Finding out whether it is any good, on a machine you own
Local model
You have a laptop with some memory to spare and no interest in creating another account before you know whether the thing works.
Your hardware
This is the honest zero, and few agents in this section have one. Nothing signs in, nothing leaves the machine, and the client is under a licence with no paid edition above it. The constraint is stated as a hard requirement rather than a recommendation: the agent refuses at startup any model offering less than 64K of context, so a local runtime left on its default context window will be rejected rather than quietly degraded.
You already pay for a coding subscription
Claude Max 5x
There is a Claude or ChatGPT plan on your card already, and buying a second allowance to run a third-party client feels like paying twice.
$100/mo
You can spend the plan you have — this client authenticates against both, and imports a Claude Code or Codex CLI setup in one command, allowlists and skills included. Read the condition before you count on it, though: the documentation scopes the Claude route to the higher plan tier plus extra usage credits, not to the entry subscription, and that is a different purchase from the one most readers assume.
One bill for the model and everything around it
Nous Portal Plus
You would rather not open accounts with a search API, an image host, a speech vendor and a browser service just to make one agent useful.
$20/mo
This is the tier the vendor pushes, and the argument for it is real: one login covers the model catalogue and the hosted tools the agent calls, which otherwise means five separate signups. The tradeoff is visible in the meter rather than hidden by it. Search, images, speech and the cloud browser draw on the same credit balance as your tokens, so an afternoon of image generation is an afternoon of not coding — the difference from a shared allowance elsewhere being that here you can watch it happen in dollars.
All month, several agents at once
Nous Portal Super
Subagents run in parallel, the gateway is always up, and the monthly credit stopped being a theoretical ceiling some time ago.
$100/mo
Worth knowing what the step up does and does not buy. It does not unlock models, tools or features: the catalogue is identical on every paid tier. What scales is the credit grant, its rollover cap, and a rate limit the vendor describes only as high. So the upgrade decision is arithmetic on your own consumption rather than a feature comparison — which is rarer, and more answerable, than it sounds.
Unattended runs, cron jobs and CI
Your own API key
The agent runs headless on a schedule or in a pipeline, and nobody is watching the terminal to approve anything.
At model rates
A key detaches the bill from any subscription and puts the ceiling wherever your provider put it. Two defaults matter more than the price here. Unattended contexts — cron, single-query and webhook sessions — refuse dangerous commands rather than auto-approving them, which is the right way round. And the default approval mode is not a pattern list but a second model asked to judge the risk, so the safety layer itself consumes tokens that the client's own cost display explicitly does not count.
A team whose source cannot leave its own infrastructure
Hermes Enterprise
Several engineers, a private codebase, and a policy that has to be answered to rather than argued with.
Contact sales
This is the only tier where the answer to "where does our data live" is your own infrastructure; everything below it runs on the vendor's. It matters more here than on most pages in this section, because the default on the consumer side is that inference payloads may be stored and used for training unless somebody finds the account setting that says otherwise. Between the two sits a self-serve team tier that shares one balance with per-member spend caps — useful for controlling cost, not for controlling where the data goes.