---

What AgentHive Is

AgentHive is not a fixed set of chatbot personas. It is a local AI roundtable runtime for strategy discussions. You can place an idea, meeting note, business context file, or decision question into one discussion space, then let different agents contribute across multiple rounds according to their roles.

It is useful for:

• Product direction debates: bring product, market, execution, dissent, and synthesis agents into one room.
• Strategy reviews: load meeting notes, sales data, or research material into context/, then ask agents to separate facts from unverified assumptions.
• Pre-decision pressure tests: ask dissent, fact-pack, policy research, and market research agents to surface risks and evidence gaps first.
• Agent Skill prototyping: create, replace, or remove role templates through roundtable-skills/.

Important: the executive-style agents, fact-pack agents, policy research agents, market research agents, and summary agents in this repository are replaceable templates. Users can create, customize, replace, or delete any agent.

Project Profile

Dimension	Details
Product form	Locally running Roundtable Web workspace
Core objects	Configurable agents, Skills, context files, discussion topics, evidence requests, synthesis output
Default frontend	Next.js 16, React 19, TypeScript
Default entry	http://localhost:3001/roundtable
License	Apache License 2.0
Data policy	Real business materials stay in local context/ and are not committed by default

Screenshots

Roundtable Discussion

Agent Registry

Core Capabilities

Capability	Details
Multi-agent roundtable	Coordinate executive-style agents, expert agents, evidence agents, dissent agents, and synthesis agents.
Configurable roles	Each agent can have its own name, avatar, responsibilities, participant status, background-research mode, and Skill document.
Skill templates	roundtable-skills/ provides role templates that can be copied and rewritten.
Context import	Load .md, .txt, .json, .jsonl, .csv, and similar files from local context/.
Evidence requests	Track facts, unverified assumptions, uncertainty, and follow-up validation tasks during discussion.
Discussion export	Export the full discussion record and decision-oriented synthesis.

Discussion Workflow

Input topic / context material
        |
        v
Select or edit participating agents
        |
        v
Multi-round discussion: views, challenges, evidence requests, risks
        |
        v
Fact-boundary cleanup and unverified-claim labeling
        |
        v
Export discussion record and decision synthesis

Quick Start

Option 0: Windows desktop app (no command line)

Download either asset from GitHub Releases:

• AgentHive-Setup-<version>-x64.exe: installer edition. Complete the setup wizard, then launch AgentHive from the desktop or Start menu.
• AgentHive-Portable-<version>-x64.zip: portable edition. Extract it, keep the AgentHive folder intact, and run AgentHive.exe.

The desktop app starts an internal service on a random loopback address. It does not require or expose a fixed localhost URL. Context, copied role Skills, discussion history, and logs remain in the current Windows user's app-data directory rather than the installation folder.

Option 1: Windows launcher

Run from the repository root:

.\Start-Roundtable-UI-Only.cmd

Then open:

http://localhost:3001/roundtable

Option 2: Manual start

cd workspace/frontend
npm install
npm run dev

Default URL:

http://localhost:3001/roundtable

Runtime Dependencies

The desktop app bundles Electron and the frontend runtime, but it does not bundle model-provider CLIs, login state, API keys, or Tavily credentials. To run the default role set, install and authenticate these dependencies under the same Windows user:

Dependency	Required	In-app use	Health check
Codex CLI	Yes	Powers roles labelled Codex CLI	codex --version
Claude Code CLI	Yes	Powers roles labelled Claude Code CLI	claude --version
Tavily CLI	Yes for web evidence research	Lets background evidence agents retrieve public-web sources with traceable evidence	tvly --status

Install and sign in to Codex CLI and Claude Code CLI using their official guides. For public-web evidence research, install Tavily CLI with uv tool install tavily-cli or pip install tavily-cli, then run tvly login or configure TAVILY_API_KEY. Missing Codex or Claude prevents the roles assigned to that runtime from generating turns. Missing Tavily still allows local context/ materials, but disables public-web evidence collection.

In the desktop app, press Alt to reveal the menu, then choose AgentHive > Runtime Dependencies to confirm the three CLIs are detected. Running the source checkout also requires Node.js 20+ and npm.

Repository Layout

AgentHive/
|-- context/                      # Local context files; real business data is ignored by default
|-- docs/assets/screenshots/       # README screenshots
|-- packages/agenthive-desktop/    # Windows desktop app and ZIP build scripts
|-- roundtable-skills/             # Agent Skill templates
|-- scripts/                       # Launch scripts
|-- workspace/frontend/            # Roundtable web frontend
|-- Start-Roundtable.cmd
`-- Start-Roundtable-UI-Only.cmd

Custom Agents

You can customize agents in two ways:

1. Edit the agent name, avatar, responsibilities, participant status, and background-research mode in the Agent Registry UI.
2. Add a new directory under roundtable-skills/, create your own SKILL.md, and register it in roundtable-skills/manifest.json.

A useful agent Skill usually includes:

• role definition
• discussion responsibilities
• decision principles
• evidence boundaries
• output style
• trigger conditions

Privacy And Boundaries

This repository does not include real business data, private meeting notes, API keys, credentials, or local context files. The default .gitignore excludes:

• real files under context/*
• .env / .env.*
• node_modules/
• .next/
• runtime output and logs

If you use private business materials, keep them in your local context/ directory and do not commit them to a public repository.

Disclaimer

The executive-style agents in this repository are AI discussion templates abstracted from public materials. They are intended for product prototyping and strategy-discussion research. They do not represent, impersonate, or imply endorsement by any real person or company.

MCP Icon Verification

The README body icons use Lucide line icons verified through better-icons search results and referenced through Iconify URLs, keeping one icon family, size system, and visual weight. Verified icons include:

lucide:brain-circuit
lucide:github
lucide:file-text
lucide:users-round
lucide:settings-2
lucide:book-open
lucide:database-search
lucide:search-check
lucide:download
lucide:workflow
lucide:rocket
lucide:square-terminal
lucide:box
lucide:shield-check
lucide:lock-keyhole
lucide:circle-check

iconfont-mcp was also verified with public Chinese search and download flows: search keywords were 圆桌 讨论 证据 and 搜索, and the downloaded verification file is docs/assets/icons/iconfont-verification/iconfont-search-verification.svg. The downloaded iconfont asset was not mixed into the README visual system to avoid inconsistent icon styles.

License

Apache License 2.0. See LICENSE.