Capabilities

This page explains how Lobu capabilities run on top of the OpenClaw runtime and Lobu’s API runtime.

See also: Comparison, Skills, and Providers.

Runtime architecture

Lobu capability execution has two cooperating runtime layers that split agent logic from platform infrastructure.

OpenClaw runtime (worker layer)

Each worker runs an OpenClaw session — the open-source agent runtime that handles the core execution loop. OpenClaw manages:

• Session loop — receives a prompt, plans tool calls, executes them, and streams results back.
• Tool execution — the 7 core tools (read, write, edit, bash, grep, find, ls) run directly inside the OpenClaw runtime. The agent decides which tools to call based on the user’s request.
• Plugin hooks — OpenClaw’s plugin system (pluginsConfig) lets you swap in memory backends, custom tool sets, and lifecycle hooks without changing Lobu’s code.
• Workspace isolation — each session operates inside a sandboxed workspace directory. Files persist across turns within the same thread.

Workers never see real credentials. They receive opaque placeholder environment variables; the gateway proxy resolves actual secrets at request time.

Lobu API runtime (gateway layer)

The gateway sits between users and workers, handling everything outside of agent execution:

• Platform adapters — receives messages from Telegram, WhatsApp, Slack, or the REST API and normalizes them into a common job format.
• Provider proxying — routes LLM API calls through the gateway proxy, injecting the user’s provider credentials per-request. Workers only know placeholder URLs.
• Auth resolution — manages provider credentials and MCP server authentication. Third-party API auth (GitHub, Google, etc.) is handled by Owletto.
• Scheduling — ScheduleReminder and CancelReminder calls are handled by the gateway’s job scheduler, which triggers worker sessions at the scheduled time.
• Policy enforcement — network allowlists, tool permissions, and package approvals are enforced at the gateway layer before reaching the worker.
• Custom tools — the Lobu-specific tools (UploadUserFile, AskUserQuestion, etc.) are API calls from the worker back to the gateway.

This split means agents evolve via config and user approvals — no new integration codepath needed for each provider or skill.

Interactive capability demos

The demos below use the same settings panels and chat UI components used on the homepage.

Platform Connections

Platforms are connected via the admin page — paste a bot token and the agent is live. No env vars or config files.

Settings panel behavior

L

Connections

🔗Connections

Support

Alice Chen8291045832

Infra Team-100204817

Internal Marketing

#campaigns109284710

+ Add connection

Chat behavior (same component as homepage)

L

Lobu

online

Hey, I just added you on Telegram!12:01

I'm live! Your admin connected me in the admin page — no config files needed. How can I help?12:01

What platforms do you support?12:01

Telegram, Slack, Discord, WhatsApp, and Teams. Your admin can add more anytime from the admin page.12:01

Example prompts: Hey, I just added you on Telegram! | What platforms do you support?

Model Setup and Provider Selection

Agent asks for setup, user picks provider/model in settings, then the same conversation continues immediately.

Settings panel behavior

L

Models

🤖Models

Groq

Connected

llama-3.3-70b-versatile

Gemini

Connected

gemini-2.0-flash

Together AI

Connected

meta-llama/Llama-3.3-70B-Instruct-Turbo

NVIDIA NIM (free)

Connected

nvidia/moonshotai/kimi-k2.5

z.ai

Connected

Auto model

ElevenLabs

Connected

Auto model

Fireworks AI

Connected

accounts/fireworks/models/llama-v3p3-70b-instruct

Mistral

Connected

mistral-large-latest

DeepSeek

Connected

deepseek-chat

OpenRouter

Connected

anthropic/claude-sonnet-4

Cerebras

Connected

llama-3.3-70b

OpenCode Zen

Connected

anthropic/claude-sonnet-4

xAI

Connected

grok-3

Perplexity

Connected

sonar-pro

Cohere

Connected

command-r-plus

OpenAI

Connected

gpt-4o

Chat behavior (same component as homepage)

M

Internal Marketing

online

Help me write a blog post12:01

I need an AI model first. Set me up below.12:01

Open Settings

Done, let's go12:01

Here's an outline: 1. Hook with a question 2. Key insight + examples 3. Clear takeaway Expand any section?12:01

Example prompts: Help me write a blog post | Use groq/llama-3.3-70b for this task

Network and Security Permissions

Blocked domains trigger an approval flow. After approval, the agent retries with updated network policy.

Settings panel behavior

L

Security

🛡️Domain allowlist

github.comAlways

registry.npmjs.orgAlways

🔌MCP proxy

gmail-mcpSecrets proxied

github-mcpSecrets proxied

🔧Allowed tools

BashSandboxed

file_uploadAllowed

Chat behavior (same component as homepage)

I

Support

online

Clone my repo and install dependencies12:01

Can't reach github.com — not in allowed domains. Grant access?12:01

Allow for 1 hour

Allow it permanently12:01

github.com added to always-allow list. Cloned the repo and ran npm install. Ready to go.12:01

Example prompts: Clone my repo and install dependencies | Allow github.com and rerun the command

Skills and Integrations

Agent proposes skill installation with required integrations, user approves once, then the capability is active.

Settings panel behavior

L

Skills

⚠️

Pending changes from your agent

Review and approve the requested configuration changes.

⚡Skills

skillOps Triage

system/ops-triage

Triage inbox, PRs, and issues

oauthGoogle

oauthGitHub

networkgmail.googleapis.com, api.github.com

Approve AllDismiss

Chat behavior (same component as homepage)

O

Coding

online

Install my triage skill and summarize inbox + PRs12:01

Found `ops-triage`. Needs: • Gmail MCP (OAuth) • GitHub MCP (OAuth) • Linear integration (API key) Install and connect now?12:01

Install Skill

Yes, connect everything12:01

Done. OAuth connected for Gmail/GitHub, Linear API key saved. Summary: • Inbox: 3 urgent emails • PRs: 2 need review • Linear: 5 high-priority issues12:01

Example prompts: Install ops-triage and summarize inbox + PRs | Connect Gmail and GitHub now

Persistent Memory Runtime

Memory plugins persist context between sessions and auto-recall relevant facts on new turns.

Settings panel behavior

L

Memory

🧠Memory Plugins

owletto-memorydefault

Enabled

pgvectorBM25 searchentity systemauto-recallauto-capture

native-memoryoptional

Disabled

Chat behavior (same component as homepage)

R

AnalystBot

online

Remember that my weekly report should focus on customer churn12:01

Saved to persistent memory via Owletto (default memory plugin). I will use this preference in future reports.12:01

Can I use a different memory backend later?12:01

Yes. You can switch to another OpenClaw memory plugin, such as native memory, by changing `pluginsConfig`.12:01

Example prompts: Remember that weekly reports should focus on churn | What preferences do you remember about my reports?

Reminders and Recurring Jobs

Natural language schedule requests become managed jobs with approval, status, and cancellation support.

Settings panel behavior

L

Schedules

⏰Reminders

Check open PRs and summarize review queue

pendingrecurringEvery Mon 9:00 AM

Cancel

Review Q1 deck

pendingTomorrow 2:00 PM

Cancel

Chat behavior (same component as homepage)

T

Personal Assistant Bot

online

Every Monday 9am, check my PRs12:01

⏰ Mon 9 AM — check open PRs Approve?12:01

Approve Schedule

Yes. Also remind me tomorrow 2pm to review Q1 deck12:01

Two schedules active: 1. PR summary — Mon 9 AM (recurring) 2. Q1 deck — Tomorrow 2 PM (one-off) I'll ping you when they trigger.12:01

Example prompts: Every Monday at 9am check my open PRs | Remind me tomorrow at 2pm to review the deck

On-demand System Packages

Agent requests package installation, user approves, and tools become available in a reproducible environment.

Settings panel behavior

L

Packages

📦Packages

ffmpegUninstall

gifsicleUninstall

imagemagickUninstall

Chat behavior (same component as homepage)

N

Internal Marketing Bot

online

Convert this video to a gif under 5MB12:01

I need ffmpeg and gifsicle. Install them?12:01

Approve Installation

Go for it12:01

Installed via Nix. These persist across sessions. Converting now...

Done! 4.2 MB (from 28 MB). Adjust frame rate?12:01

Example prompts: Convert this video to a gif under 5MB | Install ffmpeg and gifsicle first

Core runtime tool overview

From packages/worker/src/openclaw/tools.ts. These tools run directly inside the OpenClaw runtime.

Tool	What it does	Typical use
read	Reads file contents from the workspace.	Inspect source/config/log files before editing.
write	Writes full file contents.	Create new files or replace file content entirely.
edit	Applies targeted text replacements in a file.	Small, surgical code changes.
bash	Runs shell commands in the workspace (policy-controlled).	Build, test, lint, run scripts, inspect runtime state.
grep	Searches file contents with pattern matching.	Find symbols, config values, or error strings quickly.
find	Finds files/directories by path patterns.	Locate files across large repositories.
ls	Lists files/directories.	Quick workspace structure discovery.

Lobu custom tool overview

From packages/worker/src/openclaw/custom-tools.ts. These tools call back to the Lobu API runtime.

Tool	What it does	Typical use
UploadUserFile	Uploads a generated file back to the user thread.	Share reports, charts, documents, exports, media.
ScheduleReminder	Schedules one-time or recurring follow-up tasks.	Deferred tasks and recurring automations.
CancelReminder	Cancels a scheduled reminder by ID.	Stop previously scheduled jobs.
ListReminders	Lists pending reminders and schedule IDs.	Audit or pick reminder to cancel/update.
SearchSkills	Searches for skills and MCP servers, or lists installed capabilities (empty query).	Discover capabilities and check installed state.
InstallSkill	Installs or upgrades skills from the registry with bundled dependencies.	Add capabilities (skills, providers, MCPs) via settings approval.
InstallPackage	Requests system package installation (nix) with user approval.	Install ffmpeg, imagemagick, or other system tools.
RequestNetworkAccess	Requests domain access grants with user approval.	Unblock API endpoints for curl/fetch calls.
GenerateImage	Creates images from text prompts and uploads them in-thread.	Illustrations, posters, product mockups, and visual concepts.
GenerateAudio	Converts text to speech and returns audio.	Voice responses or spoken summaries.
GetChannelHistory	Fetches prior messages in the conversation thread.	Recover context from earlier discussion.
AskUserQuestion	Sends structured button-based questions to the user.	Branching choices and approvals without free-text ambiguity.

How each tool works

Each section below shows the tool description and an example conversation demonstrating how the agent uses it.

Core runtime tools (OpenClaw)

These run inside the OpenClaw session. The agent calls them directly without leaving the worker sandbox.

read

Reads file content from the current workspace path. It is used before edits to inspect source and config safely.

A

Agent

online

Open `src/server.ts` and summarize what it does.12:01

Here's src/server.ts — an Express app on port 3000 with 3 routes: /health, /api/users, /api/auth. The auth middleware checks JWT tokens from the Authorization header.12:01

write

Writes full file content. Best for creating new files or replacing an entire file when structure changes a lot.

A

Agent

online

Create `scripts/release.ts` with a release helper.12:01

Created scripts/release.ts with version bump, changelog generation, and git tag helpers. Run it with `bun run scripts/release.ts`.12:01

edit

Applies targeted replacements to existing files. It is optimized for small surgical changes with minimal diff.

A

Agent

online

Rename `oldFlag` to `featureFlag` in `src/config.ts` only.12:01

Done. Renamed oldFlag → featureFlag in src/config.ts (3 replacements applied).12:01

bash

Runs shell commands in the agent sandbox under your policy rules. Used for build, test, lint, and automation tasks.

A

Agent

online

Run tests for `packages/gateway` and report failing cases.12:01

42 passed, 1 failing: auth.test.ts — token expiry assertion off by 1s. Want me to fix the flaky test?12:01

grep

Searches text patterns in code and logs. It helps the agent quickly locate symbols, errors, and references.

A

Agent

online

Find all usages of `sendToWorker` in this repo.12:01

Found 7 matches in 4 files: • routes/jobs.ts:42,87 • orchestration/queue.ts:15 • services/worker.ts:33,58,91 • tests/worker.test.ts:1212:01

find

Searches for files and directories by path pattern. Useful in large monorepos before reading files.

A

Agent

online

Find all `docker-compose` files in this project.12:01

3 compose files found: • docker/docker-compose.yml • docker/docker-compose.dev.yml • docker/docker-compose.test.yml12:01

ls

Lists directory contents to inspect structure, generated artifacts, and available scripts.

A

Agent

online

List files under `packages/worker/src/openclaw`.12:01

packages/worker/src/openclaw/ • processor.ts • tools.ts • custom-tools.ts • worker.ts • plugins/12:01

Lobu custom tools (API runtime)

These are API calls from the worker to the gateway. Auth, secrets, and platform delivery are handled by the gateway.

UploadUserFile

Sends generated files back to the user chat thread.

A

Agent

online

Create a CSV report and upload it.12:01

Report generated (24 rows, 3 columns). Uploading churn-report.csv to your thread… File uploaded.12:01

ScheduleReminder

Creates one-time or recurring reminder jobs from natural-language instructions.

A

Agent

online

Remind me every Monday at 9am to review PRs.12:01

Scheduled: Mon 9 AM — review open PRs Recurrence: weekly ID: sched_a1b212:01

CancelReminder

Cancels an existing reminder by schedule ID.

A

Agent

online

Cancel reminder `sched_123`.12:01

Cancelled reminder sched_123. It was: "Check deployment status" — daily at 6 PM.12:01

ListReminders

Lists active reminders so users can audit or cancel them.

A

Agent

online

Show my pending reminders.12:01

2 active reminders: 1. sched_a1b2 — Mon 9 AM: review PRs (weekly) 2. sched_c3d4 — Tomorrow 2 PM: review Q1 deck (one-off)12:01

SearchSkills

Searches discoverable skills/MCPs and lists currently installed capabilities.

A

Agent

online

Find a skill for on-call incident triage.12:01

Found 3 matching skills: 1. ops-triage — incident triage with PagerDuty + Slack 2. on-call-helper — on-call runbook automation 3. alert-summarizer — alert digest and dedup12:01

InstallSkill

Installs or upgrades skills from the registry, bundling their provider/MCP dependencies into a settings approval flow.

A

Agent

online

Install the ops-triage skill.12:01

Installing ops-triage… Needs: Gmail (OAuth), GitHub (OAuth), Linear (API key) Open settings to connect?12:01

Install Skill

InstallPackage

Requests system package installation (nix) with inline user approval buttons.

A

Agent

online

I need ffmpeg to process this video.12:01

Package install request sent. Click below to approve.12:01

Approve ffmpeg

RequestNetworkAccess

Requests access to blocked domains with inline user approval buttons.

A

Agent

online

I need to call the GitHub API.12:01

Domain access request sent for api.github.com.12:01

Grant Access (1h)

GenerateImage

Generates an image from a prompt and sends it back in the chat thread.

A

Agent

online

Create a minimalist onboarding illustration with teal accents.12:01

Image generated and uploaded to the thread.12:01

GenerateAudio

Converts text responses into audio output.

A

Agent

online

Read this summary aloud and send audio.12:01

Audio generated (1:42). Sending voice message…12:01

GetChannelHistory

Loads earlier thread messages to recover context.

A

Agent

online

Review earlier thread context before answering.12:01

Loaded 12 earlier messages. Context recovered — you previously asked about deployment timelines and CI config.12:01

AskUserQuestion

Sends structured button-based prompts for explicit human-in-the-loop decisions.

A

Agent

online

Ask user whether to run full migration or dry run.12:01

Before proceeding: Run full migration or dry run first?12:01

Full MigrationDry Run

Additional tool sources

Beyond built-ins, agents can also receive tools from:

• OpenClaw plugins (pluginsConfig) loaded at runtime
• Configured MCP servers (proxied through the gateway)

Memory plugin runtime

Lobu uses OpenClaw’s plugin system for memory. The default plugin is Owletto (@lobu/owletto-openclaw, slot memory).

You can switch to another OpenClaw memory plugin, such as @openclaw/native-memory, via pluginsConfig.