LibrAgent
A local agent operating environment. Select your model, install tools with one click, and choose a coordination pattern.
You do not need a vendor harness or manual JSON configuration. The software saves completed work as files on your machine.
한국어 | 简体中文 | 日本語 | Français | Español | Deutsch | Português
HD WebM · Install tools like apps. Keep your model. Keep the file.
How LibrAgent is different
Many agent harnesses require you to edit JSON configuration files, use command-line terminals, and assemble code frameworks. Other products lock you into a single vendor stack.
LibrAgent provides a desktop product for these tasks:
| Traditional Setup | With LibrAgent |
|---|
| Edit MCP configuration files manually | Extensions — One-click presets (GitHub, Brave Search, Filesystem) and import from Cursor, VS Code, Claude Code, and Windsurf |
| Assemble multi-agent systems from code | Standard coordination patterns as bundled skills (pipeline, hub-spoke, divide-conquer, consensus-delegation) |
| Restricted to one provider model and toolset | Choose your LLM (API keys or local Ollama) and your tool stack with MIT, local-first architecture |
Download the latest release · 5-minute onboarding
What You Can Do in the First 10 Minutes
1. Install tools with one click, then generate a deliverable
- Open Extensions and install a preset (such as GitHub) without manual JSON editing.
- Point Workspace to a local directory on your machine.
- Submit your request: "Review this repo for the top risk for a new contributor and save
DELIVERABLE.md"
2. Run an automated workflow recipe
- Launch the Morning Briefing recipe from Chat home or Scheduled Tasks.
- The recipe installs Hacker News and Yahoo Finance presets, configures an assistant, and schedules execution for 09:00 daily.
- Review your synthesized technology and market briefing automatically.
3. Select a coordination pattern without code assembly
- Describe your workflow, or attach a skill directly by name:
- "@skill:pipeline — research, draft, and review; generate one final report"
- "@skill:divide-conquer — divide this task into independent steps and merge the results"
- Coordination patterns are packaged skills that do not require external SDK setup. See Sub-agents & orchestration.
4. Maintain control of your models
- Cloud: Paste an API key for OpenAI, Anthropic, Gemini, or Groq.
- Local: Run
ollama pull qwen3:14b and select Ollama in settings to use the same desktop environment locally.
Coordination patterns (bundled skills)
Select a pattern that matches the structure of your task, then execute it from the chat interface:
| Skill | Pattern | When to use |
|---|
pipeline | Sequential stages | Output from each step supplies the next step (research → draft → review) |
hub-spoke | Hub-and-spoke | One coordinator delegates tasks to multiple specialized workers |
divide-conquer | Parallel split | Divide work into independent tasks, then merge the results |
consensus-delegation | Multi-perspective | Send the same question to multiple specialists, then reconcile differences |
gatekeeper / pair-programming | Review loops | Enforce strict code review or collaborative two-agent programming |
delegate | Lightweight handoff | Run one child session with lineage tracking |
teamwork → org | Persistent team | Create durable multi-agent teams with shared guidelines |
schedule / loop / call-me-back | Time & events | Execute cron schedules, in-session delays, or resume on process events |
Selection heuristics: framework-selection · Full guide: Sub-agents.
Additional default skills: setup-wizard, mcp-installer, playbook-creator, and more — Bundled Skills.
MCP platform (advanced capabilities)
- Transports: stdio, HTTP, SSE, OAuth 2.1
- 15+ built-in servers (Workspace, Shell, Browser, Knowledge, Planning, Scheduled Tasks, and others)
- One-click presets and agent-assisted installation (
mcp-installer)
- Tool isolation per session, file path guards, command guards, and optional automated execution modes
Execution substrate
| Substrate | Capabilities |
|---|
| Workspace | Line-precise editing, multi-file operations, and @file / @skill / @playbook context |
| Shell | Isolated and persistent terminal sessions with asynchronous process monitoring |
| Browser | Built-in agent browser by default. Optional Chrome extension when the agent needs sites where you are already signed in (guide) |
| Knowledge | Graph-based knowledge store and BM25 search |
| Export | Markdown reports and ATIF trajectory exports (session export) |
Long-running sessions maintain stability through context compaction, loop prevention, circuit breakers, and stale-response protection.
Real-World Scenarios
Operator — Automated daily briefing
- Run the Morning Briefing recipe (installs presets, sets up assistant, schedules execution at 09:00).
- Click Run now to verify the configuration.
- Receive the generated briefing report directly in the application without using a terminal.
Solo developer — Direct tool installation without configuration files
- Open Extensions and install the GitHub MCP preset.
- Connect a local repository using the Workspace selector.
- Prompt the agent to generate a Markdown security and code review report saved to disk.
Non-framework power user — Standard orchestration patterns
- Enter
@skill:pipeline (or hub-spoke / divide-conquer) to specify workflow structure.
- Agents coordinate according to the selected pattern.
- Review the consolidated deliverable in your workspace without maintaining external framework code.
Privacy-sensitive team — Local models with zero cloud transmission
- Run
ollama pull qwen3:14b.
- Keep Workspace and Shell interactions strictly on your local machine.
- Add cloud API keys later if necessary; the application workflow remains unchanged.
Documentation
Getting Started
Download the latest installer from the Releases page.
The 5-minute onboarding path
Step 1 — Connect a model (Settings → LLM Providers)
- Cloud: Paste an OpenAI, Anthropic, Gemini, or Groq API key.
- Local: Run
ollama pull qwen3:14b, then select Ollama in Settings.
Step 2 — Add tools without manual configuration
- Open Extensions and install a preset (such as GitHub), or
- Instruct the agent: "Import my MCP servers from Cursor"
Step 3 — Attach a workspace and generate files
- Select a local project directory in Workspace.
- Instruct the agent: "Review this workspace, then write findings to
DELIVERABLE.md."
Next — Coordination and automation
- "@skill:pipeline — research, draft, then review; save one final report."
- "Prepare a teamwork workspace for this repo."
- "Set up a scheduled daily competitor brief at 7am." (or launch the Morning Briefing recipe)
First prompts to copy-paste
- "Import my MCP servers from Cursor and show me what was added."
- "Install the GitHub MCP preset and attach it to a coding agent."
- "Review this workspace, then write findings to
DELIVERABLE.md."
- "@skill:pipeline — research this topic, draft a summary, then review; save the final report."
- "Set up a scheduled daily competitor brief at 7am."
Developer setup
git clone https://github.com/fritzprix/libr-agent
cd libr-agent
pnpm install
pnpm tauri dev
Design Philosophy
- Product over kit: The runtime is ready for immediate use without manual code assembly.
- Orchestration as skills: Coordination patterns are named, selectable, and documented.
- Independence of stack: Models and tools remain user choices; the software does not depend on a single AI vendor.
- Local First: Workspaces, sessions, skills, and browser states remain on your machine. Cloud LLM and remote MCP connections occur only when explicitly enabled.
- Harness over Model: Tools, session state, delegation, and governance provide more value than any individual model.
- Stability over Features: Isolation, context compaction, and loop prevention take precedence over feature quantity.
- Open Standards: MIT license with MCP as the foundational interoperability layer.
Contributing & License
LibrAgent is MIT licensed and built in the open. Contributions are welcome, including bundled skills, MCP integrations, bug fixes, and architectural improvements.
License: MIT
