Price terms as published

Each line is the vendor’s own wording, split from the price capture checked on 2026-08-30. Nothing here is rounded, converted or re-worded — the full sentences stay in the review below. Lines that carry only a price continue the plan named on the line above, in the vendor’s own order.

2 price terms for Mobius
#Plan / termChecked
1Free for non-commercial self-hosting + model API costs2026-08-30
2commercial use requires a paid license2026-08-30

This is the first entry in AIListScout’s hands-on test log series. What that means, in plain terms: we install the tool ourselves, run it on real work, publish the log as we go, and tell you exactly where we are in the process at every step. No demo-as-review, no “trust us, we played with it for an afternoon” — the dates below are the actual dates.

Status as of 2026-08-30: deployment verified and first run completed (Day 0-1). Full two-week review in progress; the checklist at the bottom is the published test plan and will be updated as items are checked off.

The tool under test

Mobius describes itself as a self-evolving agent OS: a persistent agent that remembers, schedules, and — the headline feature — can modify its own interface and behavior when you ask it to. It self-hosts (you bring your own model API key). We installed it on a Windows 11 box via WSL Ubuntu 24.04, Direct route (no Docker), from the main branch cloned on 2026-08-29 (v1.9.0, ~286 stars at clone time).

286 stars matters as context: this is a small project in agent-OS terms, not a mature one. Reviewing it hands-on is exactly why this site exists — the trending-list writeups make “self-evolving agent” sound like a finished product category, and it is not.

What the README promises vs. what actually happens

The README’s headline is essentially “one command and it’s running.” Here is the real timeline, minute by minute, from a fresh box:

  • 19:28 Docker Desktop installed silently, but the engine would not start: the virtualization components had just been enabled, wsl-bootstrap crashed with 0xc00000fd, and a reboot was required. Decision: skip Docker entirely, take the Direct Linux route the README also documents. (First time “one command” was wrong.)
  • Node 22 required. The official nodesource mirror would not connect (Cloudflare IP blocked on this network) — switched to npmmirror, Node 22.15.
  • Repo cloned (767 files, ~1 minute). Config generation scripts passed first try (“Configuration OK”, auto-generated random credentials).
  • npm install for root + frontend succeeded (NTFS-over-WSL is slow but it works).
  • First start, new wall: the DeepSeek harness hard-codes ~/.nvm/versions/node/v22.22.2/bin/node — installing the right Node version is not enough, it must live at that exact nvm path. Fixed with a symlink + env override. (README mentions none of this.)
  • Node 22.15 failed its >= 22.19 check → reinstalled 22.22.0.
  • Second start: the aimux bridge script requires uv; Ubuntu 24.04 pip is PEP-668 locked, so the astral.sh installer was used (as root, it lands in /root/.local).
  • Third start: the product’s own npm run build kept getting killed (exit −1; slow NTFS
    • timeout). The workaround the README does not mention: build the frontend manually first, then start.
  • PM2 then went online (mobius-system / bridge / tokenproxy, frontend on :33316, HTTP 200 from Windows side), and a user login with the bootstrap password worked.

Total: ~55 minutes, six deployment obstacles, none of them in the README. The honest summary: “one command” assumes a machine that already has the exact Node version at the exact nvm path, a functional uv, a faster filesystem, and no proxy surprises. On a fresh box, the README is a promise, not documentation.

The seventh wall, and the genuinely good part

Day 1’s first task was a doozy: the UI defaults to “route through proxy” turned on for the DeepSeek connection, and the harness then insists on a 2012-era proxychains 3.1 (a tool so old apt has nothing newer), which hangs dead in WSL — the agent session refused to start with “proxy prerequisites missing: /usr/bin/proxychains, /root/proxy_claude.conf” (the file did not even exist yet). Diagnosis and fix: install proxychains, fabricate the config pointing at the local Windows proxy, still dead; test direct connection → 401 only on bad auth, i.e. direct works; set use_proxy: false in model-access.json, restart PM2. Done. The README says nothing about a proxy toggle existing at all.

Then the first real task: ask the agent to self-diagnose its model access. The output was correct and, frankly, better than expected:

  • DeepSeek Harness (deepseek-chat) connected and in use;
  • the GLM-4.7-FlashX light API entry had a model name but no Base URL / keys — incomplete credentials, correctly flagged;
  • it volunteered that the light-API bypass lane is architect-only and does not enter the normal agent workflow, and recommended importing GLM through the proper channel;
  • no global default model was set, so it correctly fell back to the only usable one.

The notable thing is the corrective instinct: when the user’s configuration was in the wrong place, the agent said so instead of pretending. For a 286-star project’s first run, that is an honest signal that the core loop (self-check, report, correct) is real, not demo footage.

What’s honest about this tool so far

The rough edges are concentrated in installation, not in first-run behavior. The proxy landmine is the worst documentable offense — a default toggle that silently breaks the one thing you’re trying to do, with no README hint. The nvm hard-code and the uv requirement are the kind of sharp edges you expect from a young project, but they compound: six of them in a row turns a 10-minute install into an hour.

On the upside: deployment failure modes were diagnosable — every wall had a traceable cause and a documented-in-the-repo fix path. Nothing required black magic. That is the difference between “rough” and “broken,” and it is why the tool gets a provisional pass to continue the two-week test rather than a rejection.

Test plan (published, will update)

The next two weeks on this install, in order:

  • DeepSeek model integration — settings UI paths and required fields
  • Create a project, assign a task to the agent in natural language
  • Conversation retrieval — keyword search across history, jump back and resume
  • Self-evolution: Ctrl+Shift+E to select UI, state a requirement, watch it change itself
  • Skill/memory layer management (user-level vs project-level, “emphasize” pinning)
  • Multi-endpoint: TUI (npm one-liner) against the same server
  • Division of labor vs. Hermes/Codex — same task on both sides, side by side

Scoring dimensions (fixed before testing): install friction / UI completeness / whether self-evolution is genuinely usable / documentation honesty / Chinese-language friendliness / overlap with the Hermes·Codex workflow already running this site.

Verdict (provisional)

3/5, and explicitly revisable. The install friction is real and under-documented — that alone costs it a full star from a “one command” promise. But the first-run self-diagnosis is better than most mature tools deliver, the failure modes are all diagnosable, and the self-evolution headline has not yet been tested either way. The two competing agent harnesses on this site (Hermes for authoring, Codex for review) remain the daily drivers for now; Mobius is the challenger earning a two-week window. When the checklist above is done, this rating gets revised and this paragraph gets rewritten.

Methodology & sources

Hands-on install and first run performed 2026-08-29–30 on this site’s Windows 11 + WSL Ubuntu 24.04 box; timeline, errors, and fixes recorded live in the test log. Star count and version at clone time from the GitHub repo on 2026-08-29. Licensing, read from the repo’s LICENSE at clone time: not MIT — Mobius ships under its own “Mobius Open Source License” (Guoke Zhisuan (Beijing) Technology Co., Ltd., 2026), which is source-available for non-commercial use only; any commercial deployment requires a separate paid commercial license from the vendor. Price: no paid tier inside the product, model API costs borne by the user; “free” here means free for personal/non-commercial self-hosting. This is a log in progress, not a final review — the DeerFlow review next door is the official-only comparison; this piece is the hands-on one, and the two should not be read as the same level of evidence.