Dream Server installs and wires together everything you need to run AI locally, so you do not have to assemble Ollama, Open WebUI, n8n, ComfyUI, and privacy tools by hand:

• Local model inference — run open models on your own hardware
• ChatGPT-style web UI — talk to your models from any browser
• Control dashboard — manage models, services, setup, GPU status, and extensions from one place
• Voice, agents, and workflows — build automations that can listen, speak, call tools, and get work done
• RAG and search — connect local documents, private search, and retrieval workflows
• Image generation — run local image tools without sending prompts to a hosted API
• Privacy and ops — keep service auth, secrets, observability, and diagnostics in one local stack

No cloud required. No subscriptions required. Your prompts and data stay on your machine unless you choose otherwise. Cloud and hybrid API modes are optional when you want them.

Linux and macOS:

curl -fsSL https://raw.githubusercontent.com/Light-Heart-Labs/DreamServer/main/dream-server/get-dream-server.sh | bash

Windows users should use the PowerShell installer shown below or follow the Windows Quickstart.

After install, open http://localhost:3000 and start chatting.

API endpoint: Linux Docker installs expose llama-server on http://localhost:11434 by default (OLLAMA_PORT) while containers use llama-server:8080. macOS native Metal and Windows native/Lemonade paths use http://localhost:8080 unless overridden. Open WebUI stays on http://localhost:3000.

No GPU? Dream Server also runs in cloud mode — same full stack, powered by OpenAI/Anthropic/Together APIs instead of local inference:

./install.sh --cloud

Port conflicts? Every port is configurable via environment variables. See .env.example for the full list, or override at install time:

WEBUI_PORT=9090 ./install.sh

New here? Read the Friendly Guide or listen to the audio version — a complete walkthrough of what Dream Server is, how it works, and how to make it your own. No technical background needed.

Current Platform Support

Platform	Status
Linux (NVIDIA + AMD + Intel Arc)	Supported — install and run today
Windows (NVIDIA + AMD)	Supported — install and run today
macOS (Apple Silicon)	Supported — install and run today

Tested Linux distros: Ubuntu 24.04/22.04, Debian 12, Linux Mint 21.3, Fedora 41+, Rocky Linux 9, Arch Linux, Manjaro, CachyOS, and openSUSE Tumbleweed. Other distros using apt, dnf, pacman, or zypper should also work — open an issue if yours doesn't.

Testing surface: CI and tower2 Docker containers cover broad distro installer logic; tower2 Incus VMs cover systemd + Docker daemon behavior on Ubuntu, Fedora/Rocky, Arch, and openSUSE; the real hardware fleet remains the release gate for NVIDIA/AMD/Apple GPU runtime, dashboard, Hermes, UI, and capability validation.

Windows: Requires Docker Desktop with WSL2 backend. NVIDIA GPUs use Docker GPU passthrough; AMD Strix Halo runs through the platform-specific accelerated path documented in the Windows installer and support matrix.

macOS: Requires Apple Silicon (M1+) and Docker Desktop. llama-server runs natively with Metal GPU acceleration; all other services run in Docker.

See the Support Matrix for supported platform claims and the Validation Matrix for the layered test surface used to test those claims.

A handful of companies control the vast majority of global AI traffic — and with it, your data, your costs, and your uptime. Every query you send to a centralized provider is business intelligence you don’t own, running on infrastructure you don’t control, priced on terms you can’t negotiate.

If AI is becoming critical infrastructure, it shouldn’t be rented. Self-hosting local AI should be a sovereign human right, not a career choice.

Because running your own AI shouldn't require a CS degree and a weekend of debugging CUDA drivers. Right now, setting up local AI means stitching together a dozen projects, writing Docker configs from scratch, and praying everything talks to each other. Most people give up and go back to paying OpenAI.

We built Dream Server so you don't have to.

• One command — detects your GPU, picks the right model, generates credentials, launches everything
• Chatting in under 2 minutes — bootstrap mode gives you a working model instantly while your full model downloads in the background
• Full service stack, pre-wired — chat, agents, voice, workflows, search, RAG, image generation, privacy tools, observability, and developer tools. All talking to each other out of the box
• Fully moddable — every service is an extension. Drop in a folder, run dream enable, done

git clone https://github.com/Light-Heart-Labs/DreamServer.git
cd DreamServer/dream-server
./install.sh

Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass
git clone https://github.com/Light-Heart-Labs/DreamServer.git
cd DreamServer
.\install.ps1

git clone https://github.com/Light-Heart-Labs/DreamServer.git
cd DreamServer/dream-server
./install.sh

• Open WebUI — full-featured chat interface with conversation history, web search, document upload, and 30+ languages
• llama-server — high-performance LLM inference with continuous batching, auto-selected for your GPU; Linux Docker host API defaults to localhost:11434, native macOS/Windows paths use localhost:8080, and container API runs on 8080
• LiteLLM — API gateway supporting local/cloud/hybrid modes
• TEI Embeddings — text embedding service for RAG and search workflows

• Whisper — speech-to-text
• Kokoro — text-to-speech

• Hermes Agent — default local-first autonomous/browser agent with memory, skills, and a magic-link-gated proxy
• OpenClaw — deprecated legacy autonomous agent, still opt-in during the migration window
• n8n — workflow automation with 400+ integrations (Slack, email, databases, APIs)
• APE — Agent Policy Engine for auditing and governing autonomous tool calls
• OpenCode — browser-based AI coding assistant wired to the local stack
• Memory Shepherd — host/systemd helper for agent memory lifecycle management

• Qdrant — vector database for retrieval-augmented generation (RAG)
• SearXNG — self-hosted web search (no tracking)
• Perplexica — deep research engine
• Brave Search — optional paid Brave Search API integration

• ComfyUI — node-based image generation

• Privacy Shield — PII scrubbing proxy for API calls
• Dashboard — real-time GPU metrics, service health, model management
• Dashboard API — service health, setup, status, metrics, and management API behind the dashboard
• Token Spy — token usage monitor for local and proxied LLM traffic
• Langfuse — optional LLM observability and tracing

The installer detects your GPU and first assigns a deterministic hardware tier. Linux and macOS then run the versioned catalog selector (dream-server/scripts/select-model.py), while Windows uses the PowerShell catalog selector in dream-server/installers/windows/lib/tier-map.ps1; both read dream-server/config/model-library.json to choose the best installable GGUF for the detected memory envelope. The final choice is written to .env as LLM_MODEL, GGUF_FILE, MAX_CONTEXT, and MODEL_RECOMMENDATION_*.

MODEL_PROFILE=qwen is the default non-Gemma catalog profile, so the effective pick can be Qwen, Phi, or DeepSeek depending on what fits best. MODEL_PROFILE=gemma4 forces Gemma 4 where available, and MODEL_PROFILE=auto uses Gemma 4 on NVIDIA, Apple Silicon, and Intel Arc tiers. Override tier selection with ./install.sh --tier 3; override the model family with MODEL_PROFILE=gemma4 ./install.sh or MODEL_PROFILE=auto ./install.sh.

When Hermes is enabled, which is the default agent path, installers keep the first-run bootstrap model at a 64K context floor and promote the full local model context to 128K where the selected model supports it. That avoids Hermes's hard 64K minimum while preserving the under-2-minute first chat experience. The examples below are current catalog-selector outputs for common hardware envelopes; exact installs can differ with detected VRAM/RAM, host architecture, existing downloads, or explicit profile overrides. Throughput still needs a local benchmark after first launch.

The selector routes unified-memory hosts away from Qwen3 Coder Next when that model would otherwise be selected, because current repo policy documents correctness issues on those backends.

Gemma 4 profile tiers remain in the installer tier maps: E2B on entry hardware, E4B on midrange hardware, 26B-A4B on pro hardware, and 31B on large/ultra hardware.

No waiting for large downloads. Dream Server uses bootstrap mode by default:

1. Downloads a tiny 1.5B model in under a minute
2. You start chatting immediately
3. The full model downloads in the background
4. Hot-swap to the full model when it's ready — zero downtime

The bootstrap model starts with a 64K context window so Hermes can work during the first session. After the background download finishes, Dream Server swaps to the full model and restores the Hermes/full-model context target.

Skip bootstrap: ./install.sh --no-bootstrap

The installer picks a model for your hardware, but you can switch anytime:

dream model current              # What's running now?
dream model list                 # Show all available tiers
dream model swap T3              # Switch to a different tier

If the new model isn't downloaded yet, pre-fetch it first:

./scripts/pre-download.sh --tier 3    # Download before switching
dream model swap T3                    # Then swap (restarts llama-server)

Already have a GGUF you want to use? Drop it in data/models/, update GGUF_FILE and LLM_MODEL in .env, and restart with the CLI:

dream restart llm

Or restart the container directly from the installed dream-server directory:

docker compose restart llama-server

Rollback is automatic — if a new model fails to load, Dream Server reverts to your previous model.

Dream Server is designed to be modded. Every service is an extension — a folder with a manifest.yaml and a compose.yaml. The dashboard, CLI, health checks, and compose stack all discover extensions automatically.

extensions/services/
  my-service/
    manifest.yaml      # Metadata: name, port, health endpoint, GPU backends
    compose.yaml       # Docker Compose fragment (auto-merged into the stack)

dream enable my-service     # Enable it
dream disable my-service    # Disable it
dream list                  # See everything

The installer itself is modular — 19 library modules, a shared service registry, and 13 ordered phases. Want to add a hardware tier, swap a default model, or skip a phase? Start with the installer architecture map so you update the Linux, macOS, Windows, upgrade, and host-agent writers together.

Full extension guide | Installer architecture

The dream CLI manages your entire stack:

dream status                # Health checks + GPU status
dream list                  # All services and their state
dream logs llm              # Tail logs (aliases: llm, stt, tts)
dream restart [service]     # Restart one or all services
dream start / stop          # Start or stop the stack

dream mode cloud            # Switch to cloud APIs via LiteLLM
dream mode local            # Switch back to local inference
dream mode hybrid           # Local primary, cloud fallback

dream model swap T3         # Switch to a different hardware tier
dream enable n8n            # Enable an extension
dream disable whisper       # Disable one

dream config show           # View .env (secrets masked)
dream preset save gaming    # Snapshot current config
dream preset load gaming    # Restore it

Other tools get you part of the way. Dream Server gets you the whole way.

Dream Server exists because people chose to build instead of wait. Every contributor here is part of something bigger than code — a growing resistance against the idea that AI should be rented, gated, and controlled by the few. These are the founders of the sovereign AI movement, proving that one person, one machine, and one dream is enough.

Dream Server has been recognized by the local AI and developer community, including AMD Featured Developer recognition, selection as a May 2026 AMD Lemonade Developer Challenge winner, and a feature at (Co)nnect: Philly's AI Ecosystem Summit at Pennovation Works.

Thanks to lhl for strix-halo-testing — the foundational Strix Halo AI research and rocWMMA performance work that the broader community builds on.

Tony363 (Tony Siu), a former Google University Research lead and founder of Code & Coffee Philadelphia, has been one of the biggest reasons Dream Server reached broader awareness in the Philly AI ecosystem. Code & Coffee is now a 4,000+ developer community, and Tony's support helped bring Dream Server into that world, including multiple Pennovation Works features. More than any title, he cares deeply about local AI, empowerment for the masses, and seeing this project succeed. We are also eternally grateful to Ahmad Osman for featuring and publicly supporting Dream Server; after his endorsement, the project grew from roughly 500 to nearly 1,500 GitHub stars in four days, bringing a wave of visibility, adoption, and encouragement we will not forget.

• llama.cpp (ggerganov) — LLM inference engine
• Qwen (Alibaba Cloud) — Default language models
• Open WebUI — Chat interface
• ComfyUI — Image generation engine
• FLUX.1 (Black Forest Labs) — Image generation model
• AMD ROCm — GPU compute platform
• Strix Halo Testing (lhl) — Foundational Strix Halo AI research and rocWMMA optimizations
• n8n — Workflow automation
• Qdrant — Vector database
• SearXNG — Privacy-respecting search
• Perplexica — AI-powered search
• LiteLLM — LLM API gateway
• Kokoro FastAPI (remsky) — Text-to-speech
• Speaches — Speech-to-text
• Strix Halo Home Lab — Community knowledge base

• Yasin Bursali (yasinBursali) — DreamServer's most prolific contributor across 370+ merged PRs spanning nearly every layer of the stack. Built the extensions lifecycle system, compose security scanner, host-agent install/runtime flow, dashboard extension management, and broad extension-library hardening. Carried major portability, reliability, and safety work across macOS, Windows/WSL2, Apple Silicon, AMD/Lemonade, dream-cli, installers, dashboard-api, Privacy Shield, observability, model activation, and release-contract tests. His work turned DreamServer from a service bundle into a safer, extensible, heavily tested local AI platform.

• Youness Yachouti (y-coffee-dev) — Designed and built the full-stack multi-GPU system: NVIDIA topology detection, topology-aware service placement, docker-compose.multigpu.yml generation, dashboard GPU Monitor, and five dream gpu CLI subcommands with completions. Also fixed installer custom-mode opt-outs, added curl retries, longer llama-server startup grace, and phase-state markers that make long installs easier to resume and diagnose. Tested on real multi-GPU hardware including 4x RTX 4060 Ti, 4x RTX 4080, and 8x RTX 5060 Ti configurations.

• SSignall (Android Dev) — Added early ARM64 / NVIDIA Grace Blackwell support for DGX Spark-class systems: llama.cpp architecture mismatch detection, GB10/GB200 unified-memory fallback, docker-network service URL injection for diagnostics, setup wizard resilience when diagnostics fail, and QR helper fixes so terminal onboarding points at the actual dashboard URL.

• Tony363 (Tony Siu) — Raised dashboard-api test coverage to 95% with 3,500+ lines of tests across 14 files including comprehensive endpoint coverage for setup, privacy, workflows, updates, agents, and GPU monitoring, plus 7 BATS shell test suites covering logging, constants, path-utils, bootstrap-model, nvidia-topo, ui, and background-tasks. Added comprehensive architecture overview documentation (ARCHITECTURE.md) with Mermaid diagrams for service topology, installer pipeline, and compose layering. Fixed the pre-existing ThemeProvider CI failure that was blocking every PR frontend check. Reported the PyYAML import crash on Manjaro/Arch (resolve-compose-stack.sh) with clear root cause analysis. Drives developer outreach and ecosystem growth as head of Coffee and Code Philadelphia. Earlier work: hardened service-registry.sh against shell injection, improved PII scrubber with Luhn check, fixed token-spy settings persistence with atomic writes, fixed SSH command injection in session-manager.sh, narrowed broad exception catches across dashboard-api, and authored CLAUDE.md with project instructions and design philosophy. Built three AI-powered GitHub Actions workflows: consolidated code review with fork detection and protected file enforcement, label-gated issue-to-PR automation with 4-job pipeline (validate/implement/guardrails/create-pr) and prompt injection hardening (anti-injection preamble, 4000-char truncation, tool restrictions, secret scanning), and nightly AI scanners for code review/docs/autonomous scanning with budget caps and manual-only triggers. Fixed unified APU name fallback in GPU detection for Strix Halo. Prototyped a full Rust/Axum rewrite of the dashboard-api with 285 tests, constant-time auth middleware, 3-crate workspace, and ~25MB Docker image (work-in-progress — extension security features being ported). Fixed pipefail-safe hostname fallback in installer phase 13 for Arch/Manjaro compatibility

• latentcollapse (Matt C) — Security audit and hardening: OpenClaw localhost binding fix, multi-GPU VRAM detection, AMD dashboard hardening, and the Agent Policy Engine (APE) extension

• Nolan Makatche (nolanmak) — Hardened Dream Server's agent and privacy surfaces across three security-focused PRs: added persistent APE windowed governance with human-approval escalation, restart-safe state, one-shot approval grants, circuit breaker protection, warmup handling, and a 44-test contract suite (#1272); made OpenClaw device auth secure by default with explicit opt-in for dangerous disablement, removed the always-LAN unauthenticated gateway path, protected LAN-bound token exposure, and added device-auth regression coverage (#1273); and rebuilt Privacy Shield response handling around streaming-safe PII restoration, binary/gzip passthrough, oversized-response fallback, authenticated WebSocket passthrough, and 50 regression tests covering SSE, token boundaries, non-UTF-8 bodies, and auth gates (#1275).

• Igor Lins e Silva (igorls) — Stability audit fixing 9 infrastructure bugs: dynamic compose discovery in backup/restore/update scripts, Token Spy persistent storage and connection pool hardening, dotglob rollback fix, systemd auto-resume service correction, removed auth gate from preflight ports endpoint for setup wizard compatibility, added ESLint flat config for the dashboard, cleaned up unused imports and linting across the Python codebase, and resolved CI failures across dashboard and smoke tests

• Nino Skopac (NinoSkopac) — Token Spy dashboard improvements: shared metric normalization with parity tests, budget and active session tracking, configurable secure CORS replacing wildcard origins, and DB backend compatibility shim for sidecar migration

• Glexy (fullstackdev0110) — Hardened the installer and operations layer: safe .env loading across scripts, removal of unsafe eval, dream-cli status/config/mode commands, extension manifest compatibility validation, CPU-only support, Linux portability docs, structured Linux preflight diagnostics, extension runtime checks in install summary and CI, Windows compose failure reports, broader PowerShell lint coverage, and secret-redaction warnings in diagnostic output.

• bugman-007 — Parallelized health checks in dream status for 5–10× speedup using async gather with proper timeout handling, benchmark and test scripts, integrated backup/restore commands into dream-cli, added preset import/export with path traversal protection and archive validation, added preset diff command for comparing configurations with secret masking, quarantined broken edge quickstart instructions replacing them with supported cloud mode path, added SHA256 integrity manifests and verification for backups, added restore safety prompts requiring backup ID confirmation, added backup/restore round-trip integration test, added preset compatibility validation before load, added service registry tests to CI, added Python type checking with mypy, added disk space preflight checks to backup/restore with portable size estimation, and added session-level caching to compose flags resolution for performance, expanded dashboard-api test coverage for privacy, updates, and workflow endpoints, added structured logging to agent monitor replacing silent exception swallowing, added bash completion for dream-cli with dynamic backup ID resolution, added automatic pre-update backup with rollback command and health verification, fixed gitleaks CI to use OSS CLI instead of paid license action, added disk space preflight checks to backup/restore, and replaced disabled VAD patch with AST-based Python patcher for safe Whisper voice activity detection

• norfrt6-lab — Replaced 12+ silent exception-swallowing patterns with specific exception types and proper logging, added cross-platform system metrics (macOS/Windows) for uptime, CPU, and RAM, plus Apple Silicon GPU detection via sysctl/vm_stat

• boffin-dmytro — Hardened downloads, installers, and diagnostics: SHA256 GGUF verification, corrupt-file re-downloads, retrying network/Docker pulls, timeout hardening, port/Ollama preflight checks, compose validation, cross-platform path utilities, healthcheck resilience, manifest validation, doc-link CI, restore checksum verification, redacted support bundles, dashboard-api settings refactors, and setup/test coverage.

• takutakutakkun0420-hue — Added log rotation to all base services preventing unbounded disk growth, and added open-webui startup dependency on llama-server health ensuring the UI never shows a broken state

• reo0603 — Fixed Makefile paths after dashboard-api move and heredoc quoting bug in session-manager.sh SSH command, narrowed broad exception catches to specific types across dashboard-api, parallelized health checks for 17× faster execution, added compose.local.yaml for dashboard/open-webui/privacy-shield service dependencies, added .dockerignore files to all custom Dockerfiles reducing build context, fixed H2C smuggling vector in nginx proxy and added wss:// for HTTPS in voice agent, added comprehensive extension integration and hardware compatibility test suites, and hardened secret management with .gitignore patterns for key/pem/credential files and SQL identifier validation in token-spy

• Arifuzzaman Joy (Arifuzzamanjoy) — Pinned yq and docker-compose versions in the bootstrap Dockerfile replacing floating /latest/ tags with reproducible ARG-based version pins, added Draft7Validator compatibility for jsonschema 3.x on Ubuntu 22.04/24.04, added compatibility blocks (dream_min version bounds) to 25 extension library manifests, added missing gpu_backends to 8 extension manifests, added cpu and none to the gpu_backends schema enum enabling CPU-only service declarations, fixed gpu_backends on 13+ extension manifests resolving schema validation failures, added missing required fields (icon, category, requirements, priority) to localai features, fixed env_vars schema compliance (name to key) in bark and rvc manifests, corrected privacy-shield service ID to match schema pattern, and fixed typo in baserow manifest tags

• nt1412 — Wired dashboard-api agent metrics to Token Spy with background metrics collection, added TOKEN_SPY_URL/TOKEN_SPY_API_KEY env vars, fixed missing key_management.py in privacy-shield Dockerfile, and added ui_path to dashboard sidebar links so extension services open at their correct UI page

• evereq — Relocated docs/images for cleaner monorepo root

• championVisionAI — Added Alpine Linux (apk) and Void Linux (xbps) package manager support to the installer abstraction layer, hardened hardware detection with JSON output escaping and container/WSL2 detection, rewrote healthcheck.py with retries, HEAD-to-GET fallback, status code matching, and structured JSON output, hardened Docker phase with daemon start/retry logic and compose v1/v2 detection, added cross-platform python3/python command resolution with shared detection utility, and hardened env schema validation with robust .env parsing, enum validation, and line-number error reporting, added sim summary validation test suite with 10 test cases covering help, missing files, invalid JSON, and strict mode, hardened hardware detection with JSON output escaping and container/WSL2 detection, hardened healthcheck.py with retries and HEAD-to-GET fallback, hardened Docker phase with daemon start/retry and compose v1/v2 detection, fixed Windows python3/python command resolution, added extension audit workflow with 838-line Python auditor and 'dream audit' CLI command, added duplicate key detection to env validation, added compact JSON output mode and --help flag to hardware detection, and failed env validation on duplicate keys preventing silent config corruption

• buddy0323 — Ported Windows installer phases 01-07 to native PowerShell decomposing the monolithic script into focused phase files, added Intel Arc SYCL tier map (ARC/ARC_LITE) with docker-compose.intel.yml overlay, detection logic, tier-map tests, and SHA256 verification, added Intel Arc oneAPI SYCL compose overlay with two-stage llama-sycl Dockerfile, added Intel Arc detection checks (lspci, Level Zero runtime, render nodes, group membership), and authored the Intel Arc support matrix documentation and setup guide

• blackeagle273 — Enhanced macOS installer with idempotent .env and config generation preserving existing secrets across re-installs

• eva57gr — Fixed bash syntax error in Token Spy session-manager.sh SSH heredoc command, and unified port contract across installer, schema, compose, and manifests with canonical ports.json registry

• cycloarcane — Fixed unbound variable crash by guarding service-registry.sh sourcing in install-core.sh, health-check.sh, and 04-requirements.sh

• Rowan (rowanbelanger713) — Enhanced llama-server with configurable batch-size, threads, and parallel request knobs, added TTL caching and async threading to dashboard-api status endpoints, pooled httpx connections for LiteLLM, lazy-loaded React routes with memoized components, scoped CSS transitions to interactive elements, paused polling on hidden tabs, and split Vite output into vendor/icons chunks for faster loading

• gabsprogrammer — Designed the dashboard's "liquid metal" refresh and kept expanding the operator experience: secure Settings env editor with apply-changes flow, service telemetry and restart controls, integrations topology view, service readiness summaries, compose failure reports, llama-server image tag validation, long-context/MoE tuning knobs, Android Termux and iOS a-Shell paths, Caddy image maintenance, README accuracy fixes, and Windows/installer stability patches.

• Octopus (octo-patch) — Added MiniMax provider support for cloud and hybrid modes through LiteLLM, including minimax / minimax-fast model aliases, MINIMAX_API_KEY env/schema plumbing, Linux/macOS/Windows installer preservation, compose propagation, and user docs.

• onyxhat — Fixed missing variable initialization in installer scripts If we missed anyone, open an issue. We want to get this right.

Apache 2.0 — Use it, modify it, ship it. See LICENSE.