Fosowl/agenticSeek

AgenticSeek – A Fully‑Local, Voice‑Enabled AI Assistant

What it is

• An open‑source, privacy‑first AI agent that can browse the web, run code, and orchestrate multi‑step tasks entirely on your own hardware. It is positioned as a local alternative to commercial services like Manus AI.

Key capabilities

• Autonomous web browsing – search, read pages, extract information, and fill web forms without user intervention.
• Coding assistant – generate, debug, and execute code in several languages (Python, C, Go, Java, …) directly on your machine.
• Dynamic agent selection – the system decides which specialised “agent” (e.g., web‑searcher, coder, planner) is best for a given request.
• Task planning & execution – break down complex goals (trip planning, project scaffolding) into steps and carry them out using multiple agents.
• Voice interaction – experimental speech‑to‑text and text‑to‑speech so you can talk to the assistant.
• Local‑only mode – runs with locally hosted LLMs (Ollama, LM‑Studio, etc.) and stores all files, conversations, and search results on your device. Cloud APIs are optional.

How it works

1. Backend – a Python service that talks to a language model (local or remote) and coordinates agents.
2. Frontend – a React web UI (served on http://localhost:3000) that communicates with the backend via an API token.
3. Supporting services – Docker‑compose brings up:
   • SearXNG – an open‑source meta‑search engine used for web queries.
   • Redis – a lightweight store required by SearXNG.
   • Frontend & backend containers (optional; you can also run the backend directly on the host in CLI mode).
4. LLM providers – you can plug in any compatible model:
   • Local: Ollama, LM‑Studio, or any OpenAI‑compatible server (e.g., llama.cpp).
   • Remote APIs: OpenAI, Google Gemini, Deepseek, TogetherAI, HuggingFace, OpenRouter, MiniMax, etc.
5. Configuration – environment variables in .env define service URLs/ports, while config.ini selects the LLM provider, model, and behavioural flags (speech, personality, language list, etc.).

Installation & quick start

# 1. Clone and prepare
git clone https://github.com/Fosowl/agenticSeek.git && cd agenticSeek
mv .env.example .env

# 2. Edit .env – set ports, WORK_DIR, and any API keys you need
#    (leave LLM keys empty if you run a local model)

# 3. Start Docker services (full includes backend)
./start_services.sh full   # macOS/Linux
# or on Windows
start start_services.cmd full

# 4. Open the UI
open http://localhost:3000   # or visit in a browser manually

• The first UI load will ask for an API token; retrieve it from the backend logs (docker compose logs backend | grep "API token").
• For CLI‑only use, run ./install.sh (or install.bat), adjust SEARXNG_BASE_URL to http://localhost:<port>, then start services and launch uv run cli.py.

Running a local LLM

• Install Ollama (or LM‑Studio), start the server (ollama serve), and expose it via OLLAMA_HOST=0.0.0.0:11434.
• Edit config.ini:

  [MAIN]
  is_local = True
  provider_name = ollama
  provider_model = deepseek-r1:14b   # or any model your hardware can handle
  provider_server_address = 127.0.0.1:11434
  

• Minimum hardware: a GPU capable of running ~14 B parameter models (e.g., Magistral, Qwen, Deepseek 14B).

Using a cloud API

• Populate the appropriate *_API_KEY env vars, set is_local = False in config.ini, choose provider_name (e.g., openai) and a model name (e.g., gpt-3.5-turbo).

Typical workflow

1. Type or speak a request, e.g.:
   • “Search the web for top cafés in Rennes, France and save three with addresses to rennes_cafes.txt.”
2. AgenticSeek selects the web‑search agent, performs the query via SearXNG, parses results, and writes the file into your WORK_DIR.
3. For coding tasks, the coder agent generates source files, runs them (if safe), and returns output.
4. Complex goals trigger the planner agent, which creates a step‑by‑step plan and delegates sub‑tasks to the appropriate agents.

File handling

• Uploaded files appear under WORK_DIR/.user_uploads/ with automatic name disambiguation.
• The system respects a 50 MB per‑file limit.

Limitations & status

• Early prototype – agent routing may mis‑assign tasks; explicit prompts help.
• Form‑filling in the browser is experimental.
• Voice features (listen, speak) are still under development and work only in CLI mode.
• No formal roadmap or funding; community contributions are encouraged.

License

• GPL‑3.0 (see badge in README).

Where to get help

• Discord: https://discord.gg/8hGDaME3TC
• GitHub Issues & Pull Requests
• Multi‑language READMEs (Chinese, French, Japanese, etc.) for broader community.

---

Bottom line: AgenticSeek is a self‑hosted AI assistant that combines web‑search, code generation, and task planning into a single, privacy‑preserving package. It is suitable for users who have a capable GPU and want to keep all AI interactions on‑device, while still offering the option to fall back to cloud APIs when needed.