Thermal Platform API
Three supported ways to work with Thermal records from outside the app: the HTTP API, the command line with an agent skill, and an MCP connection for assistants. All three use the same organization permissions, branch scope, rate limits and preview-then-apply controls. The full reference is below and as OpenAPI 3.1. Contract changes are listed in the API changelog.
Supported and unsupported channels
| Channel | Status |
|---|---|
HTTP API (/api/v1) | Supported. Organization API keys and OAuth access tokens. |
Command line (@thermal/cli) and agent skill, from shell-capable hosts such as Claude Code and Codex | Supported at launch. Not yet published; this page is updated when it is. |
| MCP connection, from MCP-capable assistants | Supported once the MCP connection is confirmed in production; assistants that support remote MCP with sign-in are in scope. Hosts are listed here when verified. |
| Assistants without shell or remote MCP support | Unsupported. |
| Scheduled or unattended agent runs, loops, and any tool outside the published operations | Unsupported. Every call is made on behalf of a signed-in person. |
| Personal Thermal sessions or API keys used on the MCP connection | Unsupported. The MCP connection accepts OAuth sign-in only. |
Reconsideration record
An unsupported channel is reconsidered only on evidence. Owner: the Thermal Platform API owner. Evidence threshold: a recorded run in which a real host of that kind completes sign-in, performs a read and a preview-then-apply change through the published operations, is refused outside its organization and permissions, and leaves a matching activity record. Without that record the channel stays unsupported.
Command line and skill setup
Requires Node 24 or later on macOS, Linux or Windows. Support window and upgrade rules are in the CLI support window.
npm install --global @thermal/cli thermal login # opens your browser; refresh material stays in the OS credential store thermal login --key # headless mode with an organization API key
Every command accepts --json and prints one versioned JSON value on stdout. Changes above the risk threshold need preview first, then apply with the proposal, an idempotency key and --reason. The agent skill installs into Claude Code or Codex and follows the same rules; the server enforces them regardless of what the skill says.
MCP service endpoint and sign-in callback
Two separate addresses:
- Service endpoint, where your assistant sends requests:
https://mcp.trythermal.com/v1 - OAuth callback path, reserved for the sign-in return:
https://mcp.trythermal.com/oauth/callback. It never accepts tool calls. Do not enter it as the server address.
Supported-host statement: an assistant is supported when it supports remote MCP over Streamable HTTP with OAuth sign-in. No host is listed as verified yet. Setup steps are in Connect Thermal to your assistant, and permission rules in API access and webhooks.