Every AI coding tool eventually offers both: bring your own provider key, or buy credits from the tool. The choice is usually presented as a preference. It is closer to an arithmetic problem with a security question attached.
Here are both, using our numbers, including the one that is unflattering.
The rates
Your own key: 0% markup. The provider bills you directly. Codoro takes nothing on the tokens.
Hosted credits: 40% markup on cost, with published per-model overrides. The constant is
CODORO_RETAIL_MARKUP_BASIS_POINTS = 4_000.
Being blunt about what that 40% is: after Stripe's 2.9% plus 30 cents, a 40% markup on cost nets something in the region of 25% of revenue. It pays for one bill instead of five, no key management, and the credit ledger. It does not pay for better models or better rates — the catalogue is the same either way.
If you already have provider accounts, bring the keys. That is not modesty, it is the honest reading of the number. The zero-markup path is not a loss-leader with a catch; it is what a product whose value is the interface rather than the tokens should charge for tokens.
What the plans actually include
| Tier | Price | Included credit |
|---|---|---|
| Hobby | $0 | None — your own keys |
| Pro | $20/mo | $20 |
| Pro+ | $60/mo | $60 |
| Ultra | $200/mo | $200 |
| Teams | $40/seat | $40 per seat, one pooled budget |
Top-ups are $28, $70 and $140, and the credit value equals the price exactly. They never expire and they never auto-charge.
Note what the paid tiers are: the credit included equals the price paid. You are not buying a subscription that unlocks features and separately buying tokens — the subscription is the tokens, at the 40% markup, prepaid. Hobby exists so that the bring-your-own path costs nothing at all rather than being gated behind a seat fee.
The security difference, which is the real one
Your own key lives in the macOS keychain and is transmitted only to the vendor that issued it. It does not reach our servers. If we are compromised, your provider key is not part of the incident.
Hosted credits mean our keys, on our infrastructure, and your prompts passing through them. That is a larger trust surface and no amount of policy language shrinks it.
There is a trade in the other direction worth stating. With your own key, the blast radius of a mistake is your provider account. A misconfigured agent burning tokens bills you directly at the provider, and the provider's own limits are the last line rather than ours. The per-run and per-day caps still apply — they are a property of the runner, not of the billing relationship — but they are our software protecting your money at somebody else's vendor.
What is deliberately not supported
This is where a bring-your-own-key article usually lists twenty providers. Ours is bounded, and the bounds are decisions rather than backlog:
- No Ollama. No local model endpoints.
- No OpenRouter. No aggregators.
- No arbitrary OpenAI-compatible base URL.
All three are excluded by the security model. An editor that will send your repository's contents to any host a config file names is an editor that can be pointed at a host you did not choose — by a teammate, by a merged branch, by a template. The allowlist is short because the failure mode is exfiltration and it is silent.
Also worth saying plainly: Codoro does not resell or borrow a provider's consumer subscription.
There is no path where your Claude Pro or ChatGPT Plus plan powers the editor. Every provider row in
the catalogue is subscription: false. Products that offer this are doing something their provider's
terms usually do not permit, and it tends to stop working.
The providers that are supported: Anthropic, OpenAI, DeepSeek, Google, xAI and Moonshot with hosted credits; bring-your-own additionally covers Z.ai, Alibaba and Mistral.
When each one is right
Bring your own key if you already have provider accounts, if your employer's procurement has already approved a vendor, if you have negotiated rates, or if the security answer above matters to you. This is most professional developers.
Hosted credits if you do not want five vendor relationships for one tool, if you want a single predictable monthly number, or if you are on a team with a pooled budget and a per-seat ceiling the owner sets. The 40% is the price of that, and it is legible rather than hidden in a seat fee.
Both is allowed and is what I would expect a team to end up doing: keys for the people who have them, credits for the people who do not, one editor.
The thing that makes either choice safe
Whichever you pick, the caps are the same: $2.00 per run by default, editable per task; $25.00 per developer per day, set by the team owner. At the cap a run pauses and offers a cheaper model rather than silently downgrading — a run that changed models mid-task would give you a diff to review under assumptions that stopped being true halfway through.
And the reservation behind those caps is taken at the dearest price the request could attract, then settled at the real one. That asymmetry is deliberate: under-reserving defeats the cap, and settling at the maximum would charge a regional premium the request may never have incurred.
Status
No public download. The macOS build is in preview and the download page is a waitlist until a signed release manifest says otherwise, so everything above is the launch model rather than a checkout you can reach today. The model catalogue is real — roughly 216 models synced from Vercel AI Gateway, with a CI check that fails when committed rates drift from published ones.