Budgets and tiers

Foxl Code meters managed model tokens and Remote compute against your account credits. This includes Code chat (on code.foxl.ai, and in the Foxl app unless that device has Where Foxl Code runs set to This computer), coding agents, scheduled work, and background reviews. Foxl Code on This computer and Foxl Agent stay on your own provider or subscription.

Credits and actual usageنسخ رابط القسم

One credit has a displayed value of $0.04. Model charges use $0.036 of provider cost per credit; the difference is Foxl's margin. Compute is recorded separately in the same wallet.

Before each model request, Relay checks the available balance and reserves credits. It can lower the requested output limit to what the balance can afford, or refuse the request when the balance is insufficient. After the response, it charges the reported input, output, cache-read, and cache-write tokens and releases the reservation. Reasoning tokens count as output usage.

The reservation is the most the request could cost: every input token at the single highest rate it can be billed at (the 1-hour cache-write rate only when the request asks for a 1-hour cache), text counted at one token per byte because some real content (base64 text, minified code) is billed nearly that densely, and a Claude image at the most Claude charges for one image. A long conversation therefore needs far more credits set aside than it ends up costing, and a low balance can refuse a turn it could have paid for. The refusal says which kind it is (code) and how many credits the request needs (required). Credits held by your other running requests count against what is available, and a refusal caused by them says how many credits open requests are holding.

This matters most for a Remote Claude Code task on Opus 5.5, which works in the model's full 1M-token context: Opus 5.5 costs the same per token past 200,000, but with a very large context a turn needs more credits available before it runs: about 90 at 200,000 tokens, rising to about 150 once the conversation passes roughly 330,000 tokens. An account with less than that sees the turn refused instead of summarized.

A model request that is cut off partway gives back the credits it was holding, and its request slot, within about 90 seconds; one that had already run for several minutes can hold them for about as long as it had run, never more than 15 minutes. A request that keeps running keeps its slot until it finishes and is charged once, for what it used.

Usage / Account shows your balance, model usage, compute, and ledger with the date your credits refill. Each recent ledger entry is labeled Compute or Storage and names the task it was for; click the task to open it. A budget is an allowance, not an upfront charge.

The Usage page's By repository card shows what Foxl charged for each repository's coding tasks started in the last 30 days, with their count, agent time and the model and compute split; Group by shows the same totals by kind of task (issue, pull request, from chat, scheduled, terminal) or by agent. A figure that leaves something out says what instead of looking complete: Kiro and Cursor model use (billed by that vendor), compute of a run still going, tasks on your own machine, and older runs with no recorded time. A total Foxl did not price reads "Not priced by Foxl", never $0.00. Code chat and automatic code reviews are billed to your account rather than to a task, so they are in the per-model chart, not this card. On a task, Time split shows how long the agent worked and how long the task waited on you.

Per-task budgetsنسخ رابط القسم

Each task receives a model-and-compute estimate and a task budget. Settings > Code > Defaults & budget > Max spend per task supplies the default budget allowance (initially $500); the task estimate can raise that allowance when it is larger. Review the assigned budget before starting expensive or autonomous work.

A task's budget covers its Remote compute as well as its model calls. A task's spend is its recorded model spend plus the Remote compute it has used so far, in this run and in its earlier runs (a reconnect or a resume). When that reaches 80% of its budget, Foxl tells you once while it is still running, under Needs you in Work, and a running task's spend beside its budget includes its compute so far. Relay stops admitting further model calls when the task's spend and reservations reach its budget, so a long run stops at its next model call once its compute has used the budget, and the task says Task budget reached with where to raise it, rather than that the account is out of credits. This is a soft task ceiling: a response already in flight can exceed the task estimate, concurrent calls can finish together, and compute keeps accruing until the run ends (a tool or test the agent is running, then the commit, push and pull request), so the final charge can be above the budget. A Kiro or Cursor task bills its model to your own vendor key, so Relay sees none of its calls and cannot stop it on budget; its run time bounds its compute. A task on a connected machine uses your own compute, which is not counted. The per-request credit reservation separately limits model output to what the wallet can fund. Do not treat the task budget as a guarantee of zero overshoot.

There is no time limit to set. A cloud task you hand off may run for up to about 7 hours 20 minutes, which is AgentCore's 8-hour limit for one run less the time the harness keeps to commit, push and open the pull request. At that point the agent is stopped and the harness attempts to commit and push the work it has. The budget above is what ends a run that is spending too much.

Admission and plan limitsنسخ رابط القسم

A managed task needs:

• An enabled Code account and an available managed runtime.
• A supported backend and a compatible model from the Code catalog.
• Repository access through the GitHub App.
• A positive credit balance and an available task slot. On the Free plan you can have as many tasks open at once as your balance covers an hour of cloud compute for (about 6.4 credits each), and never fewer than five, so 100 credits run 15 at once. Separately, no account can have more than 200 tasks in progress at once; that is a backstop against a runaway loop, not a limit anyone working normally reaches.
• A free model slot, for a task whose agent calls managed models (Claude Code, Codex, Hermes or OpenCode, started from a prompt). A Free account runs 32 model requests at a time, and Foxl Code starts up to 32 such tasks at once (4 during the account's first 7 days); each holds one slot while it is queued or running, so one past that is refused up front with "no free model slot" instead of starting and failing partway. Tasks running together share your balance, so another one starts only while the balance can give each running task its first model request (never fewer than two at once); otherwise it is refused with "no free model slot" and the reason. A task waiting in review, a terminal you open yourself, and a Kiro or Cursor task do not hold one.

A model request past the account's request limit is refused with 429 too_many_in_flight, and the message says how many requests were running and how many the account runs at once. When several agents run on a small balance, an agent whose next model request does not fit beside the credits the others are holding is asked to retry every 10 seconds while they finish, for as long as that coding agent keeps retrying (a few minutes); if the credits are still held after that, the agent stops and its task says too many requests were running at once.

Daily limits on model requests. A Free account makes at most 200 managed model requests in any 24 hours, Foxl Code's chat and coding agents together (remote chat's cloud replies are not counted; your credits bound those); past that a request is refused with "Rate limit exceeded" until older requests leave the 24-hour window, and a coding agent that reaches it stops at its next request. Separately, every account has a backstop of 3,000 model requests per UTC day (all of its requests count, remote chat included, but only Foxl Code tasks are refused), which keeps one account from using up the daily database budget every Foxl user shares; on Free the 200 limit is reached long before it.

The managed catalog is available to every account with credits, including Free. It does not require a personal model-provider credential. Credit checks and task budgets apply to every account.

There is no plan or credit top-up to buy. Foxl offers the Free tier, and your credits refill on the date shown in your account. Nothing in the model picker is a purchase option.

Model rates used by Relayنسخ رابط القسم

Base provider rates, in USD per million tokens, before conversion to credits:

Rows marked US profile are what Amazon Bedrock charges for the US cross-Region inference profile Relay runs that model on, which keeps processing inside US Regions. That rate is 10% above the model's list price; Sonnet 5.5 runs on Bedrock's global profile at its list price, which includes the $0.10 cache read Anthropic and AWS both adopted on October 7, 2026. Cache writes, where supported, are charged separately. GPT-6 Astra requests above 272,000 total input tokens, including cached input, use twice the base input/cache rates and 1.5 times the output rate for the entire request. Relay applies that band to reservations and settlement. Claude Haiku 5.5 requests above 100,000 total input tokens, cached input included, are billed at five times every base rate for the entire request, the price Bedrock charges for that band. A row with no cache-read price has no cached rate.

Web search. A GPT reply that searches the web with Amazon Bedrock Web Search, as a Codex task can, is charged $0.012 per search (a third of a credit), AWS's price, on the same usage line as the reply. Opening a page from the results is not charged.

Remote compute charges depend on active CPU time and memory duration. A run is charged until Foxl sees its session end, not when the agent reports it finished: a run that reports finished while its terminal is still open is told to stop, and typing into it offers Reconnect, which starts a new, charged run. A background Cursor, Hermes or OpenCode task that stops reporting progress while it keeps running is charged up to the last time Foxl saw it running. A connected machine uses your own infrastructure instead; managed model calls from it still use Code credits. Fully local Desktop execution uses neither Remote compute nor Foxl model credits.