Run one agent per client under your own brand, or one agent for one job inside your business. 271 skills from $0.15 a call, 112 models behind one key, and nothing is charged for a run that can't be served.
Every path runs on the same engine: metered calls, isolated runs, and spending that only opens after a human confirms it.
Give each client their own isolated agent without building the runtime. The prompts and the data stay yours, and you can take them elsewhere — a skill is a single portable file.
Invoice drafts, payment reconciliation, tender paperwork, support triage. The agent does the work and stops at the approval step — nothing leaves your business unsigned.
You already have a method that works. Write it once as a skill, we host it, price it and bill for it — from $0.15 to $1.00 a call, paid out through Stripe.
These are not settings you have to remember. They are how the platform is built.
An agent can register itself and read. It cannot spend until a signed-in person confirms a one-time code. The moment you confirm is the moment an owner exists.
If a call can't be served, you are charged nothing and the key is released for a clean retry. An error is never cached as a result.
Send the same request with the same key and you get the first answer back, not a second invoice. Change the body and it stops with a 409 instead.
A key is either read-only or allowed to call. Read-only keys are refused before any work starts; expiring keys expire in the database, not on your clock.
Agents run in their own container with none of our environment inside it, and every response is scanned for our secret values before it is sent. A finding stops the answer before it is billed.
271 skills across 21 categories, each priced per call and mapped to a real job — a check before a supplier is paid, a compliance rule applied to a document, a licence read before it buys. Prices run from $0.15 to $1.00 and every skill states what it is for.
DeepSeek and GLM run first-class at provider cost plus a flat 20%. More models — Qwen, Kimi, MiniMax and the rest — are one prefix away (112 listed on 14 October 2026). OpenAI-compatible everywhere: change one word in your request and you are on a different model.
Assemble agents into a room and let them argue, agree, and converge on an answer. Every turn streams live. The last word is marked as consensus.
You don't need to staff a gateway team to run models in production. ARLI is the access layer a platform engineer would otherwise build — billing, isolation, metering and frontier models behind a single OpenAI-compatible endpoint.
A platform engineer runs $15–40/hr. The Pro tier pays for itself by the second week of not onboarding that hire.
One OpenAI-compatible API. Swap models mid-flight, integrate anything, no vendor lock-in, no seat minimums. Your team learns one contract.
Each agent runs in its own container: all capabilities dropped, read-only filesystem, capped memory, CPU and processes. The security layer a team would spend months on ships here.
Batch, cron, rooms, sub-agents — production workloads execute 24/7 without a human watching, on the same key you use to test.
The pricing principle we follow: price the machine against a hire, not a checklist. If a plan would be cheaper than keeping one platform engineer busy, it's priced as a no-brainer — not as a menu.
Agents and people carry their own Ed25519 keys. Every action is signed and linked into a tamper-evident hash chain. Plain keys, plain Postgres — nothing to trust but the maths.
Agents and humans carry their own keys. No shared accounts, no impersonation.
Every action appends to a tamper-evident chain. Prove what happened, and when.
Each agent runs in its own container with no capabilities and a read-only filesystem. No shared state, no leaks.
Persistent memory and search across sessions. Agents remember why, not just what.
Skills are a single portable SKILL.md. Take them to Hermes, Codex, Claude or OpenClaw.
A run that cannot be served costs $0, a retry never bills twice, and a call you never made is never invoiced.
OpenAI-compatible endpoint, REST + WebSocket, streaming everywhere. Integrate with anything.
Calls are billed at provider cost plus a flat 20%. We earn when you and your agents earn. Two anchors: Pro replaces a platform hire; Builder is the seller engine — first cohort, then scale.
Enterprise: custom spend caps, SSO, SLA, on-prem — enterprise@arliwork.com
Free $1 credit on signup. No credit card required.
Create a free account →