AgenticSeek: A Private, Local Alternative to Manus AI (GitHub, Scanned)

AgenticSeek is an autonomous agent in the style of Manus that is built to run on your own machine. You give it a task in a web interface or the terminal, and it picks the right sub-agent: one that searches and browses the web through a private SearXNG instance and a real Chrome, one that writes and runs code in Python, C, Go, Java and other languages, one that works with files in a folder you choose, and a planner that splits bigger jobs across them. Voice input and output are experimental.

The point is privacy and cost. It is tuned for local reasoning models served by Ollama, LM Studio or any OpenAI-compatible server, so no prompt or file has to leave your computer, and there is no monthly bill. Cloud APIs such as OpenAI, DeepSeek, OpenRouter, Google and Anthropic are optional for people without the hardware.

The limits are real: the README asks for a GPU able to run a 14B model at minimum, setup involves Docker, a .env file and a config.ini, and the author describes it as a side project with no roadmap or funding. It is GPL-3.0 with about 27,500 stars.

• Repository: github.com/Fosowl/agenticSeek
• Licence: GPL-3.0 (GNU General Public License v3.0)
• Language: Python. Stars: 27.5K. Forks: 3,088. Last push: Oct 22, 2026.
• Scan: safe, Oct 26, 2026, commit 4f5f924

Who it is for

Tinkerers with a capable GPU who want to try an autonomous browsing and coding agent without sending their data to a cloud service, and anyone curious how agents like Manus work underneath.

Getting started

1. Clone the project and create your settings file (then edit .env and config.ini as the README describes)

git clone https://github.com/Fosowl/agenticSeek.git && cd agenticSeek && mv .env.example .env

2. Start a local model provider, for example Ollama

ollama serve

3. Start the services and web interface (Windows: start start_services.cmd full)

./start_services.sh full

Then open http://localhost:3000. Docker must be running, and WORK_DIR in .env is the folder the agent can read and write, so point it at a dedicated folder rather than your home directory.

Safety scan

We cloned Fosowl/agenticSeek at commit 4f5f924 on Oct 26, 2026 and ran the checks described on the GitHub Tools page: credential patterns, decode-and-execute code, install-time scripts, committed binaries, risky CI workflows, every host the code talks to, known vulnerabilities in pinned dependencies, and project hygiene. A person read every hit. This is what we found.

• No secrets and no bare-IP URLs. Of the 9 pattern hits, only llm_server/install.sh actually pipes a download to a shell (Ollama's official installer, for the optional remote model server); two are echo lines suggesting the uv installer, and six are AgenticSeek's own command checker, sources/tools/safety.py, and its tests, which exist to block commands like curl | sh.
• start_services.sh and searxng/setup_searxng.sh use sudo for Docker on Linux. In Docker mode the agent's commands run as an unprivileged user that owns WORK_DIR.
• The browser runs with stealth_mode on by default, using undetected-chromedriver to avoid bot detection; the bot-test pages among the outside hosts (bot.incolumitas.com, fingerprintjs) are there to check it. import_user_profile is off by default; turning it on copies your own browser's cookies and sessions into the agent's browser, which the README itself warns hands the agent your mail and bank logins.
• 413 known advisories, 17 rated critical: 234 in uv.lock, the Python backend that actually runs (certifi, anyio, nltk and others), and 177 in the React front end's package-lock.json. Pinned versions this old are a maintenance gap rather than a sign of bad intent.
• No GitHub workflows at all, so no automated tests or checks run on changes. A security policy and licence are present.

What the scanner counted

The raw findings

Every hit the scanner wrote out, with a link to the exact line at the scanned commit. Secrets candidates are redacted.

By the numbers

Lines by language: Python 14.1K, JSON 7,849, Markdown 6,529, JavaScript 2,760, CSS 2,178, Shell 519.

Questions

Is AgenticSeek free?

Yes. AgenticSeek is GPL-3.0 open source and free. Running it with local models costs only electricity; if you choose a cloud API instead, that provider bills you for usage.

What hardware do I need for AgenticSeek?

The README sets the minimum at a GPU that can run a 14B model such as Qwen, DeepSeek or Magistral, so plan for a recent gaming GPU or an Apple Silicon Mac with plenty of unified memory. Smaller models tend to fail at planning and tool use. Without that hardware, use a cloud API.

Is it safe to let AgenticSeek run code and browse?

Treat it like handing your keyboard to an assistant. Keep WORK_DIR pointed at a dedicated folder, run it in Docker mode, leave import_user_profile off unless you understand that it shares your logged-in sessions, and watch what it does on the first few tasks.

This post is part of GitHub Tools, where every repository is cloned and scanned before it is written up. The scan is a snapshot of one commit on one day; the repository has moved on since, so check it before you install.