Click here to see the hive in action
HiveAI is a multi-agent LLM hivemind you can run on your own hardware (or rented GPUs) while keeping everything fully private.
Instead of relying on a single model to answer your questions, HiveAI spins up a swarm of specialised agents called "bees" that think together, debate, critique, and refine ideas. At the centre sits the Queen, a coordinating agent that manages context and turns the swarm’s raw thinking into a clean, structured answer.
Powered by Parallax, HiveAI talks to fully private LLM endpoints (local or remote tunnels), so your data never leaves your control.
No SaaS APIs, no telemetry, just your own hive.
At a high level, HiveAI is a hivemind chat system:
- You create a Hive for a conversation or project.
- Inside that hive you define a set of Bees (specialised agents) and a Queen (coordinator).
- When you send a prompt, the hive operates in rounds: each bee responds, bees discuss and debate, and then the Queen synthesizes the swarm's thinking into a single answer.
A Hive is a collaborative reasoning space that bundles:
- Bees - individual AI agents with specific roles (e.g. Creative, Critic, Analyst).
- Queen - a special agent that manages context and synthesises outputs.
- Memory - persistent conversation history that the Queen can pull from.
- Cluster - a set of private LLM endpoints exposed by your Parallax setup.
Bees are the specialised workers of the hive. Each bee has:
- A Role - the lens it brings to the problem (e.g. "Realist", "Creative").
- A Model - a Parallax-backed endpoint (local or remote tunnel).
- Injections - probability-based behavioural tweaks that can fire on certain rounds.
The Queen is responsible for turning multi-agent chaos into something useful, while staying neutral:
- Context provider - fetches relevant memories from past discussions before each round and provides it to the bees.
- Aggregator - reads all bee outputs and produces one clear, structured, actionable reply.
- Neutral summariser - does not interpret the prompt herself or take a side; she only synthesises what the bees say back to you.
Instead of a single-turn Q&A, HiveAI thinks in rounds:
- Prompt - you send a request.
- Context - the Queen pulls relevant history.
- Collaboration - bees respond over N rounds, reading and building on each other.
- Aggregation - the Queen compresses the swarm into a single answer.
- Storage - the whole interaction is saved for future context.
On top of their normal roles, bees can optionally have injections attached to them during a discussion. These are small, temporary behavioural nudges that sometimes fire on a given round.
Instead of being part of the core loop, injections sit on top of it: they’re ephemeral, probability-based modifiers that occasionally push a bee to respond in a slightly different way (for example by challenging the previous point or adding a metaphor).
An injection is a temporary instruction, constraint, or "creative spark" added to a bee’s prompt for a single round.
- Ephemeral - automatically removed after the round.
- Non-destructive - never permanently changes a bee’s identity or memory.
- Additive at definition time - you can define multiple injections for a single bee.
Each injection has an interval value (X), representing a 1/X chance of activation per round:
interval = 1→ 100% (always hits on the roll)interval = 2→ 50% chance each roundinterval = 10→ 10% chance each round
For a given bee:
- Roll independently for each defined injection using its own
1/Xprobability. - Collect all injections whose rolls hit on this round.
- If no injections hit, the bee behaves normally.
- If exactly one injection hits, apply that injection.
- If multiple injections hit:
- Prefer the one with the lowest probability (i.e. the largest
intervalvalue). - If several share the same lowest probability, pick one of those at random.
- Prefer the one with the lowest probability (i.e. the largest
Only one injection is ever applied to a bee in a given round. This prevents contradictory instructions (e.g. “Be extra verbose” vs. “Don’t contribute to the discussion this round”) from being active at the same time.
When a round begins, the system performs a roll for every injection on every bee. Active injections are appended to a structured section of the system prompt, for example:
# Special directive (Must follow):
Temporarily think out loud this round.
This injection system adds controlled randomness and variation to how bees respond, instead of being a simple prompt tweak:
- Controlled stochasticity boosts exploration without losing coherence.
- Divergent behaviour stops the hive from collapsing into repetitive loops.
- Ensemble diversity forces different perspectives for the Queen to synthesise.
A single LLM is good, but it thinks in one voice. A hive lets you orchestrate many perspectives working together, which:
- Improves reasoning depth and structure.
- Reduces blind spots through critique and debate.
- Produces richer, more creative emergent outputs.
This unlocks patterns like:
- Collaborative workflows where agents play different roles, like a real team.
- Multi-step pipelines where each bee specialises in a stage of the workflow.
- Divergent + convergent thinking, with some bees exploring and others refining.
- Brainstorming swarms that generate, debate, vote, cluster, and evolve ideas.
Two concrete examples you can build with HiveAI:
- Bees (3):
- Dreamer - generates wild, unconventional ideas and creative directions.
- Pragmatist - filters ideas for feasibility and real-world constraints.
- Scout - rapidly scans for overlooked angles, constraints, or opportunities that the others missed and brings them back to the hive.
Based on the General Debate Hive [3vs3] config in this repo.
- Bees (6):
- Team A - Ethics - argues in favour from values, fairness, and long-term wellbeing.
- Team B - Ethics - argues against from ethical risks, harms, and unfairness.
- Team A - Pragmatist - focuses on practical implementation and feasibility for the pro side.
- Team B - Risk & Safety - surfaces worst-case scenarios and safety concerns for the con side.
- Team A - Evidence & Impact - brings data, evidence, and impact analysis for the pro side.
- Team B - Practical Skeptic - questions feasibility, costs, and unintended side-effects for the con side.
In both setups, the Queen remains neutral and only synthesises the bees’ reasoning back to you.
git clone https://github.com/FabsOP/HiveAI.git
cd HiveAI
pip install -r requirements.txt- Launch the UI - run
python app.py. - Create a Hive - use the interface to spin up a new hive.
- Add Model Endpoints - connect your Parallax endpoints (e.g.
http://localhost:3001). - Create Bees - add bees to your hive through the UI.
- Attach Models - assign models to your Queen and bees.
- Define Roles - give each bee a clear role to guide behaviour.
- Add Injections (Optional) - configure random injections to add variety.
- Start Collaborating - ask a question and watch the bees think together in real time.
hiveConfiguration.mp4
Some ideas for where HiveAI could go next:
- New bee types - new type of bees such as tool calling abilities
- Document ingestion - let hives read and reason over PDFs and text files.
- Richer memory - structured, long-term hive memory for projects.
This project is licensed under the MIT License - see the LICENSE file for details.
Join the revolution in private, collaborative AI.
Powered by Parallax