FR: Claude Code over Agent SDK

[NightHammer1000]

It is possible to use Claude Code as a Provider with the Agent SDK that makes it possible to use Subscriptions.

You need to limit the turn count to 1 and format the Request in Anthropic Format to Contain both Assistant and User Roles and send that as the Message.

System Message can be set to overwrite Claude Codes internal over the designated field In the SDK.

This gives full access to all the Models available.

I run it like this in my own Proxy I wrote.

Same can be done via spawning the claude code process directly and handing the Data via command line Arguments.

KiloCode Sourcecode is a good place to look on how this is done correctly

[mirrobot-agent]

@NightHammer1000 Thank you for submitting this issue. I am now beginning my analysis and will report back shortly.

[mirrobot-agent]

Initial Analysis Report

Summary: This is a feature request to add Claude Code as a provider using Anthropic's Agent SDK, enabling access to Claude subscription models through the proxy by utilizing the Agent SDK's subscription-based authentication.

Issue Validation: Confirmed — This is a valid and actionable feature request. The author provides concrete implementation guidance (turn count limiting, Anthropic format messages, system message override).

Reproducibility Assessment: Not Applicable — This is a feature request, not a bug report.

Related Issues:

• Anthropic API Endpoint #33 (Anthropic API Endpoint) — Related but different scope. Issue Anthropic API Endpoint #33 requests a generic Anthropic Messages API endpoint for Claude Code usage, while this issue (FR: Claude Code over Agent SDK #38) specifically describes implementing Claude Code integration via the Agent SDK with subscription support. These are complementary but distinct approaches.
• FR : add "grey provider" opencode kilocode #24 (grey provider opencode kilocode) — Marked invalid and closed. The author mentions KiloCode's source code as a reference implementation (the codebase already references KiloCode in src/rotator_library/providers/qwen_auth_base.py:30 for CLIENT_ID).

Technical Analysis:

The codebase already has sophisticated infrastructure for supporting similar providers:

1. Provider Architecture (src/rotator_library/providers/provider_interface.py):

   • Abstract ProviderInterface class with support for custom completion logic (has_custom_logic(), acompletion())
   • OAuth-based authentication patterns (get_auth_header(), proactively_refresh())
   • Credential prioritization and tier management

2. Existing Similar Implementations:

   • Qwen Code Provider (src/rotator_library/providers/qwen_code_provider.py) — Uses custom acompletion() logic with SSE streaming, tool schema cleaning, and response format conversion
   • Antigravity Provider — Already provides Claude Opus 4.5/Sonnet 4.5 access through Google's internal API with thinking mode support

3. Provider Registration (src/rotator_library/provider_factory.py):

   • Simple provider map structure for adding new providers
   • Auth base classes for OAuth credential handling

Implementation Approach (Based on Author's Guidance):

The author suggests two approaches:

1. Agent SDK Method — Use Anthropic's Agent SDK with turn_count=1, Anthropic format messages (Assistant + User roles), and system message override
2. CLI Process Spawning — Spawn Claude Code CLI directly and pass data via command-line arguments

Suggested Labels: enhancement, needs discussion

• Justification: This is a feature enhancement that requires architectural discussion on:
  • Agent SDK vs CLI spawning approach
  • Subscription credential management
  • Integration with existing credential rotation system
  • Whether this complements or overlaps with Anthropic API Endpoint #33

Proposed Next Steps:

1. Clarify Relationship with Anthropic API Endpoint #33 — Determine if this should be merged with or supersede issue Anthropic API Endpoint #33 (Anthropic API Endpoint), as both aim to enable Claude model access but via different mechanisms.

2. Architectural Decision — Choose between:

   • Agent SDK approach (recommended by author): Cleaner integration but requires understanding of SDK authentication
   • CLI spawning approach: More isolated but higher process overhead

3. Implementation Plan (if proceeding with Agent SDK):

   • Create claude_code_provider.py following the pattern of qwen_code_provider.py
   • Implement ClaudeCodeAuthBase for subscription token management
   • Add provider to provider_factory.py
   • Handle Anthropic message format conversion in acompletion()

4. Reference Implementation — Review KiloCode source for implementation patterns (as suggested by author)

Missing Information (if any):

1. Agent SDK Details — Which specific Anthropic SDK package/version is being referred to? Is this anthropic PyPI package or a separate Agent-specific SDK?
2. Authentication Flow — How are subscription credentials obtained and stored? Is this OAuth-based or API key-based?
3. Message Format — Could you provide a sample of the "Anthropic Format" message structure with Assistant + User roles?
4. Available Models — Which specific models become accessible through this subscription method?

---

This analysis was generated by an AI assistant.

[NightHammer1000]

There is also the Possibility to use the OAUTH2 Token directly, like opencode does it. Almost like we do with Gemini.

https://github.com/sst/opencode-anthropic-auth

But there are Reports of Account Bans after using OpenCode with it.
It is unclear if it's really because of this. I have been using it for a while now without issue.