npm install and go: the DCode CLI and the real cost of the DipoleML API

Some of the best engineers we know have never opened our desktop app. They live in a terminal, six tmux panes deep, and the idea of clicking through a GUI to talk to a model makes them tired. This post is for them. It doubles as the cost breakdown we wish other AI vendors had the nerve to publish.

The CLI, in one command

npm install -g @dipoleml/dcode-cli drops the full DCode agent into your shell. It is not a chat wrapper. Underneath sits the same planner and shell runner as the desktop app; the prompt just replaces the window. One account, one set of Raven usage windows, shared with the desktop app. Start a session over SSH and it picks up wherever the IDE left off.

What it is like to use

Point it at a directory, describe what you want, and step back. It indexes the repo, plans, edits, runs your build and your tests, and reads its own failures well enough to fix them. We have thrown monorepo dependency bumps at it that we did not have the stomach for by hand. Over SSH, on some remote box, while your local editor sleeps. The CI pipeline that fails at 3 am? The agent reads the failure log before anyone's alarm goes off. And it runs from scripts, for those multi-repo chores you refuse to do twice.

The examples that sold us internally were all boring, which is rather the point. Migration scripts. Test triage. Dependency bumps across a monorepo. The kind of work nobody misses doing by hand. And because the tool is just a CLI, it slots into whatever you already run.

The API behind it

Under the hood, everything talks to the DipoleML API at api.dipoleml.com. It speaks both the OpenAI and Anthropic SDK formats natively, so if you have working client code, migrating is a one-line base URL change; the full protocol walkthrough is in our gateway article. If you are building your own thing rather than driving our agent, that is the surface you want. The developer overview lists every endpoint.

The price list

Billing is pay-as-you-go with prepaid credit packs at $5, $10, $20, and $50. The credits never expire, which we consider a moral position as much as a pricing one. Per-token rates, published in full on the pricing page:

• Raven Flash: $0.50 per 1M input tokens, $2.50 per 1M output tokens
• Raven Max: $1.25 per 1M input tokens, $6.25 per 1M output tokens
• Cached input: $0.05 per 1M on Flash, $0.13 per 1M on Max, applied automatically with no flags to set

When your input repeats, and most production input does, the cache quietly cuts input costs by up to 90%. No engineer should ever have to think about this, and with us nobody does.

Management keys

Teams get a second layer that has saved our own late nights: management keys. These provision regular API keys with individual USD spend limits and expiry dates, from the dashboard or from /v1/management/keys directly. Give the intern a key capped at $5 that dies Friday. Give the production service a key with a real budget. Revoke either one in a click when the intern becomes a senior engineer with stronger opinions. API keys cannot manage, and management keys cannot chat. Both rules hold from the first token onward, with no exceptions.

From install to first task

The impatient path is short: install the CLI, sign in, hand it a task in your messiest repo. If you care more about the API, mint a key in the dashboard, aim an existing OpenAI or Anthropic client at our base URL, and rerun a request from yesterday. Both SDK families have copy-ready samples in the docs, and the playground exists for poking the models before you install anything.

There is no sales call and no "contact us for pricing" wall anywhere in the flow. Install it, build something, and if it turns out not to be worth the fifty cents, the credits keep until you need them.