TL;DR
A public GitHub Gist publishes a TypeScript test that claims Anthropic's Claude OAuth flow rejects the system prompt phrase "You are OpenCode" while accepting other identity statements. The script runs OAuth against Claude Pro/Max tokens and reports blocked versus allowed second-system-block values; the claim is based on those automated tests and not independently verified.
What happened
A developer posted a Gist containing a Bun/TypeScript script that exercises Anthropic's Claude OAuth-based messaging endpoint to validate which system-prompt phrases are accepted. The script authenticates via OAuth (using a provided client ID) and sends requests with a required first system block string and different second-block identity lines. According to the script's test suite, the exact phrase "You are OpenCode" (and the lowercase variant) consistently fail, while other identity lines such as "You are Cursor.", "You are Pi.", and "You are Droid." succeed. The author prints pass/fail results and concludes the behavior represents a targeted block against OpenCode. The Gist includes the API endpoint, model identifier used in requests, and instructions for reproducing the tests. The report in the Gist is the sole source for these findings and has not been corroborated by other evidence in the provided material.
Why it matters
- If accurate, selective blocking of a competitor's name in system prompts raises questions about platform moderation and competitive behavior.
- Developers building tooling that relies on system prompt identity may encounter unexplained failures or interoperability problems.
- The finding highlights transparency concerns for AI platform policies and how those policies are enforced in programmatic interfaces.
- Open-source projects that integrate with Claude's OAuth flow could be affected if targeted filters exist.
Key facts
- The evidence is a public GitHub Gist containing a Bun-compatible TypeScript script designed to test Anthropic's OAuth/messaging API.
- The script requires a Claude Pro/Max OAuth account to run and performs an OAuth code exchange to obtain an access token.
- A required first system block string in the tests is: "You are Claude Code, Anthropic's official CLI for Claude."
- The script reports that second-system-block phrases "You are OpenCode." and "You are opencode." fail, while "You are Cursor.", "You are Pi.", and "You are Droid." pass.
- Requests in the script target https://api.anthropic.com/v1/messages?beta=true and specify a model labeled "claude-sonnet-4-20250514" and several beta headers.
- The Gist prints a pass/fail summary and an explicit conclusion asserting that Anthropic blocks the OpenCode phrase.
- All details and the conclusion derive from the single Gist; independent verification or responses from Anthropic are not included in the source.
- The Gist includes step-by-step instructions for reproducing the tests locally using Bun and the provided OAuth client ID.
What to watch next
- Formal response or clarification from Anthropic regarding prompt filtering and the specific blocking behavior (not confirmed in the source).
- Independent reproduction of the tests by other developers or researchers to confirm whether the behavior is consistent (not confirmed in the source).
- Any policy documents or changelogs from Anthropic that describe filters applied to OAuth system prompts (not confirmed in the source).
Quick glossary
- System prompt: A message block provided to a language model that sets behavior or identity prior to user instructions.
- OAuth: An authorization protocol that enables applications to obtain limited access to user accounts on an HTTP service.
- API endpoint: A URL exposed by a service where requests are sent to interact with the service's features or data.
- Beta flag: A header or parameter that enables experimental or pre-release features in an API.
Reader FAQ
Does the Gist prove Anthropic intentionally blocked OpenCode?
The Gist's author asserts targeted blocking based on automated tests, but that claim is not independently verified in the source.
How were the tests performed?
The script runs an OAuth flow to obtain a Claude access token and sends messages with two system blocks; test outcomes are recorded as pass/fail in the Gist.
Is there confirmation from Anthropic in the source?
Not confirmed in the source.
Can other identity lines be used in system prompts?
According to the script, examples like "You are Cursor.", "You are Pi.", and "You are Droid." succeeded in the same tests.
Instantly share code, notes, and snippets. R44VC0RP/anthropic-screws-opencode.ts Last active Star 0 (0) Fork 0 (0) Code Revisions 2 Embed Clone this repository at <script src="https://gist.github.com/R44VC0RP/bd391f6a23185c0fed6c6b5fb2bac50e.js"></script> Download ZIP Raw anthropic-screws-opencode.ts #!/usr/bin/env…
Sources
- Anthropic Explicitly Blocking OpenCode
- Anthropic blocks third-party use of Claude Code …
- Using opencode with Anthropic OAuth violates ToS & …
- Anthropic made a big mistake – The Software Archæologist
Related posts
- 9to5Mac Daily: Jan 14, 2026 — New Details on Apple–Google Deal and More
- Musk denies knowledge of Grok underage sexual images as California launches probe
- Scaling Autonomous Coding: Running Hundreds of Agents for Weeks