A hardware-aware, local-first desktop studio for language models.
Run models on your own machine, connect API providers when you choose, and keep one clear interface for models, agents, memory, search, and evaluation.
| Section | What you will find |
|---|---|
| Download | Published desktop bundles and release notes |
| Installation | Quick install and the recommended source start |
| Two ways to run | Local runtimes and optional API providers |
| Highlights | Engine, PhiloBots, memory, search, and marketplace |
| Engine demo | A visible hardware-aware fallback decision |
| Features | Complete feature and maturity overview |
| Architecture | Components, interfaces, and data flows |
| Privacy | Local storage and external network boundaries |
| Transparency | Provider choices and AI-assisted development |
| Current scope | Available modules, alpha features, and phases 1–5 |
| Documentation | Roadmap, troubleshooting, development, and contribution guides |
Important
PhiloEngine is in Phase 1 alpha. Chat, model management, local inference, API providers, memory, and the marketplace are usable today. Some surfaces are still evolving, and future modules remain visibly locked rather than pretending to be finished.
Note
This documentation follows the current development tree, which can be ahead of the latest packaged release. For an installed package, its release notes are the authoritative feature list.
|
PhiloEngine detects RAM, GPUs, and available VRAM before a model starts. It proposes a context size, selects a compatible runtime, and can step down to a safer configuration when the first plan does not fit. |
Use the same marketplace and chat workflow with hosted models. Provider keys are stored in the local application data and sent only to the provider involved in that request. |
| Runtimes: llama.cpp · vLLM · Transformers | Providers: OpenRouter · Featherless |
|
Model fingerprinting, memory estimates, runtime selection, context planning, GPU/CPU placement, guarded startup, and visible fallback decisions. |
Create assistants with their own prompt, style, trigger words, model binding, planning mode, project-aware file tools, permission prompts, and readable diffs. |
Recall across sessions with per-user and per-project storage, SQLite FTS5, vector retrieval, context budgeting, and optional embedding backends. The default is deterministic local hash embeddings; optional ONNX embeddings require a separately configured source-tree sidecar, which Quick Install does not bundle. |
|
Search through DuckDuckGo, Brave, Google, Bing, or Wikipedia. Public pages can be fetched, guarded against local-network targets, and condensed to Markdown. |
Browse local and hosted models together. Local candidates show formats, quantizations, estimated resource use, and a hardware-fit verdict when enough metadata is available. |
A requested 64k context does not fit, so the engine retries with 32k and keeps the decision visible.
Model Studio |
Marketplace |
flowchart LR
UI["Flutter desktop app<br/>Material 3 · DE/EN"]
API["Go backend<br/>Fiber · HTTP/JSON · SSE"]
PLAN["Hardware planner<br/>RAM · VRAM · runtime recipes"]
WORKERS["Local workers<br/>llama.cpp · vLLM · Transformers"]
DATA[("Local application data<br/>accounts · settings · chats · memory")]
CLOUD["Optional network services<br/>providers · search · news · datasets · updates"]
SKILLS["Skills client"]
UI --> API
API --> PLAN --> WORKERS
API <--> DATA
API -. "feature-specific, including documented background refreshes" .-> CLOUD
SKILLS -. "limited gRPC surface" .-> API
Both application servers bind to 127.0.0.1 by default. The main desktop flow
uses HTTP/JSON and server-sent events; the current gRPC surface is limited to
Skills. See Architecture for module and data-flow
details.
“Local-first” means that accounts, settings, chats, memory, and model files are kept on your machine. It does not mean that every feature is offline:
| Action | Where data goes |
|---|---|
| Chat with a local model | Inference stays with the local backend and worker; separately configured online tools or remote embeddings can still create their own network requests |
| Chat with an API model | The selected API provider |
| Search, news, or benchmarks | The selected/public source used by that feature |
| Browse or download models | Hugging Face or the configured provider |
| Check for application updates | GitHub release infrastructure |
News, benchmark refreshes, and update checks may contact their documented sources automatically. Read the full privacy and network matrix before using PhiloEngine in a restricted environment.
Warning
PhiloBot adds application-level path checks and approval prompts, but command execution is not an operating-system sandbox. Bind projects carefully and review proposed actions and diffs.
Download the Quick Install archive for your platform from the
releases page. Its
filename ends in -<release-target>-quickinstall followed by the archive
extension. Do not use the similarly named update archive for a first
installation. Extract the Quick Install archive once, and start
myphiloengine (myphiloengine.exe on Windows).
The launcher verifies the published archive size and SHA-256, installs updates
atomically, and can roll back a version that fails its initial health check.
| Published target | Quick Install filename ending |
|---|---|
| Linux | x64 · -linux-x64-quickinstall.tar.gz |
| Windows | x64 · -windows-x64-quickinstall.zip |
| macOS | Apple Silicon / ARM64 · -macos-arm64-quickinstall.tar.gz |
Quick-install users do not need Flutter or Go. See the installation guide for platform steps, update behavior, runtime prerequisites, and troubleshooting.
On Linux, after completing the one-time setup described in the installation guide, change to the repository directory and start PhiloEngine with the project launcher:
cd /path/to/philoengine
./start.shstart.sh opens the development console and manages the backend and frontend
in the intended order. On a clean main checkout it may apply a safe
fast-forward update first; it does not overwrite local changes.
Source development requires Go 1.25+, Flutter 3.44+ / Dart 3.12+, Python 3, and the native toolchain required by the selected local inference runtime.
| Status | Modules |
|---|---|
| Phase 1 — available | Chat, Engine, Marketplace, authentication, user preferences, PhiloBots, memory, text search, settings, skills administration |
| News — alpha | AI and technology feeds with search, filters, and saved articles |
| Benchmark — alpha | LMArena text leaderboard with ranking, model details, and comparison views |
| Phase 1 — active development | Improve existing functionality, fix bugs, refine frontend design and usability, strengthen verification, and expand the documentation |
| Phase 2 — planned | Extend and improve current features, with usable connections to external servers |
| Phase 3 — locked preview | Guided full fine-tuning, fine-tuning, and quantisation workflows |
| Phase 4 — locked preview | Image and video generation plus a game-development workspace |
| Phase 5 — long-term direction | Opt-in sharing of self-hosted AI models and compute capacity, following the project's commitment to keep the PhiloEngine software free and open source |
There are no promised dates for future phases. The detailed roadmap separates working functionality from previews and planned work. Always compare this development overview with the notes for the release you actually install.
| Guide | Purpose |
|---|---|
| Installation | Quick install, source setup, first run, runtimes, and updates |
| Features | Detailed feature and maturity matrix |
| Architecture | Components, modules, data flows, and security boundaries |
| Privacy | Local storage and every class of external connection |
| Project transparency | Why these runtimes/providers exist and how AI assists development |
| Troubleshooting | Startup, runtime, memory, provider, and update problems |
| Development | Repository structure, commands, tests, and conventions |
| Roadmap | Current phase, active work, and future modules |
| Contributing | CLA, DCO sign-off, pull requests, and verification |
Contributions are welcome, especially reproducible bug reports, fixes, hardware compatibility feedback, tests, and documentation improvements. Start with CONTRIBUTING.md; contributions require acceptance of the CLA and signed-off commits.
Report vulnerabilities privately to security@fillystudio.com as described in
the security policy.
PhiloEngine source code is licensed under the GNU AGPL-3.0. Project names and logos are not granted by the code licence; see the trademark policy. Models, runtimes, and bundled dependencies retain their own terms and licences.
Built by fillystudio · Powered by PhiloEngine