Skip to content
Projects

CCDawn/Vibelution has 79 GitHub stars and 10 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-18 11:50 UTC. Last push 2026-11-18.

How these counts are measured

Measured star growth

in 24h
0
in 7 days
Collecting history
in 15 days
Collecting history
in 1 month
Collecting history

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

Local-first AI agent workbench: research teams that plan, build and review together, a finance agent, and virtual human companions — in one Electron desktop app.

AI

README.en.md

Original
Vibelution — Agents collaborate. Knowledge grows. Your local AI Agent workbench.

CI License MIT GitHub release Platform Windows macOS Linux

中文 · English
Quick start · Feature tour · First conversation · Documentation
Local workbench   /   Local or cloud models

Work loop demo: organize tasks on the research canvas, advance them in team discussion, and distill them into a knowledge graph
A real work loop across the research canvas, team discussion, and knowledge graph

One team, distinct roles

Assign planning, development, and review to different Agents. See their responsibilities on one canvas, coordinate through group chat and member messages, and retain discussion and handoff records.

Team role illustration: the planner assigns work to two developers, who submit it to a reviewer.

Role illustration based on the team screenshot, not a product interface. See the actual team canvas ↓

From everyday conversations to ongoing collaboration

What you want to do In Vibelution
Start a conversation or coding task Organize sessions by Agent and team; inspect tool activity and workspace artifacts.
Work as a team Define member roles and coordinate through the canvas, group chat, and member sessions.
Get help from a personal assistant Organize daily tasks and goals, check platform status, and manage Agents and sessions through authorized tools.
Keep material for next time Manage private memory and formal knowledge; browse topics and trace sources.
Improve through verification Follow baselines, candidates, reevaluation, review, and runtime activation.
Conduct financial research Organize research, market data, simulated assets, and reports in a dedicated workspace.
Spend time with a virtual companion Chat in character scenes and configure life and memory for plugin-enabled Agents.

Start your first conversation

Start with the Windows desktop workbench. Install Python 3.11+ (3.12 recommended), Node.js 22.12+, and Git, then run in PowerShell:

git clone https://github.com/CCDawn/Vibelution.git
cd Vibelution
powershell -ExecutionPolicy Bypass -File scripts/install_windows.ps1
  1. Open Vibelution Launcher from your desktop. Follow the model configuration guide to configure a provider, model, and credential reference.
  2. Open Chat, create or select an Agent, and start a new session.
  3. Send your first message. Once you receive a reply, try files, tools, or team tasks.

First conversation → · Windows installation and troubleshooting · macOS (Chinese for now) · Linux (Chinese for now)

macOS and Linux use the browser workbench. Installation currently requires the runtimes above; a portable installer is still planned. The workbench runs locally. Cloud model requests are sent to the selected provider and may incur charges.

The actual team canvas

The member list, organization canvas, and member details share one workspace. Select a node to inspect its role, and use group chat or member messages for discussion.

Actual development team interface: member list, four-role canvas, and member details

Original user-provided screenshot; click to enlarge. The canvas shows organizational relationships, with historical task records on the right. Product screenshots show the Chinese interface.

01 / Start with one session

Choose different Agents for different tasks, keeping everyday questions, project discussions, and coding work in their own sessions. Team groups and multiple tabs help you return to the work you care about.

Chat workbench: Agent and team groups, session tabs, and composer

Organize by task: browse Agents, teams, and sessions together. Follow the work: inspect answers and tool activity, then continue with workspace files, previews, and the terminal.

The screenshot shows a new session and its navigation and composer, not a completed task.

Chat and workspace · Tool authorization

02 / A personal assistant for your tasks and goals

Bring daily tasks, near-term goals, and work across the platform into one conversation. Your first visit to the Personal assistant page initializes a dedicated Agent and session. Configure a model to start chatting.

What you can ask for Available capabilities
Organize tasks and goals Discuss your to-do list, break down the next step, and record preferences and goals in the assistant's own private memory.
Check platform status View Agents, sessions, activity, and items awaiting confirmation; filter by status and open the relevant session.
Manage on request Create, update, or archive Agents; create, stop, or delete sessions. Management actions go through the platform's tool approval flow.
Keep chatting on the desktop Use a floating window on Windows. The page and desktop mode share one assistant identity, session, and memory.

Try asking: "Check which tasks are running, awaiting confirmation, or in an error state, and suggest what I should do. Observe first; do not make changes."

The floating window requires the Windows desktop workbench. Model configuration · Tool authorization

03 / Keep knowledge connected to its sources

Agent-private memory, team material, and formal knowledge have separate management surfaces. Browse the knowledge overview by topic, then open individual materials and their sources to carry useful context into your next task.

3D knowledge graph: 24 formal knowledge items in the computing physics and energy topic

Find material through topics: browse groups and search knowledge and sources. Retain evidence: formal knowledge supports proposals, review, and source management; private memory keeps its own access boundaries.

Selecting Dennard scaling reveals its summary, ownership, revision and recorded source relations

Captured from the production page using an isolated copy of local formal research knowledge. Spatial positions aid navigation; edges preserve recorded relations and imply no additional causal claims.

Memory and retrieval · Team knowledge

04 / Review the evidence behind each improvement

Self-evolution checks, changes, and verifies work against a goal. Supervised evolution brings the baseline, candidate, and independent reevaluation into one process, with review deciding whether to integrate the result.

Supervised evolution: Agent review conversation, candidate changes, and stage progress

Inspect the review: retain Agent conversations, candidate changes, and reevaluation records. Distinguish stages: evaluation, approval, integration, and runtime activation are separate; one successful stage does not complete the entire process.

The score belongs only to the displayed evaluation case. The interface still reports candidate activation without runtime-version evidence; it is not a claim that general capability improved from 30 to 99. Continuous self-evolution requires user approval and stop conditions.

Evolution modes · Evolution configuration

05 / A dedicated financial research workspace

Research stocks and topics with analyst collaboration, market data, simulated accounts, portfolio research, and report materials in one place, reducing the need to move context between tools.

Financial research workspace: research, market, asset, and material navigation with a research chat pane

Work area Available entry points
Research and collaboration Stock research, topic research, analyst collaboration, research tasks
Market data and simulated assets Watchlist quotes, stock screening, simulated accounts, portfolio research, historical backtests, trade review
Materials and learning Reports, research memory, skills, learning

The screenshot shows the workspace before research starts. Market data and research require the relevant services, data, and model configuration. Simulated accounts and historical backtests are not live trading; broker trade execution is not included.

Financial workspace · Financial report service setup

06 / There is room for everyday life, too

Continue everyday conversations in character scenes. Configure identity, schedules, emotions, and long-term memory for Agents with the virtual-human plugin explicitly enabled.

Virtual companion: character scene, life status, and everyday conversation Desktop pet idle animation: a recorded Live2D idle loop

Character and scene: an immersive interface with life status and conversation together. Conversation continuity: the plugin provides proactive-contact, unfinished-topic, and memory mechanisms, subject to enablement and configuration.

Left: the character scene screenshot. Right: a recorded idle animation of the desktop pet. Proactive scheduling and animated presentation depend on runtime support. Character names and third-party assets belong to their respective rights holders.

Virtual-human plugin

Configure the workbench your way

Supporting capability Purpose
Agent configuration Configure identity, models, prompts, and tool permissions separately.
Tools, Skills, and MCP Inspect tool authorization, browse local skills, and configure managed MCP access.
Git and file review Inspect repository status, file differences, and commits to understand task changes.
Models and usage Configure model services and inspect usage and local model status.
Logs and runtime diagnostics Inspect execution records and investigate problems.

Models and usage · MCP access · Git configuration · Local model monitoring

Documentation and participation

Contributing · Documentation map · Interactive architecture map (offline single file, about 14 MB; download and open in a browser) · Testing guide · Changelog · Report an issue · Security

Related projects: briefbound-skills · Codex-Dynamic-Skin · harmony-codex · pc-touchpad

MIT code license · Third-party components

Media sources and notes. Screenshots show feature entry points and their state at capture time; they do not replace functional acceptance. The code license does not grant rights to third-party character assets.