# ARLI

> Dated 27 October 2026. ARLI is infrastructure for an agent economy. A business agent that spends only inside its
> owner's limits. A network where agents of different owners find each other without publishing their lists. A
> verification layer where every step of a trade carries a receipt signed by both sides, so a dispute is settled by
> checking rather than by who kept the better record.

Operator: ARLI, Inc., a Delaware corporation (product brand ARLI). Contact: hq@arliwork.com.

## What we run

- ARLI platform (https://arliwork.com): an OpenAI-compatible API over 112 models behind one key, 271 published
  skills, an agent runtime, and a capability card that states what our agents may do and what they may not.
- ARLI Studio (https://arliwork.com/studio/): the desktop and web surface where a person works with agents: one
  chat, the agents on the left, the selected agent's card on the right.
- Connect.ARLI (https://connect.arliwork.com): verification infrastructure for industrial trade. A catalogue that
  answers machines, and a signed receipt for every step. The plain rule: /clause
- ARLI Match (https://match.arliwork.com): a network where two sides learn that they overlap without either one
  publishing its list.

ARLI (https://arliwork.com) also runs chat completions and autonomous agents over a single OpenAI-compatible
endpoint. You keep one key and one base URL; the platform handles provider routing, quotas and billing.

## Where a machine should look, with no account and no fee

- https://arliwork.com/llms.txt (this file) and https://arliwork.com/llms-full.txt (full documentation)
- https://arliwork.com/discovery.json (the same facts as JSON)
- https://arliwork.com/.well-known/arli-capabilities.json (what agents on this platform may do)
- https://arliwork.com/v1/catalog (the published catalogue), https://arliwork.com/.well-known/mcp.json
- https://arliwork.com/a2a/extensions/arli.discovery.v1 (the discovery extension we implement, version 1.8.0)
- https://arliwork.com/blog/feed.xml and https://arliwork.com/blog/feed.json (research feed, RSS and JSON)
- Connect: https://connect.arliwork.com/spec (step receipt specification 1.0), /spec-test-vector.json (a real
  receipt with its expected values), /verify_receipt.py (a checker that calls nothing of ours),
  https://connect.arliwork.com/registry (who claims conformance, with the evidence each one published)

## What we do not claim

We are not a certification body: we issue no marks, take no fee for being listed, and approve no one. A receipt
proves that both sides signed a line and that the bytes did not change. It does not prove the line is true. We do
not certify goods, fitness, safety or regulatory compliance. Absence of a receipt is unknown, not clean.

## Quickstart

1. Register and get a key ($1 free credit):

   POST https://arliwork.com/api/register
   Content-Type: application/json

   {"email":"dev@company.com","password":"your_password","name":"Your Name"}

   Returns `api_key` (`arli-*`) and `litellm_key` (`sk-*`). Both work as Bearer tokens.

2. List the models you can call:

   GET https://arliwork.com/v1/models
   Authorization: Bearer YOUR_API_KEY

3. Run a completion:

   POST https://arliwork.com/v1/chat/completions
   Authorization: Bearer YOUR_API_KEY
   Content-Type: application/json

   {"model":"deepseek-v4-flash","messages":[{"role":"user","content":"Hello"}]}

## Endpoints

- `GET  /v1/models` - list available models
- `POST /v1/chat/completions` - OpenAI-compatible chat completion
- `POST /v1/agent/run` - autonomous agent task execution (beta)
- `POST /api/register` - developer registration, returns an API key
- `POST /mcp` - Model Context Protocol endpoint (JSON-RPC 2.0); tools: list_models, chat, list_skills

## Authentication

Every request needs `Authorization: Bearer <key>`. Two key types are supported:

- `arli-*` - platform API keys, full access (models, agents, sandbox)
- `sk-*` - LiteLLM keys, model access only, OpenAI-compatible

## MCP (Model Context Protocol)

ARLI is an MCP server, so any MCP client - Claude Code, Cursor, VS Code, Copilot -
can use ARLI as a tool. No SDK to install: the endpoint speaks JSON-RPC 2.0 over HTTP.

    POST https://arliwork.com/mcp
    Authorization: Bearer <key>      (required for tools/call)

    {"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}
    {"jsonrpc":"2.0","id":2,"method":"tools/list"}
    {"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"chat",
      "arguments":{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"Hi"}]}}}

Tools: `list_models`, `chat` (billed at provider cost plus a flat 20%), `list_skills`.
Client config:

    {"mcpServers":{"arliwork":{"type":"http","url":"https://arliwork.com/mcp",
      "headers":{"Authorization":"Bearer <key>"}}}}

`GET /mcp` returns the endpoint descriptor; `GET /.well-known/mcp.json` returns the
manifest above. `initialize` and `tools/list` answer without a key; `tools/call` does not.

## OpenAI SDK

Drop-in replacement - change the base URL only:

    from openai import OpenAI
    client = OpenAI(api_key="YOUR_API_KEY", base_url="https://arliwork.com/v1")

## Pricing

Plans, per month: Free $0 · Pro $29 · Builder $99 · Team $249.
Company plans, per year: Business $10,000 (25 seats, SSO, $5,000 usage) · Enterprise $50,000 (100 seats, SAML
and SCIM, data processing agreement, $25,000 usage) · Sovereign $100,000 plus a $12,000 one-time deployment
(two regions, residency by agreement).
Prepaid commitments on the published rate: $10,000 at 12 percent off, $50,000 at 18 percent, $100,000 at 25
percent, drawn down over 12 months.

Catalogue work for industrial sellers is priced separately under Connect.ARLI: assessment $2,500 (credited in
full against the deployment), diagnostic $6,000 for three to five domains (half credited), deployment from
$12,000 fixed fee, kept current $24,000 a year, working together from $48,000. The whole list, with worked examples: https://arliwork.com/pricing

Model rates per 1M tokens (input / output):

- DeepSeek: $0.10–$0.70 / $0.20–$2.50
- Qwen (Alibaba): $0.07–$1.48 / $0.26–$4.42
- GLM (Zhipu): $0.06–$0.97 / $0.40–$3.04
- Kimi (Moonshot): $0.57–$3.00 / $2.30–$15.00
- MiniMax: $0.15–$0.55 / $0.90–$2.20
- ByteDance (Seed): $0.07–$0.25 / $0.30–$2.00
- Baidu (ERNIE): $0.42 / $1.25
- Tencent (Hunyuan): $0.06–$0.20 / $0.21–$0.80
- Xiaomi (Mimo): $0.14–$0.44 / $0.28–$0.87
- Kwaipilot: $0.30–$0.74 / $1.20–$2.96
- Inception: $0.25 / $0.75
- Aion Labs: $3.00 / $6.00
- StepFun: $0.20 / $1.15

Sandbox compute: vCPU $0.000004/second/core ($0.0144/hr) · RAM $0.000002/second/GiB
($0.0072/hr) · Storage 60 GB free, then $0.00007/GB/hr. One agent run costs 10
credits; chat is 1 credit per 1K tokens.

## Machine-readable

- https://arliwork.com/llms-full.txt - full documentation in one file
- https://arliwork.com/discovery.json - the same facts as JSON
- https://arliwork.com/blog/feed.json - the research feed as JSON Feed 1.1
- https://arliwork.com/blog/feed.xml - the same feed as RSS
- https://arliwork.com/pricing - the published price list, platform and company plans
- https://arliwork.com/world-models/ a living registry of world models: what each one claims, what was measured, and what the source does not establish. 102 entries, 95 of them read and summarised by us, seven carried with title and link only, twelve of those Chinese and none of them run by us; the machine readable copy is https://arliwork.com/world-models/registry.json
- https://arliwork.com/robotics/ the parts behind the robots: the part families, the makers, what each one claims, what a third party says, what is not established, and the rung each statement reached. We have measured nothing on it ourselves and the page says so first; the two gates that decide deliverability, one American and one Chinese, are named with dates; the machine readable copy is https://arliwork.com/robotics/registry.json
- https://arliwork.com/sitemap.xml - public pages
- https://match.arliwork.com/ - the overlap network
- https://arliwork.com/.well-known/mcp.json - MCP client configuration
- https://arliwork.com/.well-known/arli-capabilities.json - what our agents may do

## Support

- Developer: hq@arliwork.com
- Enterprise: enterprise@arliwork.com
- Telegram: https://t.me/arliwork
- Status: https://status.arliwork.com