Manus

Manus suits broad tasks: it researches, writes, builds a website and drops a slide deck in the same session, which no coding agent here will do, and the daily trait is that the same sandbox holds all of it, so the report cites the data the app just fetched. For demo day, that is the whole pitch.

You will not like it on a real repository, where a generalist wanders through a specialist's job and the diff shows it. Pick it for demo day and the one-off deliverable. Pick Devin or Codex cloud for the sprint, where the output is a pull request into code that already exists.

The risk is the write path. The sandbox has internet access and the project syncs two-way to a private GitHub repository, so a bad page becomes a bad push, and an agent that browsed the wrong instructions has write access to your history. Nothing in the docs describes a review gate between them.

The consequence: point the sync at a repository nobody deploys from, and treat the branch as untrusted input until a human diffs it. A documented approval step before sync would change this verdict. What it does right: each task runs in its own sandbox with a persistent file system, so one runaway task cannot corrupt another.

Manus publishes no benchmark and names its own models, Manus 1.6, Manus 1.6 Max and Manus 1.6 Lite, without a methodology for how they differ, so a buyer cannot tell whether Max is a bigger model or a longer budget. Documented: 1. A sandbox where software can be installed at run time. 2. Wide Research, which fans one task out across parallel agents. 3. Scheduled tasks.

The consequence is that capability is described entirely by what the agent may do, never by how often it does it correctly, and parallelism multiplies whatever the base rate is. The observation: parallelism is not evidence.

The cap table is the story. Butterfly Effect agreed to sell Manus to Meta for more than $2B in December 2025, Chinese regulators blocked the deal in April 2026, and by August 2026 the company said it would stay independent. That is a founder who wanted out and a regulator who said no, which is a company whose strategy is now being written by lawyers.

The next acquirer has to clear the same regulator, which narrows the list considerably. The pivot is toward enterprise in a market that will let it sell. Position: rent it, never build on it, and reprice the moment the ownership changes again.

The demo is an agent with its own computer. Procurement: Team at $20 per seat is $1,200 a month for sixty, then the credits, the real bill, which the seat price does not include and the page does not forecast; concurrency runs from one task on the free plan to 20 at the $200 tier, so throughput is bought per user.

No SSO, SCIM, audit log or retention terms appear on the page, and a sandbox with internet access and our GitHub credentials needs all four before a pilot. CI reaches it only through the REST API, one more integration to own. Onboarding is a login. Not yet.

Closed and then some: no model choice, no key field and no local anything, so the sandbox is theirs, the model is theirs, and the only thing I own is the prompt. What I can script: a REST API for driving tasks from my own code, and custom MCP servers, which only an admin can attach, so personal accounts are shut out of the one extension point.

Desktop apps exist for macOS and Windows, the same box with a frame. Nothing to fork, nothing to fix, nothing to run at home. Useful as a service I call, never as a tool I hold.