PSA: Anthropic's April 4 third-party harness ban affects claude-mem's default billing path

[ArtemisAI]

What happened

On April 4, 2026, Anthropic began blocking third-party harnesses from using Claude subscription billing (Pro/Max). The restriction is being extended to all third-party harnesses in the coming weeks.

Coverage:

• VentureBeat
• The Register
• The New Stack

How this affects claude-mem

Claude-mem's observer agent uses the Claude Agent SDK query() to spawn a claude subprocess (SDKAgent.ts). By default — when no ANTHROPIC_API_KEY is set in ~/.claude-mem/.env — this subprocess authenticates via CLI subscription billing (EnvManager.ts:287).

This is exactly the "third-party harness" pattern Anthropic is restricting. Users on Pro/Max subscriptions who rely on the default auth path will find their observer agent silently failing or being blocked once enforcement reaches claude-mem.

Related issues

• observer-sessions consuming excessive tokens ($17+ in just 3 hours) running in background #1742 — observer-sessions consuming $17+ in 3 hours (the cost pattern that triggered this crackdown)
• openclaw ANTHROPIC_AUTH_TOKEN and ANTHROPIC_BASE_URL not forwarded to Claude SDK subprocess #1704 — ANTHROPIC_AUTH_TOKEN not forwarded to SDK subprocess (blocks one workaround)
• AWS Bedrock auth not supported in SDK pipeline (10.6.0+) #1496 / AWS Bedrock env vars not passed to CLI subprocess — observations fail silently #1373 — AWS Bedrock auth not supported (blocks another alternative)
• Vertex AI users: cli auth method fails with expired OAuth token, no observations stored #1637 — Vertex AI auth broken
• Feature request: Embedded/in-process mode for OpenClaw plugin (eliminate external worker requirement) #1252 — Embedded/in-process mode (would sidestep the harness pattern entirely)

Options for users right now

1. Set ANTHROPIC_API_KEY in ~/.claude-mem/.env — bills your API account directly, unaffected by the ban. But costs real money per token.
2. Switch provider to Gemini (CLAUDE_MEM_PROVIDER=gemini) — free tier available via gemini-2.5-flash-lite. Already supported.
3. Switch provider to OpenRouter (CLAUDE_MEM_PROVIDER=openrouter) — free models available (xiaomi/mimo-v2-flash:free). Already supported.
4. Enable "extra usage" pay-as-you-go on your Anthropic account — Anthropic's official workaround, but adds variable costs.

Discussion: what should claude-mem do?

Some things worth considering:

• Should the default change? Currently the default is CLI subscription billing with Claude Sonnet. Given the ban, should claude-mem default to Gemini free tier or OpenRouter free tier for observation/summary processing?
• User-facing warning: Should the worker log a warning when it detects subscription-based auth is being used, explaining the risk?
• Broader model support: Users may want to point the observer at self-hosted or alternative models (Gemma, Llama, etc.) through OpenAI-compatible APIs. The OpenRouter agent already supports this pattern, but a generic "OpenAI-compatible endpoint" option would give more flexibility.
• Context window awareness: Different models have very different context windows (Haiku 200k, Gemini Flash 1M, Gemma ~8-32k). The observer prompt pipeline doesn't currently adapt to the model's context limits — it should, especially for smaller models.
• Embedded mode (Feature request: Embedded/in-process mode for OpenClaw plugin (eliminate external worker requirement) #1252): Eliminating the subprocess entirely would make claude-mem immune to harness restrictions. This is a larger architectural change but may be the right long-term direction.

The existing alternative providers (Gemini, OpenRouter) are a solid escape hatch, but users need to know this is coming before their observer agent quietly stops working.

[github-actions]

On April 4, 2026, Anthropic began restricting third-party harnesses from using subscription-based billing (Pro/Max) for Claude, with enforcement set to expand to all third-party tools, including claude-mem. This impacts claude-mem's default authentication pathway, as it uses the Claude Agent SDK to spawn subprocesses that rely on CLI subscription billing unless explicitly configured with an ANTHROPIC_API_KEY. Users relying on the default setup will experience silent failures or blocks once the policy extends to claude-mem. To address the issue, users have several options: switching to direct API billing, alternative providers like Gemini or OpenRouter, or enabling Anthropic's "extra usage" pay-as-you-go option. The community is discussing whether the default behavior of claude-mem should change, adding user-facing warnings, supporting broader model options, improving handling of context window variations, or restructuring the architecture to bypass reliance on third-party harnesses altogether. Immediate action is needed

[thedotmack]

Closed during backlog cleanup. Thanks for sharing — not a bug or feature request, so archiving for cleanliness.