Anthropic TOS Compliance
Forge & Ralph Automation

1. How Forge Uses Claude — Technical Specifics

The Binary

All Claude calls made by Ralph (the Forge automation engine) go through the official claude binary installed via Anthropic's published npm package (@anthropic-ai/claude-code). No browser automation. No scraped web interface. No reverse-engineered APIs.

The Command Pattern

Ralph invokes Claude using the documented non-interactive (headless) mode:

• claude -p "$prompt" --output-format text — standard pipe mode
• --dangerously-skip-permissions — Anthropic's own named flag for automated/unattended execution, documented in Claude Code's help output

Both of these are documented features of the Claude Code CLI. Anthropic named --dangerously-skip-permissions, documented its purpose, and ships it in every release. Using a flag Anthropic deliberately built and named is not a terms violation.

The Model Routing Cascade

Ralph uses a 4-tier routing system designed to minimize API cost and respect Claude Max's intended primary role:

Claude Max is T0 — the primary path. This architecture defers to Claude Max and only falls back to paid API when the subscription tier is rate-limited. It does not circumvent rate limits — it respects them by routing elsewhere.

Rate Limit Handling

When Claude Max rate limits are hit, Ralph's claude-max.sh library:

• Detects the rate limit signal from the CLI
• Sets a 15-minute pause semaphore
• Sends a Telegram alert to the account holder
• Routes subsequent calls to T1/T2 until the pause expires

This is graceful degradation — the exact behavior Anthropic's rate limiting infrastructure is designed to handle. It is not circumvention; it is compliance with the throttling signal.

2. The Compliance Arguments

Claude Code Is Anthropic's Own Product

Claude Code is not a third-party tool or a hack built on top of claude.ai. It is an Anthropic product, maintained by Anthropic engineers, distributed through official Anthropic channels (npm, GitHub), and documented at docs.anthropic.com/claude-code. Using it is not "exploiting" anything — it is using the product as sold.

Claude Max Explicitly Includes Claude Code

Anthropic's pricing page bundles Claude.ai and Claude Code access under the same Max subscription tier. They are not two separate products with separate entitlements — Anthropic has deliberately packaged them together. The Max subscription is sold partly because it includes Claude Code for development and automation use.

The -p Flag Is Documented for Exactly This Use

From Anthropic's Claude Code documentation, the -p (or --print) flag is described as enabling "non-interactive" and "piped" use — output to stdout, no TTY required. The docs explicitly show examples of piping Claude output to other commands and using it in scripts. Ralph's usage is textbook -p mode.

Single Account, Single User, Single Subscription

The TOS concerns around automated use typically target:

• Sharing credentials across multiple users
• Reselling access or sub-licensing
• Scraping responses for competitive model training
• Multi-tenant services where one subscription serves many end users

None of these apply here. One Claude Max account. One human subscriber (Jason MacDonald). All automation runs tasks that benefit only that subscriber. No third-party users are being served by this subscription.

Budget and Rate Governance Are Active

Forge has an explicit cost governance layer that prevents runaway spending:

• Daily API budget cap configured in config/budget.yaml
• Per-task cost cap enforced before each Ralph execution
• Anthropic API calls require explicit approval (not auto-approved)
• Usage logged to logs/claude-max-usage.jsonl with full audit trail

This is the behavior of a responsible subscriber, not an actor trying to maximize extraction from a subscription.

3. The Contrast — What Would Actually Violate TOS

4. Key Distinctions from "Gray Area" Automation Tools

Some third-party tools (e.g., OpenClaw, various browser-based automation scripts) access Claude by:

1. Automating a Chromium/browser session against claude.ai
2. Intercepting and replaying the internal WebSocket or REST calls that the claude.ai web app makes
3. Injecting cookies from a real browser session into automated HTTP requests

These approaches are problematic because they: (a) access an interface designed for humans, not automation; (b) rely on reverse-engineered internal APIs that Anthropic never documented or intended to be stable; (c) impersonate a human browser to bypass controls that are specifically designed to detect automated access.

Ralph does none of this. The claude CLI communicates with Anthropic's servers via officially documented channels. The CLI is Anthropic software. The authentication is Anthropic's own login flow. There is no reverse engineering involved at any layer.

5. Relevant Anthropic Statements & Documentation

Claude Code's Own Documentation

From Anthropic's Claude Code docs, on the -p flag: enables "non-interactive mode" for use in "scripts, CI/CD pipelines, and automated workflows." This is verbatim the use case Ralph is built for.

Claude Code in CI/CD

Anthropic explicitly documents using Claude Code in GitHub Actions, CI/CD pipelines, and automated testing workflows. These are, by definition, non-human, automated, high-frequency invocations. Ralph's pattern is equivalent.

The Max Plan Intent

Claude Max is marketed as the plan for "power users, developers, and teams who use Claude heavily." "Heavily" includes automation. The plan's pricing reflects high-volume use. A subscriber using it heavily, even via automation, is using the product as sold to them.

6. Account & Subscription Summary