nc://anthropic

A Nodecode cell that lets a Claude Pro or Max subscription serve the anthropic provider. Sign in once with /anthropic login. When no Anthropic key is set, anthropic/<model> rounds then run on the subscription.

The core already serves anthropic with an API key, and that path does not change. A key in the config, a key /connect saved, ANTHROPIC_API_KEY or ANTHROPIC_AUTH_TOKEN all still win over the sign-in. A key round goes out exactly as it did before the cell was installed, with the same address, headers and bytes.

Ported from oh-my-pi's anthropic sign-in and the Claude Code transport its anthropic provider uses for a subscription token. See NOTICE.

Before you use it

• Anthropic's terms. The Claude Agent SDK documentation says third-party products may not offer claude.ai login or plan rate limits without Anthropic's approval. This cell signs in to claude.ai with Claude Code's public client and makes each request look like Claude Code's. It is meant for running your own subscription on your own machine. Read Anthropic's current terms yourself before relying on it, and expect that Anthropic may refuse such requests or act on the account.
• Billing. Whether a round draws from your plan or from extra usage is decided by Anthropic's service. Turn off extra usage in your account if you don't want overage charges. Nodecode prices anthropic rounds at the API list price whatever the credential, so the Dashboard shows what a round would have cost on the API, not what the plan charged.
• Long context. The context-1m beta is never asked for, because a seat has no long-context credit. Models whose window is natively 1M serve it without the beta.
• The grant lasts about 30 days. Anthropic ends a sign-in's refresh tokens about a month after you sign in, however often they are refreshed. Then a round fails with a refresh error naming /anthropic login. Sign in again.

Sign in

/anthropic login

The command answers with an address on claude.ai. Open it in a browser and sign in. claude.ai sends the browser back to http://localhost:54545/callback, or to another free port if 54545 is busy. This cell listens there for five minutes, and a notice says when you are signed in.

If the browser runs on another machine, claude.ai shows the code on its own page. Copy it, along with the #state after it, and send:

/anthropic code CODE#STATE

The address the browser ended on works too.

The token is kept in Nodecode's shared auth.json under oauth_tokens.anthropic:

{"provider": "anthropic", "access_token": "sk-ant-oat01-...",
 "refresh_token": "...", "expires_at": 1791234567, "account_id": "...",
 "email": "...", "org_id": "...", "org_name": "...", "installation_id": "..."}

expires_at is in epoch seconds, saved five minutes early as omp saves it. A token due within a minute of it is refreshed before a round sends it, and the refreshed entry is written back. Every other field of auth.json is kept, and the file stays mode 0600. When the token response does not say whose account it is, the cell asks Claude Code's bootstrap endpoint. /anthropic status says who is signed in and whether a key outranks the sign-in. /anthropic logout takes the entry out.

What a subscription round carries

Anthropic accepts a subscription token only on a request shaped like Claude Code's own. A round on the anthropic lane whose credential is this cell's sign-in, or whose key is a subscription token (sk-ant-oat..., as omp tells them apart), goes out like this:

• Authorization: Bearer instead of x-api-key, to /v1/messages?beta=true
• Claude Code's headers: User-Agent: claude-cli/2.1.280 (external, cli), x-app: cli, the Stainless SDK headers, X-Claude-Code-Session-Id, and Claude Code's beta set (oauth-2025-04-20, and claude-code-20250219 for a request with tools or thinking)
• the system prompt opens with Claude Code's billing header (its version, a fingerprint of the first user message, and a cch attestation: the low 20 bits of the body's XXH64) and its one-line identity, ahead of Nodecode's own prompt
• every tool name carries a _ prefix, which is taken back off the answer's tool calls
• a metadata.user_id names the device, the session and the account
• cache marks live an hour, Claude Code's policy for a seat

The identity block carries no cache mark of its own. In omp it does, but Nodecode already spends Anthropic's four cache marks per request, and the mark on the last system block covers it.

The catalog

The anthropic row gains omp's bundled models that models.dev does not list, such as older Claude 3 models and the Mythos models. Nothing models.dev lists changes, and no base is added, so the core's lane keeps its own address.

Install

Copy this folder into ~/.nodecode/cells/nodecode-anthropic/ and restart Nodecode, or run (restart-cells).

Configure

Nothing to set. The section is on unless it says "enabled": false:

{
  "anthropic": {}
}

Not ported

• omp's Foundry mode (env hook="anthropic-foundry"): enterprise gateway credentials, not a sign-in.
• Adopting a newer Claude Code version from a claude_code_version_too_old refusal. The version is pinned at 2.1.280 here. If Anthropic starts refusing it, the round fails with Anthropic's reason.
• Usage and quota reporting.
• omp's mid-conversation tool and effort controls, compaction and server-side fallback. The core builds the body, and this cell only reshapes it.

Test

From this repository's root:

tools/test-cell.sh nodecode-anthropic

The sign-in's callback listens on a free loopback port the test dials. Every Anthropic endpoint is a stubbed dex:post or dex:get, so nothing reaches Anthropic. One test runs the same key round with and without the cell started and compares the address, headers and body byte for byte.

MIT licensed.