Connectors, Subscriptions, and Token Costs
Understand the open AlonChat connectors, hosted runtime, subscription boundary, and AI usage
Connectors, Subscriptions, and Token Costs#
AlonChat uses an open connector, hosted runtime model. The official contracts, TypeScript SDK, CLI, and MCP server are small clients that connect software and AI tools to AlonChat. The AlonChat application, agent runtime, workflow execution, source processing, billing system, and infrastructure remain part of the hosted AlonChat service.
What Is Open#
These npm packages use the Apache License 2.0:
@alonchat/contracts@alonchat/sdk@alonchat/cli@alonchat/mcp-server
You may inspect, use, modify, and redistribute those packages under the license terms. The license does not grant rights to use AlonChat names or branding in a way that implies an official AlonChat product.
The packages contain connection code and public contracts. They do not contain AlonChat's backend, provider credentials, private prompts, customer data, workflow engine, retrieval system, or billing implementation.
@alonchat/contracts is shared metadata for the other three packages. It is not a separate business
API. Most integrations should install @alonchat/sdk. An operation present in the installed
metadata is not permission to execute it.
What the Subscription Pays For#
Installing a connector package is free. Calling the hosted AlonChat service requires a valid project API key and is governed by the project's plan, permissions, limits, and credit balance.
The subscription and credits cover hosted work such as:
- Running AlonChat agent turns
- Retrieval and embedding work
- Source ingestion and processing
- Tool, Procedure, and Workflow execution
- Jobs, storage, logs, and reliability infrastructure
The clients send typed requests for the published operation catalog. AlonChat authorizes and executes them. New agent, source, Procedure, and Workflow capabilities appear on these connectors only after they are part of that catalog. The implementations stay in the hosted service.
The agent workbench now exposes bounded setup and conversation reads plus review-gated design sessions. An outside model can propose linked Q&A, Procedure, Workflow/form, alert, and Tool changes without receiving AlonChat's private runtime source or unattended publish authority.
Who Pays for AI Tokens#
There can be two independent AI calls:
| Usage | Who runs it | Who is charged |
|---|---|---|
| An outside agent decides to call AlonChat MCP | Codex, Claude, or another client provider | The user of that outside provider |
chat_with_agent runs the AlonChat agent | AlonChat's hosted model provider account | AlonChat, recovered through the project's AlonChat credits |
The MCP adapter itself does not create a second hidden subscription. get_subscription,
get_usage, list_plans, and list_invoices read AlonChat data and do not run an AlonChat
language model. chat_with_agent does, and it consumes the project's AlonChat credits. Do not
assume every other operation is free of credit use. Read billing usage when you need the project
balance.
Billing Through CLI and MCP#
The CLI exposes:
alonchat billing status --json
alonchat billing usage --json
alonchat billing plans --json
alonchat billing invoices --input '{"limit":20}' --json
The SDK methods are getBillingSubscription(), getBillingUsage(), listBillingPlans(), and
listBillingInvoices(). MCP exposes the matching get_subscription, get_usage, list_plans, and
list_invoices tools when the connected credential allows them.
All four reads require billing_read, so the key permission means what it says: a key without it
cannot see the plan, renewal date, trial state, usage, or invoices. A chat-only integration can
still tell it is running low, because every chat response returns credits_remaining.
Card details, provider secrets, direct cancellation, and direct plan changes are not exposed to unattended agents. Consequential billing actions continue in the authenticated owner dashboard.
Recommended Key Setup#
For an agent connection that also monitors spend:
In Project Settings → API Keys, enable chat, read_agents, and billing_read. Leave
read_sources, conversation reads, agent-design writes, and activity-case writes off unless that
integration needs them. File and spreadsheet uploads additionally need the knowledge.write
automation scope. They are not included in read_sources.
Use billing_read off when the integration only needs chat. A chat response still includes
credits_remaining. Restrict the key to selected agents whenever the integration does not need the
entire project.