Our previous comparison named a winner, narrowly: Hermes Agent, ahead of OpenClaw, on the ground of local, durable agentic work. The logical question follows on its own: how does that free champion hold up against a paid rival of the same kind? Manus is the most directly comparable answer available today — an autonomous agent able, since March 2026, to reach your files and your terminal, exactly as Hermes does. The difference is that Manus is paid for in credits, and that it has just come through a rare geopolitical saga: bought by Meta for $2 billion, an acquisition Beijing blocked and then fully undid. Here is the match, twists included.
The verdict, in short
Across our 7 criteria the score is 3-3, with one tie — but Hermes Agent wins on the tie-break. Manus wins clearly on measured capability, native Windows access and ease of getting started: it is the more finished product, built for a general audience. Hermes Agent wins on what matters most when you hand your files to an agent that runs alone, over time: price, real control of your data, and markedly more stable governance — Manus has just been through an acquisition cancelled by government order.
| Criterion | Hermes Agent | Manus | Advantage |
|---|---|---|---|
| Price and business model | Free software (MIT); only tokens and hardware cost anything | $20 to $200/month in credits; the cost of a task is never announced in advance | Hermes |
| Measured agentic capability | No equivalent public third-party benchmark | GAIA: 86.5% / 70.1% / 57.7% (levels 1-3), at launch | Manus |
| Native file and terminal access on Windows | Windows still « experimental », WSL2 recommended | « My Computer », native on Windows and macOS since March 2026 | Manus |
| Security and control of data | Self-hosted, containers, you know where your data goes | Documented prompt leak; stated jurisdiction (Singapore) ≠ actual servers (China) | Hermes |
| Governance and durability | Open source MIT; keeps working even if Nous Research disappears | Meta acquisition blocked by China, undone on 11 June 2026; founders barred from leaving the country | Hermes |
| Ecosystem and integrations | MCP, VS Code/Zed/JetBrains, 6 messaging channels | Web, mobile, desktop, Telegram (suspended by Telegram shortly after launch) | Tie |
| Ease of getting started | Command-line script, developer profile required | Web/desktop app, no technical skill required | Manus |
Le Recul verdict. Manus is objectively the more finished and more accessible of the two — but « finished » and « accessible » do not mean « the one you should hand your files to without thinking ».
Hermes Agent wins on the tie-break (3-3, 1 tie), because the three criteria it wins — price, control of data, governance — are precisely the ones that matter most when you give an agent your shell and your files and let it work on its own, over time.
Manus remains a defensible choice if you are not a developer, if you want a finished product rather than a project to configure, and if you accept paying for that comfort — while watching your credit consumption.
Two worlds: the community project against the Sino-Singaporean startup
You have to know where these two tools come from to understand why they are not quite playing the same match.
Hermes Agent is published by Nous Research. Under the MIT licence since early 2026, the project passed 214,000 GitHub stars and nearly 40,000 forks in early July 2026. Development remains very active: version v0.18.0 (1 July 2026) added « Mixture-of-Agents » patterns, completion contracts and goal verification proofs; the « Quicksilver » update (20 July 2026) cut time to first token by around 80% and made the desktop app 14 times faster at rendering streamed text. One fact to fold into the equation: Nous Research is itself mid-fundraise, seeking at least $75 million at a $1.5 billion valuation, led by Robot Ventures with USV participating — Hermes is therefore no longer quite the disinterested community project it might have looked like a year earlier, even if the software itself stays free.
Manus comes from an entirely different world. The agent is built by Butterfly Effect, founded in Beijing in 2022 by serial entrepreneur Xiao Hong, with Yichao « Peak » Ji (who created the Mammoth browser at 17, after dropping out of high school), Henry Yang and Tao Zhang. The company passed $100 million in annual recurring revenue in December 2025, and moved its headquarters from Wuhan and Beijing to Singapore in mid-2025 after a Series B round — a separate Singaporean structure taking over operation of the product outside China. That is where the story turns: in December 2025, Meta announced the acquisition of Manus for around $2 billion. An acquisition that would never see completion.
The Meta-Manus saga: a deal bought, then undone by Beijing
This is the angle most AI agent comparisons ignore entirely, and it deserves telling in detail, because it says something important about what « choosing an AI agent » actually means in 2026.
As early as January 2026, days after the announcement, the Chinese authorities opened an investigation into the deal, on national security and export control grounds. The central grievance: Meta and Manus had not notified the Chinese authorities before closing in December, and Beijing held that Manus’s intellectual property — developed on Chinese soil before the move to Singapore — remained subject to Chinese export control law, whatever the company’s current registered office. On 27-28 April 2026, the National Development and Reform Commission (NDRC) formally ordered the deal cancelled. The acquisition, already closed at that point, was fully undone on 11 June 2026: Meta had to be repaid, Manus became an independent company again, and its early investors (Tencent, ZhenFund) may, according to several analysts, end up better off than with a Meta exit.
The most striking detail, almost unnoticed in Western coverage: in March 2026, mid-procedure, Xiao Hong and chief scientist Ji Yichao were barred by the authorities from leaving Chinese territory. That is not an anecdote — it is concrete proof that an agentic AI company, even relocated to Singapore, stays exposed to great-power politics in a way a self-hosted open-source project structurally is not.
What each one can do
On paper the two promise the same thing — an agent that plans, acts and delivers finished work without constant supervision. The architectures diverge sharply.
| Capability | Hermes Agent | Manus |
|---|---|---|
| Underlying model(s) | Your choice (Claude, GPT, Gemini, Qwen, DeepSeek, a local model…) | Fixed: Claude (Opus 4.x line) + a fine-tuned Qwen for cheap sub-tasks |
| File and terminal access | 40+ tools, snapshots before modification | « My Computer » (16 March 2026): reading, editing and organising files, terminal commands, per-command approval |
| Parallel tasks | 3 sub-agents in parallel by default | Up to 20 concurrent tasks (paid plans) |
| Memory and learning | Persistent, self-generated skills (claimed ~40% gain, unverified) | Persistent memory per agent, custom skills (« Agents » feature) |
| Messaging channels | 6 native channels | Telegram (16 Feb 2026, suspended shortly after), web, mobile |
| Code editor integration | VS Code, Zed, JetBrains (ACP protocol) | Not code-specialised; oriented to research, reports, deliverables |
| Public benchmark | No equivalent third-party benchmark identified | GAIA (levels 1-3): 86.5% / 70.1% / 57.7% at launch, against 74.3% / 69.1% / 47.6% for the reference Manus cited at the time |
Manus’s GAIA score needs one methodological caveat: it dates from the product’s launch in 2025 and has not been republished with the same figures since — treat it as evidence that Manus aimed high from the start, not as a fresh measurement of its current capability on the Opus 4.x line it now uses. No independent, dated 2026 republication was found to confirm whether the gap has held, narrowed or reversed against today’s best agents.
That table reveals a fundamental difference: Manus ships a fixed backend, where Hermes lets you plug in the model of your choice. That choice is not neutral — it radically changes the agent’s cost, speed and reliability.
The price, without the small print
Hermes Agent is free software. No subscription for the tool itself; you pay only for the tokens of whichever model you plug in, or for electricity if you run locally, plus, if you want it, the optional subscription to the Nous Portal gateway (around $20/month according to third-party sources).
Manus runs on credits, across four tiers:
| Plan | Price | Credits |
|---|---|---|
| Free | $0 | 300 credits/day |
| Pro / Standard | $20/month | 4,000 credits/month |
| Customisable | $40/month | 8,000 credits/month |
| Extended | $200/month | 40,000 credits/month |
(Annual billing: roughly 17% saving on all paid plans; all paid plans include a 300-credit daily top-up and up to 20 simultaneous tasks.)
The real problem is not the list price, it is the opacity of consumption. A simple conversation costs 5 to 15 credits; a deep research task can swallow 500 to 900 credits, sometimes more than 1,000 — and Manus does not show an estimated cost before launching the task. A user on the $20/month plan can therefore burn a whole month’s quota on a handful of ambitious tasks without seeing it coming.
Security, data and governance: the real dividing line
This is where the comparison earns its keep, and where the gap between the two philosophies becomes concrete.
Hermes Agent inherits from its open-source design a simple but decisive property: you know where your data goes, because you (or the model provider you chose) host the execution. Its container isolation (Docker, Singularity) and default authorisation model hold.
Manus accumulates several distinct warning signs, each documented separately:
- A researcher demonstrated that a simple prompt could leak part of the internal system prompt, by asking it to display the contents of a directory — a sign that the isolation between internal instructions and user input is not airtight.
- Before the Meta acquisition, Manus’s privacy policy claimed Singaporean jurisdiction while data in fact transited through servers located in China — a documented gap between the promise and the technical reality, the kind of false sense of security that is more misleading than a plainly stated risk.
- Users report automated account suspensions for « risky behaviour » (fake friend invitations being flagged, for instance), with no apparent relation to their actual use — a detection system that sometimes seems to punish legitimate accounts.
- And, of course, the Meta episode: an acquisition cancelled by government order, with executives barred from leaving the country during the procedure. Nothing even remotely equivalent on the Hermes Agent side.
None of this means Manus is a malicious tool — it is a serious commercial product used by hundreds of thousands of people. But for a use that means handing your files and your terminal to an autonomous agent, the sum of these signals weighs clearly in favour of the solution you host and control yourself.
Ease of getting started: where Manus takes the lead back
Conversely, it is only fair to say what Manus does better. User accounts (freelancers, small companies of 10 to 50 staff) describe a tool able to produce a complex report in under 30 minutes, with an interface usable with no technical skill at all — no terminal, no API key to configure, no install script. That is the exact opposite of Hermes Agent, whose command-line installation and explicit configuration (memory, skills, model) assume a genuine developer profile.
The same accounts temper the picture: Manus can « go round in circles » on a badly defined task, and there are reports of file download bugs and instability on long-running operations — teething problems comparable to those documented on the OpenClaw side in our previous comparison.
What to remember
- Final score: 3-3, one tie — Hermes Agent wins on the tie-break, on price, control of data and governance.
- Manus wins clearly on three grounds: measured capability (GAIA), native Windows access (since March 2026, ahead of Hermes), and simplicity for a non-developer.
- Manus has just been through a $2 billion Meta acquisition blocked by China and fully undone on 11 June 2026 — with founders barred for a time from leaving the country. A level of geopolitical risk a self-hosted open-source tool structurally does not carry.
- Manus’s real cost is unpredictable: a research task can consume 500 to 1,000 credits with no prior warning, on a plan that only offers 4,000 a month for $20.
- Hermes Agent is no longer a purely disinterested community project: Nous Research is currently raising at a $1.5 billion valuation — the software stays free, but the company behind it is on a commercial trajectory too.
The figure to remember
11 June 2026.
That is the date on which Meta’s acquisition of Manus — announced six months earlier for around $2 billion — was fully undone, by order of the Chinese regulator. A useful reminder before handing your files to a paid AI agent: behind the subscription there is always a company, and sometimes a contest between states that can, overnight, redefine who actually owns the tool you are using.
Our verdict. Manus is the more mature, better measured and more accessible product of the two — a real competitor, not a gadget. But on the ground this comparison is fought on — handing your shell and your files to an agent that works alone, over time — Hermes Agent keeps the advantage: free, self-hosted, and out of reach of a balance of power between states that can flip a commercial product overnight. AI promises turnkey autonomy; Le Recul checks that it is not charging you control of your own data in return.
For more on agents that work alone, see our other AI comparisons and our tested promises.