Horbot is a lightweight AI assistant stack focused on practical multi-agent orchestration, agent-specific workspaces, browser-based operations, chat channels, and persistent memory.
The project borrows and adapts ideas from several excellent open-source systems:
- Core lightweight agent patterns and implementation structure from HKUDS/nanobot
- Autonomous agent and self-improvement ideas from NousResearch/hermes-agent
- Three-layer memory ideas inspired by volcengine/OpenViking
- Earlier skill-system inspiration from OpenClaw
Horbot does not attempt to be a giant framework. The emphasis is:
- a readable agent loop
- practical Web UI operations
- agent-scoped workspace, memory, sessions, and skills
- compatibility with real operational workflows
- browser-persisted Web UI language switching for English, Simplified Chinese, and Thai
The screenshots below were captured from the current Horbot Web UI with the locale set to English.
| Dashboard | Chat |
|---|---|
 |
 |
| Skills | Teams |
 |
 |
- Create multiple agents with independent
provider,model, permission profile, workspace, memory, and skills - Build teams with ordered members, responsibilities, and lead assignment
- Support direct chat and team relay conversations in the same UI
- Let agents silently review completed work and distill reusable workflows into skill families with reference notes
- Let agents save user-requested reusable experience through a managed
save_skilltool instead of direct skill-path writes
- Maintain per-agent
SOUL.mdandUSER.md - Persist agent-scoped runtime metadata under
.horbot/agents/<agent-id>/workspace/.horbot-agent/ - Separate long-term memory, recent history, and reflection notes
- Keep team shared memory separate from private agent memory
- Markdown rendering for assistant messages
- Live Artifact cards for structured agent output such as dashboards, chart stories, map stories, process views, and interactive reports
- Zero-extra-token task workbench popover in the chat header, derived from loaded messages, attachments, execution steps, tools, and relay state instead of asking the model to repeat UI metadata
- Denser chat bubbles and tighter Markdown spacing to reduce blank space in long conversations
- Conversation history loads use no-store responses and incremental fallback recovery so refreshes or module switches do not reuse stale latest-message windows
- Inline preview for image, audio, PDF, Office, and text attachments
- High-fidelity PPTX preview through LibreOffice PDF export plus lazy per-slide PNG rendering
- Office document operations via OfficeCLI MCP, including structure-aware create/edit/view/validate flows for
.docx,.xlsx, and.pptx - Standalone remote image URLs in assistant history are promoted into normal image-card attachments when possible, instead of degrading to bare links
- Multi-image preview modals now include a bottom thumbnail strip for faster visual switching
- Drag-and-drop, paste upload, and retry flows
- Group chat history merge and recovery across legacy and current session paths
- Team relay timeline now shows clearer baton status such as who handed work to whom and whether the next turn is a continuation or a final summary
- When a direct-message agent dispatches work into a team relay, the Web UI now auto-switches into the team chat and then returns to the original direct chat once the final mirrored summary is ready, with a short-lived baton navigation banner at the top
- Multiple provider backends
- MCP integration
- Automatic OfficeCLI MCP wiring during local setup so agents can use Word, Excel, PowerPoint, and OpenXML-aware document tools without extra manual configuration
browser,web_search, andweb_fetchas the standard browser/search toolset, with Playwright as the default browser runtime- External Agent inbound-bot runtime for Feishu/Discord-style vendor or local agent platforms, plus compatibility adapters for OpenAI-compatible and generic HTTP/SSE/WebSocket agents
- Channels can also create Horbot inbound bot endpoints that issue App ID, Token, and an Inbound URL for platforms such as WorkBuddy to push messages into a fixed or request-specified internal Agent
- External channels such as WeCom, Feishu, ShareCRM, Telegram, Discord, Slack, Matrix, Email, Mochat, and others
- WeCom AI Bot support with reply-mode streaming, media upload, and inbound media download/decryption
- Web admin UI for configuration, agents, teams, status, skills, channels, tasks, and token usage
- Skills UI support for custom vs system grouping, auto-skill consolidation, promoting custom skills into built-in system skills, and focused skill graph visualization
- Agent-scoped skill graphs connect skills to references and related skills, auto-refresh after managed skill changes, and feed compact runtime hints back into agent skill summaries
- Configuration now exposes remote-image cache stats and a manual clear action for runtime-generated image-card cache files
- Built-in UI locale switcher for English, Simplified Chinese, and Thai, with the selection persisted in browser storage
- Smoke scripts for browser, chat, configuration, and agent asset flows
- Security defaults for local-only access and admin-token-gated remote access
- Web API endpoints are now split into focused route modules under
horbot/web/*_routes.py; keephorbot/web/api.pyas the composition, chat-core, and compatibility layer instead of adding unrelated feature routes back into it. - The Chat page keeps history parsing, file metadata, stream errors, execution-step mapping, and turn virtualization in focused modules under
horbot/web/frontend/src/pages/chat/.
git clone https://github.com/jenpeng/horbot.git
cd horbot
./horbot.sh install
./horbot.sh start./horbot.sh install now also installs OfficeCLI and checks LibreOffice. OfficeCLI is wired into ./.horbot/config.json for document operations; LibreOffice is used by the Web UI for high-fidelity PPTX preview.
To validate the Office document toolchain end to end, run ./horbot.sh smoke officecli. It directly checks .docx, .xlsx, and .pptx create/edit/view/validate flows with the locally installed officecli.
Default local URLs:
- Web UI: http://127.0.0.1:3000
- Backend API: http://127.0.0.1:8000
Common commands:
./horbot.sh status
./horbot.sh restart
./horbot.sh install officecli
./horbot.sh install libreoffice
./horbot.sh check libreoffice
./horbot.sh smoke officecli
./horbot.sh logs backend
./horbot.sh smoke chat-workbench
./horbot.sh smoke browser-e2e- Documentation Index
- Architecture
- API
- User Manual
- Multi-Agent Guide
- Skills
- Security
- Contributing
- Changelog
The current runtime model is agent-scoped:
.horbot/
├── agents/
│ └── <agent-id>/
│ └── workspace/
│ ├── SOUL.md
│ ├── USER.md
│ ├── .audit/
│ ├── .checkpoints/
│ ├── .state/
│ └── .horbot-agent/
│ ├── memory/
│ ├── sessions/
│ └── skills/
├── teams/
│ └── <team-id>/
│ ├── workspace/
│ ├── shared_memory/
│ └── taskboard/
├── data/
│ ├── uploads/
│ ├── sessions/
│ ├── plans/
│ └── cron/
└── runtime/
├── logs/
└── pids/
Legacy .horbot/context and .horbot/memory directories are no longer part of the active memory model and can be removed from existing local environments.
Legacy agent paths such as .horbot/agents/<agent-id>/{memory,sessions,skills} and workspace/{memory,skills} are treated as migration inputs only. Current builds merge forward from those paths but do not recreate them.
Horbot is released under the MIT License.
This license allows commercial use, modification, distribution, private deployment, and resale, provided that the original copyright notice and license text are retained.