AgenticSeek

AgenticSeek is the private Manus you run yourself, and the trait that decides it is the graphics card. The hardware table is unusually blunt about this: 14B on 12GB of VRAM is usable for simple tasks, 32B on a 24GB card succeeds at most of them, and 7B hallucinates. If you own that machine you get browsing, code execution and voice with no API bill and nothing leaving your desk.

Pick it when privacy is the actual requirement and the GPU is already paid for. Pick Manus if you want the same shape of assistant without buying a card, and accept that your work goes somewhere else to get done.

The failure is dispatch. You do not choose which agent runs. A router reads your sentence and allocates one, and the project states that this routing might not always allocate the right agent based on your query. The documented remedy is user discipline: ask for a web search instead of a question. Form filling is separately marked experimental and might fail, which is the step a browsing assistant exists for.

A misrouted request returns a bad answer, not an error, so you cannot tell them apart. What it does right: both limits sit in the README above the demo rather than buried in a closed issue.

1. Context arrives through a metasearch engine the project ships itself. Docker Compose starts SearxNG on port 8080 with Redis behind it and a frontend on 3000, so retrieval requires no vendor key. 2. Decomposition is handed to a planner agent that splits a request into steps for others. 3. Actions are programs written and executed on the host, confined to the directory named in work_dir. 4. Verification is execution: the program runs or it does not.

No benchmark accompanies any of this. The hardware table stands in for one, which is a candid substitute rather than a measured claim.

There is no cap table to read. The maintainer writes that this began as a side project with zero roadmap and zero funding, that he is not a startup or affiliated with any organisation, and that two friends maintain it with him. The monetisation on the page is a proxy vendor's banner and a search API banner, one carrying a promo code.

Distribution is real and the name is searched, which is the only asset here. The likely outcome is not an acquisition; it is the code absorbed into another product, or a maintainer taking a job. Position: use it, mirror the repository, depend on nothing.

The demo is a local assistant that browses and writes programs. Procurement never opens: there is nobody to sign with, no single sign-on, no SCIM, no audit trail and no retention policy, because there is no counterparty. Per seat the software costs nothing and the workstation hardware does, which lands in a capital budget nobody prepared for sixty people.

It cannot run unattended in our pipelines, so it contributes nothing to review or delivery. Onboarding is a pinned Python 3.10.x, Docker Compose and a per-person configuration file, so budget a day of help each. Support is one maintainer and a chat server. Not yet.

GPL-3.0, so a fork has to stay open, which suits me. Everything I care about is one file: is_local decides whether my box or a vendor answers, provider_name takes ollama, lm-studio or any OpenAI-compatible server, and provider_server_address points at whichever machine holds the card. The model lives on the tower under the desk while the assistant runs on the laptop, and no code changes to make that true.

The gap is protocols. No MCP client, no MCP server, so every extra capability is a Python file I write against the repository rather than a server I register. More work, still mine.