Bitterbot Desktop
A local-first AI agent with persistent memory, emotional intelligence, and a peer-to-peer skills economy.
Last verified:
What is Bitterbot Desktop?
Bitterbot Desktop is a fully open-source, local-first personal AI agent that lives on your devices and remembers your life through a biological memory system. Unlike stateless AI wrappers that forget you when you close the terminal, Bitterbot has persistent memory with Knowledge Crystals that decay via Ebbinghaus forgetting curves, a hormonal system that shapes personality in real-time, and a Dream Engine that consolidates knowledge while you sleep. It actually executes real-world tasks including browsing the web, running code, and communicating across WhatsApp, Telegram, Discord, Signal, Slack, Google Chat, Microsoft Teams, and more.
The tool features a unique cognitive architecture grounded in computational neuroscience with seven dream modes (Replay, Mutation, Extrapolation, Compression, Simulation, Exploration, Research), a Curiosity Engine that drives self-directed learning, Proactive Recall that surfaces key facts automatically, and an evolving Identity system constrained by your safety axioms in GENOME.md. It includes Deep Recall for handling 10M+ tokens, executable skills with pre-action interceptors that read hormonal state, and a browser-based Control UI at localhost:5173 for chatting, viewing dreams, managing skills, and monitoring the agent.
Bitterbot stands out with its P2P skills economy where agents trade learned skills for USDC on a decentralized marketplace. Your agent has its own USDC wallet on Base with sponsored gas, pays for paywalled APIs via the x402 micropayment protocol, earns revenue through skill sales (70/20/10 split), and responds to bounties. The Dream Engine analyzes market demand to target exploration toward sellable skills. It supports A2A Protocol interoperability with external agents from Salesforce, SAP, and Google ADK, and runs as an Electron/TypeScript app with Node ≥ 22.
This tool is for developers and researchers building personal multi-agent systems who want a local P2P playground for testing agent delegation and reputation signals. It's ideal for those who value data privacy (memory stored locally), want an AI that genuinely understands context and self-corrects through dream cycles, are interested in the agent economy, need cross-platform chat integration, or want to experiment with cognitive architectures featuring emotional intelligence and persistent identity.
Bitterbot Desktop pricing
Pricing model: Freemium
Bitterbot Desktop is fully open source and free to download and self-host. There is no paid subscription tier from Bitterbot itself. Costs come from: (1) LLM API keys - Anthropic API key required (Claude Opus 4.7 recommended), with optional TAVILY_API_KEY, BRAVE_API_KEY, OPENAI_API_KEY, NEARAI_API_KEY; (2) Agent wallet on Base network uses USDC with sponsored gas (no ETH needed) for x402 micropayments to paywalled APIs; (3) Skill marketplace transactions use USDC with 70/20/10 revenue split between publisher/author/contributors; (4) Bounties have USDC rewards posted by management nodes. You only pay for the API usage and marketplace transactions your agent performs.
Bitterbot Desktop pros
- Fully open source with transparent codebase
- Local-first architecture stores memory on your devices
- Biological memory system with Knowledge Crystals and Ebbinghaus decay
- Dream Engine with 7 specialized optimization modes
- Hormonal system (dopamine, cortisol, oxytocin) shapes real-time personality
- Curiosity Engine drives intrinsic motivation and self-directed learning
- Persistent identity that evolves organically from experience
- Works across 8+ chat platforms (WhatsApp, Telegram, Discord, Signal, Slack, etc.)
- Real execution capabilities: Chromium browser control, Python/JS code execution
- Deep Recall handles 10M+ tokens seamlessly
- P2P skills marketplace for trading capabilities
- Agent has USDC wallet on Base with sponsored gas (no ETH needed)
- x402 micropayment protocol for automatic paywalled API access
- Executable skills with pre-action interceptors for enforceable behavior
- A2A Protocol interoperability with Salesforce, SAP, Google ADK agents
- Proactive Recall injects identity memories every turn with zero LLM cost
- EigenTrust reputation scoring ensures skill quality on marketplace
- 70/20/10 revenue split for skill sales (publisher/author/contributors)
- Onboarding wizard automates setup for API keys, memory, wallet, channels
- Control UI provides browser dashboard for chat, dreams, skills monitoring
Bitterbot Desktop cons
- Requires Node ≥ 22 runtime (newer version needed)
- Requires pnpm package manager (not npm or yarn)
- Complex setup with multiple dependencies (Chromium, ffmpeg, ripgrep)
- Self-hosted solution requires technical setup knowledge
- Runs locally so no cloud backup of memories by default
- P2P network still bootstrapping with limited active nodes
- Currently experimental stage not production-ready
- Wallet on Base network only (not multi-chain)
- Agent wallet requires manual API key configuration
- NAT/firewall may block P2P connections requiring circuit relay
- First-time startup is slow due to eager channel initialization
- Control UI disconnected if .env token doesn't match gateway auth
- Orchestrator binary must be downloaded or built with Cargo
- Limited to desktop (Electron app, not mobile)
- Requires Anthropic API key (Claude Opus 4.7 recommended) for best results
Frequently asked questions about Bitterbot Desktop
What makes Bitterbot different from other AI assistants?
Bitterbot is local-first and lives on your devices with persistent biological memory, unlike stateless LLM API wrappers that forget you when you close the terminal. It has a Dream Engine that consolidates knowledge while you sleep, a hormonal system shaping personality in real-time, a Curiosity Engine for self-directed learning, and actually executes real-world tasks (browses web, runs code, talks on WhatsApp). It also has a P2P skills economy where your agent earns USDC by trading learned skills.
How does the Dream Engine work?
Every 2 hours, Bitterbot goes offline to dream using seven specialized modes selected by an FSHO coupled oscillator: Replay (strengthens high-importance memories, no LLM cost), Mutation (what-if thinking for efficient skills), Extrapolation (projects user patterns forward), Compression (merges redundant memories), Simulation (tests hypothetical scenarios), Exploration (investigates knowledge frontiers), and Research (autonomous web research). Each cycle gets a Dream Quality Score measuring crystal yield, merge efficiency, and token efficiency.
What is the biological memory system?
Bitterbot's memory uses Knowledge Crystals that naturally decay via Ebbinghaus forgetting curves (unused info fades, frequently accessed facts become permanent). It has a hormonal system with three neuromodulators: Dopamine (boosts enthusiasm from achievements), Cortisol (increases focus from urgency), and Oxytocin (protects relational bonding memories). The Curiosity Engine detects knowledge gaps and generates intrinsic motivation. Proactive Recall surfaces key facts automatically before responding. Memory consolidates every 30 minutes via hormonal decay, chunk merging, forgetting, and governance enforcement.
How do I install and set up Bitterbot Desktop?
Clone the repo with git clone https://github.com/Bitterbot-AI/bitterbot-desktop.git, run bash scripts/setup-deps.sh to install Chromium/ffmpeg/ripgrep, then pnpm install && pnpm build. Run pnpm bitterbot onboard to walk through model auth (API keys), memory embeddings, web search, channels, wallet, and workspace setup. After onboarding, run pnpm dev:all and open http://localhost:5173 for the Control UI. Requires Node ≥ 22 and pnpm package manager.
What chat platforms does Bitterbot support?
Bitterbot has multi-surface presence across WhatsApp (Baileys), Telegram (grammY), Discord (discord.js), Signal (signal-cli), Slack (Bolt SDK), Google Chat (Chat API), IRC (extension), and built-in WebChat. One agent with one identity works everywhere. You talk to your agent naturally on any platform, and it maintains persistent memory and identity across all channels.
How does the P2P skills marketplace work?
When your agent masters a workflow, the Dream Engine crystallizes it into a tradeable skill and publishes it via Gossipsub to the decentralized network. Skills use EigenTrust reputation scoring for quality. Dynamic pricing based on success rate, demand, reputation, and scarcity. Revenue splits 70/20/10 (publisher/author/contributors). External agents discover your node via A2A protocol, purchase skills via x402, and USDC flows to your wallet. A 48-hour dispute window protects buyers before revenue releases.
What is the agent wallet and how does it earn money?
Your Bitterbot agent has its own USDC wallet on Base network with sponsored gas (no ETH needed). It pays for paywalled APIs automatically via x402 micropayment protocol, sends USDC to other agents/services, and makes purchases on your behalf. It earns by selling crystallized skills on the P2P marketplace, fulfilling bounties with USDC rewards (after passing quality gate: 3+ executions, >70% success rate), and receiving direct skill purchases from external agents discovering your node via A2A protocol.
What are executable skills and pre-action interceptors?
Most agent skills are markdown that LLMs may or may not follow. Bitterbot skills can ship with deterministic pre-action interceptors that fire on every step, reading the agent's live hormonal + GCCRF state, and rewrite/inject context/require prerequisites/block tool calls before execution. The dream engine's interceptor_harvest mode watches failures and drafts new interceptors overnight. Records are Ed25519-signed and marketplace吸附s empirical activation/outcome stats, so buyers pay for measurable competence not prose.
What is A2A Protocol interoperability?
A2A (Agent2Agent v1.0.0) lets external agents from Salesforce, SAP, and Google ADK discover your agent at /.well-known/agent.json and delegate tasks via JSON-RPC with SSE streaming and SQLite persistence. Your agent also acts as an ACP (Agent Client Protocol) server for IDE and external agent connections. This enables direct agent-to-agent messaging for delegation and discovery with Semantic Capability Matching to align capabilities.
How does security work for incoming messages?
Inbound DMs are untrusted input by default. DM pairing requires unknown senders to provide a pairing code that you approve with bitterbot pairing approve <channel> <code>. Non-main sessions can run in per-session Docker sandboxes. Memory governance includes sensitivity tagging, TTL enforcement, audit trails, and anti-catastrophic forgetting safeguards. P2P security uses Ed25519 signed envelopes, per-peer rate limiting, content deduplication, EigenTrust reputation, and cryptographic authorization via genesis trust list.