless it actually appears in what you found.

Do NOT ask for a login email during the interview. Email is only needed when a managed connector actually requires a session (step 3) or when the user chooses cloud deployment (step 6); ask for it at that point. Treat identity only as identity; never infer a company from an email address or domain.

2. Interview with concrete, personalized options

Interview the user one question at a time; wait for each answer, do not guess, and do not batch the questions. Anchor each question in what you found in step 1. Instead of asking "what agent do you want?" as a blank slate, summarize what you see and offer two or three concrete alternatives to choose from:

• "You already have org X with connectors A and B installed and N existing agents. Do you want to (a) add a new agent on top of those, (b) extend one of the existing agents, or (c) start something standalone?"
• "For what it should do, the common shapes are (a) a triage agent that reads incoming chat and files a ticket, (b) a scheduled report that summarizes new activity each morning, (c) a research agent that answers questions from your connected tools. Which fits, or what would you change?"

Offer a default so the user can answer "the first one" — decisive users and undecided users both move faster with a concrete menu.

• What company or product is this for?
• What does it do, and who does it serve?
• What is its public domain? (optional for a personal project)
• Do I have permission to research the public website and other public sources?
• What is the agent for?
• What should we call it?
• Who uses it: just me, my team, or each of my customers?
• How should it behave, and what may it do unattended versus only with approval?
• What should it remember? Model this as one to three entity types.
• Where does its data come from? Pick one connector or custom source to start.
• Where should people talk to it? Offer Web UI, API or SDK, MCP, Slack, Teams, WhatsApp, Discord, Google Chat, and Telegram without forcing a chat platform. If Slack is chosen, ask whether to join Lobu's hosted workspace for a quick start, install Lobu into an existing workspace, or use a custom Slack app; then ask whether to test the DM with Lobu, a named channel, or both, and whether people knowledge means conversation participants, bound-channel members, or the full workspace directory.
• Should anything run on a schedule?
• Which supported LLM provider credential does the user have?

If the user gave permission to research and has a public domain, research the company before proposing the design. Summarize its product, customers, workflows, and useful signals, and cite the public sources you used. Then propose one to three entity types, their relationships, the first data source, one or two useful Automations, and suitable interaction channels. Play back a short numbered plan and wait for confirmation before scaffolding anything.

3. Discover the surface before choosing how to scaffold

Check Node is 22-24 or 26+ (only Node 25 is unsupported), then run lobu context list and identify the named Lobu Cloud context; do not assume the active context still points to cloud after a local run. If a managed or cloud connector was chosen and there is no session, ask for the login email at that point, explain that Lobu will email an approval link, and run lobu login --email <confirmed-address> --context <cloud-name>. The person approving must sign in with that same email; knowing an address never grants access. After a successful login, an empty organization membership list is a tenancy/discovery result, not an authentication failure: call client.organizations.list() and explain the result. Never revoke or repeat login with --force unless the user separately approves that re-authentication.

Use ClientSDK through Lobu MCP or lobu memory exec with the explicit cloud context. Call client.organizations.list() to find public organizations offering managed connections. Call client.catalog.listInstalled({ kinds: ["connectors"] }); if the connector is not installed, call client.catalog.listCatalog({ kinds: ["connectors"] }). Inspect the selected entry's detail.auth_schema, detail.options_schema, detail.feeds_schema, and detail.automation_events instead of guessing field names. If a matching live managed offer exists, explain it and call client.connections.connectManaged only after consent. Complete the returned user authorization, then run lobu context use <cloud-name> immediately before executing its local_bootstrap_command from the parent directory where the new project should be created; use --here only after entering an empty directory, and do not create a blank Lobu scaffold first. Ask for bring-your-own OAuth credentials only if no managed offer exists or the user prefers their own app.

If no connector in the catalog matches the data source, do not report "no integration exists" — author a custom connector: write connectors/<name>.connector.ts in the project, register it with connectorFromFile("./connectors/<name>.connector.ts"), and model it on the npm downloads example. See Connector authoring for the full contract; the end-to-end entity/event/Automation model is in Lobu concepts.

4. Scaffold and build from the confirmed plan

If managed auth supplied a local_bootstrap_command, use the project it generated. Otherwise run npx @lobu/cli@latest init with the confirmed name and provider. Postgres is built in, so lobu run starts an embedded database; ask for DATABASE_URL only if the user chooses external Postgres. Read the generated AGENTS.md completely before editing lobu.config.ts. Use the canonical reference at https://github.com/lobu-ai/lobu/blob/main/examples/lobu-crm/lobu.config.ts when needed; do not assume that example exists in the new project's filesystem. Briefly explain how the chosen feed may use checkpoints to collect incrementally, emits events, builds shared entity memory, and gives both chat and Automations the same governed context.

feeds_schema.<feed>.eventKinds are the default Automation trigger catalog for feed connectors: every declared kind is a subscribable event type (event_type = the kind slug) for any feed that emits the kind. A feed's first successful non-dry sync establishes its baseline without activation; later inserts of a matching kind activate subscribers. Feed-scoped filtering is not yet available (events dedupe by connection + origin id, so a feed_key match would miss origins first seen by another feed). automation_events (now derived from eventKinds when a connector declares none) is the trigger picker's list. If no declared kind matches the use case, use a scheduled Automation with a bounded named SQL source and read event text from payload_text. Validate the query against the running local schema before relying on it. If feeds_schema is empty, do not invent or trigger a feed; chat integrations such as Slack can deliver incoming messages directly. For Slack or Telegram, prefer defineConnection with credentialMode: "hosted" so no bot token is required; lobu run prints the exact hosted-workspace join URL, own-workspace install URL when available, and short-lived link command. Use those returned values instead of constructing URLs. A printed /lobu link <code> is one-time: redeem it in the first selected DM or channel. In an installed workspace whose installer is linked to Lobu, bind each additional surface with /lobu link <agent-id>; otherwise generate a fresh code for that surface. Mention Lobu in a channel test and verify the App Home separately. Lobu email login and Slack installation are separate: the email link authorizes Lobu CLI access, while an existing-workspace install requires the user or workspace admin to approve Slack's OAuth screen. Never describe Slack consent as an email approval. Use credentialMode: "byo" only if the user chooses their own app. Tell the user once which remaining secrets are required and add only secret("ENV_NAME") placeholders to config. Never open or request permission to read .env; inspect .env.example when present and declared secret placeholder names only. Never read, print, invent, or paste a real secret into source or chat.

If the user wants the agent to know people in Slack, ask whether that means conversation participants, members of bound channels, or the full workspace directory. Installing the app proves none of those by itself. Verify the selected scope by inspecting the people or member entities Lobu actually materialized and by asking the agent a roster question. If human-readable names or non-participants are unavailable, identify that exact product gap; never claim the complete workspace directory was synced.

5. Prove the complete local path

Change into the generated project directory and verify pwd contains the intended lobu.config.ts; never run project commands from its parent. Run npx @lobu/cli@latest validate, boot with npx @lobu/cli@latest run, and verify health and the local Web UI. With the embedded database default, lobu run creates and selects the local context and auto-applies the project; with an external DATABASE_URL, it does neither, so authenticate and apply to that runtime explicitly. Send a harmless direct chat message without tools.

Before consent to access provider data, inspect only catalog, connection, auth-profile, and feed metadata. For a browser_session profile only, npx @lobu/cli@latest connector run <key> --auth-profile <browser-session-profile-slug> --check resolves the device-bound path without execution. Never use connector run as a generic managed OAuth or env-profile check; those durable credentials stay at the gateway. Do not run a feed or connector dry-run before consent, because a dry-run may still read provider data. Ask explicitly before accessing real provider data. Only after approval, dry-run the connector or feed server-side and trigger the selected feed manually if that connector actually declares one. A feed dry-run records its run but persists no collected events, entities, attachments, checkpoint changes, or feed sync state. Run each Automation once, poll client.operations.getRun with its returned run_id, and show each completed run plus any declared result event or entity. For a run-result-only or reaction-only Automation, use the completed run and the Automation view_url returned by Lobu. For Slack, treat DM, channel mention, App Home, and requested member discovery as four separate checks. Keep actions in approval mode unless the user explicitly approves execution. If any step fails, fix it and rerun that step; do not silently substitute a different path.

6. Offer the deployment choice

After the local proof, ask the user whether to keep running locally or deploy to Lobu Cloud. Explain that the embedded local runtime and Lobu Cloud are separate targets, and that keeping it local needs no further setup. Only if the user chooses cloud deployment, ask for the login email to identify the deployer, explain that Lobu will email an approval link, and run lobu login --email <confirmed-address> --context <cloud-name>; the person approving must sign in with that same email. Because lobu apply has no context flag, use lobu context list and then lobu context use <cloud-name> to select an explicit cloud context before deployment. Run lobu apply --dry-run first, show the target organization and exact plan, and wait for confirmation before applying. After apply, use only authoritative view_url values and link commands returned by Lobu; never construct URLs. Present the working access choices: Web UI, direct agent chat, API or SDK, MCP, and any channels the user selected.

Core Model

• Lobu is the agent framework, runtime, deployment layer, and memory surface.
• Keep framework configuration in lobu.config.ts (TypeScript, defineConfig from @lobu/cli/config).
• Keep agent identity and instructions in IDENTITY.md, SOUL.md, and USER.md.
• Keep reusable capability bundles in skills/<name>/SKILL.md or agents/<agent>/skills/<name>/SKILL.md.
• Use lobu login for CLI authentication. Do not use a separate memory login command.
• Use lobu connect for MCP client and skill setup.
• Use lobu memory ... for memory operations, seeding, direct tool calls, and browser-auth capture.

Reference the canonical docs before guessing how a feature works: Lobu concepts (entities vs events, identity, end-to-end lifecycle), Automations (trigger/source/output contract), and Connector authoring (custom connectors).

Project Checklist

1. Read lobu.config.ts first.
2. Read the active agent files under agents/<id>/.
3. Check local skills under skills/ and agents/<id>/skills/.
4. Use lobu validate after config changes.
5. Discover the live SDK and connector surface with search_sdk; never guess a method, connector key, feed key, connection id, or operation key.
6. When prompt or Automation changes, run evals via promptfoo (see examples/personal-finance/evals/promptfooconfig.yaml). The in-house lobu eval command has been removed.

Common Commands

npx @lobu/cli@latest init my-agent
npx @lobu/cli@latest run
npx @lobu/cli@latest validate
npx @lobu/cli@latest login

Authentication

• Interactive (human at a terminal): lobu login runs the device-code flow with a browser approval.
• CI / your own automation: LOBU_API_TOKEN, or lobu login --token <pat>.
• Local lobu run: the CLI mints credentials automatically over loopback — no prompt.
• Headless, on a user's behalf: lobu login --email <address>. The server emails the user a one-click approval link and the CLI polls until they approve, then stores the scoped credential — no TTY, no pre-minted token. This is the auth.md "user_claimed" flow.

An external agent not using this CLI can drive the same flow over HTTP: read <origin>/auth.md (linked from the agent_auth block in <origin>/.well-known/oauth-authorization-server) for the endpoints. Today only the email user_claimed flow exists (no zero-touch ID-JAG). Never fabricate a token.

Organization Deletion

Treat organization deletion as permanent and restricted to organization owners. The supported path is the Organization settings page, opened from that workspace's settings button in the organization switcher. ClientSDK and Lobu CLI intentionally expose no organization-delete method. When asked to delete organizations, identify each exact target and its data impact read-only, surface one production organization at a time for explicit confirmation, then direct an owner to that page or use their logged-in browser only when explicitly authorized. Never invent a hosted URL, call the raw Better Auth route, bypass a failed authorization check, delete organization rows directly, or batch confirmations.

Cloud CLI OAuth/PAT credentials are not a Better Auth browser session. A 401 or 403 is a stop signal: report the authorization boundary and leave the organization intact.

Memory Defaults

Your long-term memory is powered by Lobu. Do NOT use local files (memory/, MEMORY.md) for memory.

• Lobu automatically recalls relevant memories when you receive a message.
• To save something, call save_memory with the content and an appropriate semantic_type.
• To search, call search_memory. Results include view_url links to the web interface.
• NEVER construct Lobu URLs yourself. When the user asks for a link, call search_memory to get the correct view_url.
• When the user says "remember this", save it to Lobu immediately.
• When a message changes a fact you already stored (an updated preference, status, count, location, or plan), first search_memory for the prior memory to get its id, then save_memory the new value with supersedes_event_id set to that id. This replaces the stale value so future recalls return the current one; the old value stays in history but is hidden from normal search.

Lobu Memory

Configure project-scoped memory in lobu.config.ts by setting the org on defineConfig and declaring the schema with the define* helpers:

import { defineConfig, defineEntityType } from "@lobu/cli/config";

const ticket = defineEntityType({
  key: "ticket",
  name: "Ticket",
  properties: {
    subject: {
      type: "string",
      "x-table-label": "Subject",
      "x-table-column": true,
    },
  },
});

export default defineConfig({
  org: "my-org",
  orgName: "My workspace",
  agents: [/* ... */],
  entities: [ticket],
});

Seed data records still live as YAML under ./data. Then seed or operate the memory workspace with:

lobu login
lobu memory org set <org-slug>
lobu memory health --org <org-slug>
lobu memory seed --org <org-slug>
lobu memory run search_memory '{"query":"Acme"}' --org <org-slug>

Use search_memory first when the user asks about a specific entity or workspace memory. Use save_memory to persist durable memory. To update existing knowledge, search first, then save with supersedes_event_id so the old row is tombstoned rather than deleted.

MCP Client Setup

Use the actual MCP URL for the user's runtime. Never hardcode a hosted URL unless the user explicitly asks for that instance.

Common setup commands:

# Claude Code
lobu connect claude-code --url <mcp-url>

# Codex
lobu connect codex --url <mcp-url>

# Antigravity
lobu connect antigravity --url <mcp-url>

# OpenCode
lobu connect opencode --url <mcp-url>

Run lobu connect --url <mcp-url> without a client id to choose interactively. For ChatGPT, Claude Desktop, Cursor, and other browser-managed clients, the same command applies the safe local configuration it can and prints the exact remaining client handoff. Complete OAuth in the client, not through a separate Lobu CLI login.

Browser-Authenticated Connectors

For connectors that need a real browser session, browser-auth launches a dedicated Chrome with remote debugging, stores its CDP endpoint on the auth profile, and the connector attaches over CDP at sync time (harvesting cookies live):

lobu memory browser-auth --connector <key> --auth-profile-slug <slug>
lobu memory browser-auth --connector <key> --auth-profile-slug <slug> --check

Use --dedicated-profile only when you want a non-default dedicated Chrome profile directory; use --remote-debug-port to customize the CDP port (default 9222).

Agent-Facing Tool Surface

The normal MCP surface is intentionally small. Flat administrative tools are not advertised to agents; compose through these tools instead:

• search_memory and save_memory for direct knowledge recall and persistence.
• search_sdk to discover current ClientSDK methods, signatures, connector feed keys, operations, access requirements, and examples. Pass mode: "read" when you only need methods safe for query_sdk.
• query_sdk for read-only TypeScript and query_sql for governed, paginated SQL reads.
• run_sdk for mutations and external operations. Use lobu memory exec from the CLI for the same ClientSDK scripting workflow.

Search before create to avoid duplicates. Prefer a read or dry run before mutations, never fabricate returned URLs or identifiers, and never delete from events; supersede or tombstone instead.

Connectors, Feeds, and Operations

Connectors are the primary integration path for third-party services. The end-to-end model (entities vs events, the run lifecycle) is in docs/CONCEPTS.md; authoring a new connector when none matches is in docs/connector-authoring.md. Do not freeze a connector lifecycle in project instructions because it can drift from the deployed SDK:

1. Call search_sdk with the service name immediately before acting. It returns installed and installable connectors, declared feed keys, readiness, and the caller-aware live lifecycle.
2. Always follow the methods it returns in order instead of relying on a remembered connect/feed/operation sequence. Repeat discovery after authorization or state changes so the next step reflects the live connector and current caller.
3. Treat access: operate as mcp:write and access: administer as workspace owner/admin plus mcp:admin. If discovery asks for MCP reauthorization or an owner/admin handoff, follow that instruction; never attempt methods it withheld as unavailable.

Source-backed feeds query the provider only through an explicit client.feeds.readMany({ reads: [...] }) call; search_memory stays local and reports unqueried source feeds in coverage, while client.feeds.get returns metadata without source access. Feeds with sync may persist events for search and relational queries. For connector actions, use the live discovery result's operation target and readiness. If readiness is disconnected, follow the returned next action instead of guessing a connection. Surface approval-gated pending_approval as a waiting state, not a failure.

Data Ingestion

• Use lobu memory seed for small declarative YAML datasets under ./data; it processes records sequentially and is not a bulk-backfill path.
• Use client.knowledge.save for schema-less semantic history such as messages, notes, observations, and content. Pass supersedes_event_id when replacing an existing fact.
• Use client.entities.create / client.entities.update only for strict structured records that match declared entity schemas.
• For large backfills, send bounded chunks through run_sdk / lobu memory exec and use Promise.allSettled so conflicts and partial failures remain visible.

export default async (_ctx, client) => {
  const records = [
    { content: "...", semantic_type: "note" },
    { content: "...", semantic_type: "observation" },
  ];
  return Promise.allSettled(records.map((record) => client.knowledge.save(record)));
};