Hivekeep

Your AI team. At home.
The simplicity of a consumer assistant, the sovereignty of your own server.
A self-hosted platform of persistent AI agents that remember, collaborate, and answer you everywhere, in a single container.

_{100% open-source (MIT) · self-hosted · zero external infra}

Website · Docs · Install · Plugins

_{▶ Watch the 6-minute walkthrough — onboarding, Queenie setting everything up, agents building a mini-app and a custom tool live, and a full tour.}

The 30-second pitch

Most AI tools treat every conversation as disposable. You explain yourself on Monday, they have forgotten by Tuesday. Hivekeep takes the opposite path: a team of persistent Agents that live on your server, keep their memory, and work together. Not a chatbot. A household of specialists that works like a hive.

• They never forget you. One continuous session per Agent (no "new conversation"), backed by a hybrid long-term memory that accumulates months of context. No reset, ever.
• A team, not a chatbot. Agents collaborate (request / reply), delegate to ephemeral sub-Agents, and run scheduled work, so several things move at once.
• Self-improving. Your Agents build their own tools, mini-apps, and plugins. The platform grows with you instead of staying frozen.
• Everywhere. Telegram, WhatsApp, Slack, Discord, Signal, Matrix, plus a polished PWA. An Agent can hand a channel to a specialist in real time.
• One container. Zero Postgres, Redis, Mongo, or queue broker. One process, one SQLite file. Run it and Queenie sets up the rest by conversation.
• Your secrets stay yours. An AES-256-GCM vault that is never exposed to the LLM, connected accounts that never leave your infrastructure, and token transparency so you stay in control of costs.

A peek inside

One continuous conversation Every tool call rendered inline, never a black box.	Agents ask, never see Secrets go straight to the encrypted vault. The model never sees the value.
Your agents build apps Real mini-apps, hosted right inside Hivekeep.	A shared kanban Projects and tickets your agents work alongside you.
A real workspace Browse and edit every agent's files, with a proper editor.	Every token on record Cost per agent, per model, per day. No surprises.

Take the full tour →  ·  30+ screenshots from a running hive

They never forget you

Every Agent has a persistent identity (name, role, character, expertise, model, avatar) shared across everyone on your instance, and one continuous session that never resets. Underneath sits a dual-channel long-term memory: automatic extraction during compaction, plus explicit tools (recall, memorize, update_memory, forget). Recall is hybrid, sqlite-vec KNN fused with FTS5 keyword search via Reciprocal Rank Fusion, then re-ranked by importance, temporal decay, and category intent. Compaction summarizes old context to stay within token limits and never deletes the originals, so anything can be recovered later with search_history.

One container. Nothing leaves your server

A single process, a single SQLite file, a single Docker container. No Postgres, Redis, Mongo, SearxNG, or queue broker to provision. You bring one LLM key, and the whole platform runs on your infrastructure. Upgrading is re-pulling the image (or git pull for the native install). Your conversations, memories, and secrets stay on your server.

Your agents extend the platform themselves

This is the surprising part. Hivekeep is not a fixed feature set, it is a base your Agents grow:

• Custom tools, any language (Python, Node, Bun, TypeScript, Bash, Deno) with native dependency management, and rich React renderers so a tool result shows up as a themed UI card, not raw JSON.
• Mini Apps built by your Agents: real web apps in a sandboxed iframe, with a JS SDK, 29 React hooks, 50+ themed components (DataGrid, charts, Kanban, Calendar), an optional Hono backend, KV storage with snapshots and rollback, a public gallery with clone, 14 templates, and an "Improve this" natural-language edit loop.
• Plugins over npm with a typed TypeScript SDK (@hivekeep/sdk): a built-in marketplace (any package keyworded hivekeep-plugin, live npm search) plus Git install, native provider interfaces, channel adapters, lifecycle hooks, granular runtime-enforced permissions, and a scaffolder (create-hivekeep-plugin). Fully self-hosted, no proprietary cloud.
• Dynamic MCP servers the Agents can add and manage, and toolboxes (composable named allow-lists) to scope capabilities precisely per role.

One inbox for your whole team

Six native channels: Telegram, Discord, Slack, WhatsApp, Signal, Matrix, plus the PWA as a seventh surface. The standout is real-time channel handoff: transfer_channel reassigns a channel from one Agent to another mid-conversation. The address does not change, the Agent does. The new Agent receives handoff context and the causal chain, with async delivery statuses throughout.

No black box. No cost surprises

Transparency is a first-class feature, framed as "you control your costs," never "here is the bill." A Context Viewer breaks the prompt into a stacked, color-coded bar by section. Cache observability shows hit rate and TTL. Per-Agent EMA calibration compares estimated to actual token counts, and every call is tracked. The vault is AES-256-GCM, never printed in prompts or logs. Connected accounts (mail, calendar, contacts) use OAuth or encrypted credentials, and their tokens are never seen by Agents.

Setup is a conversation, not a YAML file

Onboarding is three quick screens (identity, language, one LLM key). Then Queenie, a permanent configurator Agent with 45+ tools, takes over by conversation: she connects providers, secures secrets in the vault, and creates your first Agents. No YAML, no CLI. Secure input means secrets go UI to vault and never to the LLM (only a non-sensitive confirmation comes back). One OpenAI key becomes several auto-detected capabilities. Queenie stays accessible for life.

Build a household for your life

Only Queenie ships built in. Everyone else is yours to create, each with an auto-generated avatar. A few examples of what people build (names from the sample roster):

The same building blocks cover a DevOps copilot, a home brain, a personal knowledge base, a multi-agent dispatcher team, a business monitor, or a shared family assistant.

Providers and plugins

Bring one config per provider and Hivekeep auto-detects its capabilities (llm, embedding, image, search, stt, tts). Built in today: Anthropic (API key and Claude Max OAuth), OpenAI (API key and Codex CLI), Google Gemini, OpenRouter, xAI, DeepSeek, MiniMax, Kimi (Moonshot), and a generic OpenAI-compatible connector (your own base URL, for NewAPI / LiteLLM / llama.cpp / LM Studio / vLLM / Ollama) for LLMs; OpenAI and Gemini for images; OpenAI and the OpenAI-compatible connector (local models via Ollama, llama.cpp, etc.) for embeddings; OpenAI and ElevenLabs for speech-to-text and text-to-speech; Brave Search, SerpAPI, Tavily, and Perplexity Sonar for web search. Need more? Add any provider as a plugin through the typed SDK, no fork required.

Why Hivekeep

Self-hosted AI assistants like OpenClaw and Hermes are excellent: they win or tie on memory, omnichannel reach, and self-hosting too. Where Hivekeep pulls ahead is the team, the polished product UI, and transparency. Marks below are best-effort from public docs (June 2026); peers genuinely win or tie on several rows.

Hivekeep is production-ready for individual and small-group use, with solid foundations and UX polish that keeps advancing. We are honest about the maturity (~80%) rather than overselling it. See the roadmap for the known rough edges.

Get started

One command. No docker-compose, no YAML, no database to provision.

Native install (recommended)

curl -fsSL https://raw.githubusercontent.com/MarlBurroW/hivekeep/main/install.sh | bash

The script installs Bun if needed, clones the repo, builds the frontend, runs migrations, creates a system service (systemd or launchd), and starts Hivekeep on port 3000. It preflights disk, RAM, ports, and connectivity, and rolls back cleanly on failure.

Then open http://localhost:3000 and Queenie takes it from there: three quick screens, then she configures everything by conversation.

Docker (alternative)

docker run -d \
  --name hivekeep \
  -p 3000:3000 \
  -v hivekeep-data:/app/data \
  ghcr.io/marlburrow/hivekeep:latest

This path requires the published container image. If the pull fails with manifest unknown or a 403, the image is not public yet, use the native installer above (it builds locally and needs no registry image), or build from source. See the install docs for the Docker Compose path and reverse-proxy setup. The installer also offers a hardened Docker mode: bash <(curl -fsSL .../install.sh) --docker.

Recovery one-liners

docker run -d -p 3001:3000 -v hivekeep-data:/app/data ghcr.io/marlburrow/hivekeep:latest

# Inspect why the daemon will not start:
docker logs hivekeep            # Docker
journalctl -u hivekeep -e       # native (systemd)

# Generate a diagnostic report for bug reports (no local clone needed):
bash <(curl -fsSL https://raw.githubusercontent.com/MarlBurroW/hivekeep/main/install.sh) --doctor
# Or, if you used the native installer, the wrapper it installs also works:
#   "$HIVEKEEP_DIR"/hivekeep doctor   (HIVEKEEP_DIR defaults to /opt/hivekeep or ~/hivekeep)

Back up your encryption key. On first boot Hivekeep generates and persists an AES-256-GCM key at data/.encryption-key and encrypts all vault secrets with it. If you instead set ENCRYPTION_KEY yourself, you must provide the same value on every restart. Keep this key (or the data/ directory) backed up, or you lose access to encrypted secrets.

To expose Hivekeep on your network or behind a reverse proxy, add -e HOST=0.0.0.0 and -e PUBLIC_URL=https://your-domain. See the configuration reference for every environment variable.

Features

All capabilities are detailed on the website and in the docs. The short version:

Persistent agents that remember and collaborate

• Persistent identity, expertise, and long-term memory per Agent, shared across instance users
• One continuous session, never resets
• Hybrid memory: sqlite-vec KNN + FTS5 fused by Reciprocal Rank Fusion, automatic extraction plus explicit tools
• Progressive compaction that summarizes without ever deleting originals
• Sub-Agents (await / async), inter-Agent messaging (request / reply), unified contacts registry, stateful browser

Self-hosted and self-improving platform

• One process, one SQLite file, one container, zero external infra
• Custom tools in any language with rich React renderers
• Mini Apps built by Agents (SDK, 29 hooks, 50+ themed components, optional backend, gallery)
• Plugins over npm via a typed SDK plus a built-in marketplace, dynamic MCP, composable toolboxes

A nice agent UI (PWA)

• Installable PWA, responsive and mobile-first
• 18 OKLch palettes (aurora, ocean, forest, sunset, monochrome, sakura, neon, lavender, midnight, copper, jade, crimson, galaxy, amber, slate, rose, mint, citrus) x light/dark/system x adaptive contrast
• Glass/gradient design system, WCAG AA, i18n (10 UI languages; Agents speak 40+)
• Rich tool renders inline, @mentions with autocomplete, real-time SSE streaming, auto-generated per-Agent avatars
• Files: a built-in workspace browser/editor — browse, edit, upload, and share any file your Agents produce, with clickable file paths in chat

Conversational onboarding (Queenie)

• Minimal 3-screen onboarding, then full guided configuration by chat
• Secure input: secrets go UI to vault, never to the LLM
• One key becomes many auto-detected capabilities, avatars are auto-generated and customizable

Everywhere

• Six native channels: Telegram, Discord, Slack, WhatsApp, Signal, Matrix
• Real-time channel handoff between Agents with handoff context and causal chain

Trust, security, and transparency

• AES-256-GCM vault, never exposed to the LLM
• Connected accounts (mail, calendar, contacts) whose tokens are never seen by Agents
• Token and context transparency: Context Viewer, cache observability, per-Agent EMA calibration
• Multi-user isolation, quick private sessions, Better Auth with admin and member roles

Architecture

One process, one SQLite file, one container. Everything is here, nothing external to wire up.

• Native LLM primitives. AI providers are implemented natively, organized by capability in src/server/llm/{llm,embedding,image,search,stt,tts,core}/. There is no external agent-orchestration SDK in the loop. Plugins consume the same surface through @hivekeep/sdk.
• Queue per Agent. A serialized FIFO per Agent with user-message priority, so there are no race conditions on shared context.
• Global SSE. One real-time connection per client, multiplexed by agentId. No per-Agent connections.
• No message deletion. Compaction compresses older messages into dated summaries that merge telescopically. Originals are always preserved and recoverable.
• Secrets stay in the vault. Vault secrets are never placed in prompts. Redaction prevents leaks into summaries.
• Pluggable providers. One config per provider, capabilities (llm, embedding, image, search, stt, tts) auto-detected.

Full details in the technical docs.

Tech stack

Configuration

All values have sensible defaults, you can start with an empty .env. The most common ones:

See config.md and .env.example for the complete reference (compaction thresholds, memory tuning, queue settings, cron limits, web browsing, and more).

Development

Prerequisites

• Bun 1.3+
• Git

Setup

git clone https://github.com/MarlBurroW/hivekeep.git
cd hivekeep
bun install
bun run dev    # Vite dev server (5173) + Hono backend (3000)

Commands

Project structure

src/
  server/    # Bun + Hono backend (routes, services, llm/, tools, db, auth, sse)
  client/    # React + Vite frontend (pages, components, hooks, locales, styles)
  shared/    # Types and constants shared by client and server
data/        # Created at runtime: SQLite DB, uploads, workspaces
site/        # Marketing site (Astro)
docs-site/   # Documentation (Astro + Starlight)

Design system

The UI follows a custom design system with 18 palettes and full dark/light support. Before building any UI:

• Consult src/client/pages/design-system/DesignSystemPage.tsx (the live showcase, source of truth)
• Use semantic CSS variables (var(--color-*)) or Tailwind tokens, never hardcode colors
• Reuse components from src/client/components/ui/

Community and support

• Issues - report bugs or request features
• Discussions - questions, ideas, show and tell
• Troubleshooting - TROUBLESHOOTING.md
• Security - see SECURITY.md to report a vulnerability

If Hivekeep is useful to you, consider starring the repo. It helps others find the project.