Anyone hitting a wall with Claude on this repo?

[5eh]

Considering using a different model service such as OpenModel for this. Is there a way to have Claude be supportive of this rather than safeguarding?

[jimtje]

Everybody hits a wall with Claude on this, at least with their models. It's on the API end since overriding it with an ENV before launching claude code would fix it if the endpoint runs a model that has refusals turned off. Actually, this seems to trigger most models at some point, although at different points for seemingly different reasons with different wording. The only true no-refusal runs I've had are either with a self-hosted one that has refusals removed and a small number that are available remotely, and of course, these are all open source ones.

Since Anthropic sells a security version of their product, they have little reason to stop gatekeeping. "AI safety" etc. Of course, their models still would do some of the work, but they just won't do all of it, even on my own code, my own servers, on apks or ipas signed by my dev cert, which are paid and on my keychain and valid. Wild, innit?

[kjarada]

Same issue here

[5eh]

Kind of figured. Tried running DeepSeek Harness on an AWS Instance but ran into some context issues... Time to get silly with some workarounds. Big wild, innit. Thanks for the help <3

[Muhammadbilal83]

Hey, just to clarify from a general standpoint — Claude’s safety behavior isn’t something that can be toggled off via an ENV var or config flag, even on the API. It’s built into the model itself, not a separate filter layer sitting on top, so “overriding” it isn’t really on the table for Anthropic’s hosted models.

If you’re hitting refusals on stuff that’s actually legitimate (your own signed code, your own servers, etc.) and it feels like a false positive, the best path is to report it directly to Anthropic through their support channel (support.claude.com) or the thumbs-down feedback in the product — that’s the only way it actually gets looked at and potentially tuned.

If what you need is a model with no safety guardrails at all, that’s a fundamentally different product category — self-hosting an open-weight model (like you’re already trying with DeepSeek) is really the only route there, since that’s not something Claude is ever going to offer.

[Sohan-Zakaria]

You won’t be able to prompt around the safety layer directly on Anthropic's hosted API—the guardrails trigger at the endpoint level before output generation.
If you want to keep using the claude CLI workflow without hitting safeguard walls, you can redirect the API base URL to an uncensored OpenAI-compatible model runner:
export ANTHROPIC_BASE_URL="http://localhost:8000/v1"
export ANTHROPIC_API_KEY="dummy"
Point it to an instance running an open-weights model with refusals stripped (e.g. via vLLM, Ollama, or an uncensored DeepSeek/Llama derivative). That way you preserve the tool-calling/agent loop of the CLI while bypassing hosted safety gatekeeping.