AgenticSeek
A local task assistant for web research, file work, and code execution.
Per-dimension scores and reasoning
Evidence shows: The project claims local operation, but configuration allows setting multiple API keys (e.g., OPENAI_API_KEY) without clarifying their scope. The working directory (WORK_DIR) can be read/written by the agent, but no least-privilege mechanism is provided. User confirmation is absent; the agent can execute actions autonomously. Data flow transparency is limited; README mentions local data storage but does not detail data flows. Sensitive data handling: API keys are stored in plaintext in .env, with no encryption or secure storage advice. Dependency security: dependency list is provided, but no vulnerability scanning or security audit information. External effects: the agent can browse, fill forms, execute code, but risks are not clearly communicated. Rollback is not mentioned. Source attribution: project is open-source, but contributor information or code provenance is not detailed.
Evidence shows: The project includes multiple test files covering agent loops, browser parsing, etc., with detailed test cases, indicating some self-consistency. Dependency availability: dependency list is clear, but no lock file or version pinning strategy, risking inconsistency. Failure messages: tests cover error handling, but no user-friendly error messages or logging mechanisms are provided.
Evidence shows: README provides multiple usage scenarios (local LLM, API providers, remote server) with detailed configuration, indicating broad audience and scenarios. Capability boundaries: README mentions code execution, web browsing, but does not clearly limit or explain safe operations. Trigger precision: agent auto-selects based on query, but README admits routing may be inaccurate, requiring explicit user instructions. Environment fit: supports multiple OS and Docker, but no detailed system requirements or compatibility notes.
Evidence shows: README is well-structured with install, config, usage, troubleshooting sections, good information architecture. Install notes are detailed. Naming stability: project and config names are consistent, but no version history or changelog. Examples and FAQ are rich. Known limitations: README mentions experimental form filling, English-only STT, but not exhaustive. License is GPL-3.0, compliant. Versioning/changelog missing. Maintenance responsibility: README states side-project, no roadmap, but provides contact.
Evidence shows: Output usability: agent can generate code/files, but no output format or quality guarantees. Marginal value: provides local, privacy-preserving AI agent, unique vs cloud. Cost-benefit: claims zero API cost, but requires user hardware and Docker; cost-effectiveness depends on user environment.
Evidence shows: Claims in README (fully local, privacy) lack implementation details or test evidence. Cross-source corroboration: no third-party verification or independent test results. Fact-inference separation: README mixes product features with speculative descriptions without clear distinction.
- API keys are stored in plaintext; use environment variables and avoid committing to version control.
- Agent can autonomously execute code and browse the web, posing security risks; run in isolated environments.
- Dependencies are not pinned, risking supply chain issues; use lock files.
- Project is a side-project with no roadmap; maintenance may be unstable.
What does this agent do, and when should you use it?
AgenticSeek is a self-hosted general task assistant with both a Docker web interface and a host-based CLI. Its web mode starts a frontend, backend, SearXNG, and Redis through Docker Compose and is accessed at localhost:3000; CLI mode runs cli.py with uv. It routes requests to an agent that can search, read, and extract web information through SearXNG, while web-form filling remains experimental. It can also read and operate on files in the configured WORK_DIR and write, debug, and run programs in Python, C, Go, Java, and more. The primary deployment path uses Ollama, LM Studio, or a local OpenAI-compatible service, with optional support for OpenAI, Google, DeepSeek, Hugging Face, TogetherAI, OpenRouter, and MiniMax providers.
A user submits a task in the localhost:3000 web UI or through uv run cli.py, after which the routing system selects an agent. Web tasks connect to SearXNG using SEARXNG_BASE_URL from .env to search, read, and extract information, with experimental form-filling support. File and coding work is scoped to WORK_DIR: documented examples include finding and renaming JPGs, saving file lists, writing cafe addresses to a text file, and writing and running Go or Python programs. Model access is configured in config.ini with provider_name, provider_model, and provider_server_address; local options include ollama, lm-studio, a local openai-compatible service, and the self-hosted server option.
- A person with a local GPU who wants Ollama or LM Studio to handle web research and local file work without sending conversations and workspace data to a cloud service.
- A developer who wants to turn a natural-language request into a Python, Go, C, or Java program saved to a chosen WORK_DIR and then run it.
- A researcher or operations worker who needs SearXNG-based web search and extraction, with results saved locally as text or CSV files.
- A user deploying a personal assistant with Docker who wants the localhost:3000 frontend while running the model service on the host or another server.
- A terminal-oriented user who wants to use
uv run cli.pyand optionally try the experimental, CLI-only voice input.
What are this agent's strengths and limitations?
- It offers a local deployment path with Ollama, LM Studio, or a local OpenAI-compatible model service, so cloud APIs are optional.
- Its Docker web mode starts the frontend, backend, SearXNG, and Redis together and exposes a localhost:3000 interface.
- It combines web search and extraction, workspace file operations, code generation, and code execution rather than only providing chat.
- Its
provider_nameandprovider_server_addresssettings support local, remotely self-hosted, and multiple cloud model providers.
- Local use requires substantial hardware: the README recommends at least a GPU able to run a 14B model, does not recommend 7B models, and positions 32B-plus models as better for most tasks.
- The first Docker startup may download images for up to 30 minutes, and the backend may take about five minutes on its first run.
- Web-form filling is experimental, and the early-prototype routing system may select an unsuitable agent for a request.
- Using host Ollama from the Docker backend requires Ollama to bind to
0.0.0.0:11434and may require Docker host mapping and firewall configuration.
How do you install or deploy this agent?
Install Git, Python 3.10.x, Docker Engine, and Docker Compose. Run:
git clone https://github.com/Fosowl/agenticSeek.git
cd agenticSeek
mv .env.example .envSet a valid WORK_DIR in .env; use SEARXNG_BASE_URL="http://searxng:8080" for Docker web mode and SEARXNG_BASE_URL="http://localhost:8080" for host CLI mode, changing the CLI URL to match SEARXNG_PORT if that port is changed. Configure a local provider in config.ini, for example provider_name = ollama, a model name, and its server address; API keys may be left empty for local models. Start web mode with ./start_services.sh full, wait for the backend health check, then open http://localhost:3000/.
How do you use this agent?
For web mode, start ./start_services.sh full, then open http://localhost:3000/ and submit an explicit request such as “Search the web for top cafes in Rennes, France, and save a list of three with their addresses in rennes_cafes.txt.” For CLI mode, run ./install.sh, point .env SEARXNG_BASE_URL to the host-mapped port, run ./start_services.sh, then run uv run python -m ensurepip and uv run cli.py. Type goodbye in the CLI to exit; setting listen = True in config.ini enables experimental English speech input.
How does this agent compare with similar options?
The project positions itself as a local alternative to Manus AI, emphasizing local or self-hosted models, local workspace handling, and independence from cloud services. The repository does not provide evidence of feature parity or benchmark comparisons with Manus.
FAQ
Do I need to buy a cloud-model API?
What hardware is needed for local use?
Which files can it access or change?
WORK_DIR as the crucial directory in which AgenticSeek reads and writes files. Set it to a valid, accessible path on the local machine.Why cannot the Docker backend reach Ollama?
host.docker.internal; an Ollama service bound only to 127.0.0.1 rejects Docker bridge connections. Start it with OLLAMA_HOST=0.0.0.0:11434 ollama serve.