Little stone, big help
HiRocky User Guide
HiRocky is a local-first desktop AI agent workbench. This guide follows a practical order—from setup to daily use—so you can get productive quickly.
Quick start
Install the app, configure your first model, and send your first message.
-
Download and install
Pick the macOS, Windows, or Linux package on the home page. If downloads are not yet available, check back for updates.
-
First launch
The left sidebar lists main areas; the center shows the active page. We recommend opening Settings to configure a model first.
-
Configure a model provider
In Settings → Models, add an OpenAI-compatible endpoint: Base URL, API key, and model name. You can add multiple providers and set a default. Use Test to verify connectivity.
-
Send your first message
Open Chat, create a session, choose an agent, and type in the composer. Replies stream in real time.
API keys are stored on your device and are not uploaded to a HiRocky cloud by default.
Interface tour
The sidebar connects the modules you use every day.
- Chat: multi-session conversations with attachments and streaming.
- Agents: define roles, prompts, knowledge, and skills.
- Knowledge: import documents for retrieval in chat.
- Memories: view and edit long-term memory per agent.
- Skills: built-in and installed tool capabilities.
- Channels: Feishu, WeCom, WeChat personal, SaleSmartly, and more.
- Tasks: automation tasks (evolving).
- Stats / Logs: usage and troubleshooting.
- Settings: models, theme, language, and data locations.
Chat and agents
Sessions are where you work; agents define behavior and capabilities.
-
Create or pick an agent
On the Agents page, add a role with a system prompt and default model. Use prompt helpers to draft or refine instructions.
-
Manage sessions
Each session can bind an agent and workspace folder. Search, rename, or delete sessions. Long chats may trigger context compression with a recoverable summary.
-
Messages and attachments
Markdown is supported; attachments depend on agent capabilities. Stop generation anytime or regenerate a reply.
-
Normal vs expert view
Normal view focuses on the answer; expert view shows steps, tool calls, and timing—useful for debugging.
Use separate agents per project and bind the right knowledge bases.
Knowledge bases
Ground answers in your own documents.
-
Create a library
Add a knowledge base with a clear name (e.g. product docs, notes).
-
Import files
Supports TXT, Markdown, PDF, Word, PPT, Excel, RTF, and more. Text is extracted and indexed locally.
-
Use in chat
Bind libraries to agents so relevant chunks are retrieved automatically. Export libraries for backup.
Long-term memory
Let agents remember preferences and project context across sessions.
- Memories can be organized after conversations (behavior may vary by settings).
- Review, edit, or delete entries on the Memories page per agent.
- New sessions can reuse existing memories to avoid repeating yourself.
Memory stays on your machine—you can clear outdated entries anytime.
Skills
Tools the AI can call: files, browser, weather, presentations, and more.
- Built-ins: time/math, knowledge retrieval, filesystem, browser automation, weather, slides, controlled shell.
- Enable capabilities per agent; expert view shows each tool run.
- Install extensions from folders, ZIP archives, or URLs.
- File and shell skills are scoped to the session workspace; shell uses a whitelist.
Channels
Reply on Feishu, WeCom, WeChat, and other messaging surfaces.
-
Open Channels
Choose a platform and follow in-app instructions.
-
Authorize
Provide app credentials or scan to log in. WeChat personal credentials stay local. Feishu official APIs are recommended first.
-
Assign an agent
Route inbound messages to an agent so external chats use the same skills and knowledge.
Channel features evolve—refer to the latest in-app guidance.
Tasks, stats, and logs
Monitor usage and diagnose issues.
- Tasks: scheduled or triggered automation (in progress).
- Stats: sessions, messages, tokens, and channel volume.
- Logs: redacted local logs for safer debugging.
Settings and data
Personalize the app and find where data lives.
- Models: multiple OpenAI-compatible providers and defaults.
- Appearance: light or dark theme.
- Language: English and Simplified Chinese in the app UI.
- Data location: view database and file paths for backup.
Quick troubleshooting
Common fixes before asking for help.
- Cannot reach the model: verify URL, key, network, and model name; run Test in Settings.
- Slow or cut-off replies: rate limits or long context—try compression or a faster model.
- Skill failures: check expert view; confirm workspace path and permissions.
- Channels silent: refresh credentials and check status on the Channels page.