The only comparison on this page that is actually a choice.
Devin, Codex cloud, Claude Code on the web, Jules, Replit Agent. Everything else here composes with intentic. These do not, because they answer one question differently: whose computer holds your source.
A hosted agent you give a task and a repository to. It runs on the vendor's infrastructure, in the vendor's sandbox, and hands you a pull request.
The experience is genuinely good and the setup cost is genuinely zero. That is the trade: your code, your keys and your production access, living in someone else's sandbox for the duration.
intentic's position is that the browser experience was never the part that required giving that up. The sandbox runs on hardware you own, and the platform sits off the command path.
Where we agree
An agent should be supervised from a browser, not a terminal. It should run several jobs in parallel, show its plan before it acts and its diff before anything lands. Every one of those is a cloud-agent idea and intentic took all of them.
Where they differ
Custody, stated architecturally
Your browser holds the token that commands the sandbox; the platform never does. It stores identity, a URL, billing state and grants, encrypted with no decrypt path. A breach reads a URL and reaches nothing.
Real access, safely
An agent usually stalls at 80% because it cannot reach what it needs: the staging database, the error tracker, the box the service runs on. Capabilities give it those, with the credential kept inside a sandbox you own.
Your subscription, no meter
Cloud platforms bill you for the model and for the machine, usually with a markup on both. intentic runs on the account you already pay for, on hardware you already own, for a flat fee that never meters tokens.
An exit that is not a migration
Your repos are ordinary git on your own disk. The sandbox and CLI are MIT on GitHub and drive from the command line without ever signing in. Leaving is deleting an account, not extracting a workspace.
Head to head
| Dimension | intentic | cloud agent platforms |
|---|---|---|
| Where the agent runs | your hardware | the vendor's infrastructure |
| Who can read your source | you; the platform has no path to it | the vendor, for the duration of the run |
| Where your service credentials sit | inside your sandbox | in the vendor's secret store |
| Model billing | your own subscription, never metered by us | the vendor's plan, usually with a markup |
| Compute billing | hardware you already own | the vendor's, metered or bundled |
| Supervision | fleet board, plan mode, diff review, transcripts | fleet board, plan mode, diff review, transcripts |
| Setup | Docker, a Google account, one pasted command | ▸ connect a repo; nothing to host |
| Elastic capacity | bounded by the machine you provide | ▸ as many agents as you will pay for |
| Enterprise paperwork | a new product; read the MIT source instead | ▸ established vendors with the certifications |
Pick cloud agent platforms when
You would rather not provide a machine, the repository is not sensitive, your procurement process needs a vendor with the certifications already in hand, and you want to burst to fifty agents on a Tuesday without owning fifty cores.
Every claim above is checkable on www.cognition.ai and in this repo. Where we have it wrong, open an issue and it gets fixed in the next build.