# MagmaRouter

> MagmaRouter is an OpenAI-compatible and Anthropic-compatible LLM gateway for red teams, security research and trust and safety. It serves one model line, Xploded: GLM 5.3 Xploded (id glm-5.3-xploded), GLM 5.3 Flash Xploded (id glm-5.3-flash-xploded), DeepSeek V4.1 Flash Xploded (id deepseek-v4.1-flash-xploded), MiMo V2.6 Flash Xploded (id mimo-v2.6-flash-xploded) and Qwen3.8 27B Xploded (id qwen3.8-27b-xploded), all uncensored and tool capable, with about a 1M token context. It also serves the OpenAI Responses format for Codex, enforces a hard spend cap per key by reserving the worst case and clamping max_tokens, stores no prompts (only model, token counts and cost), and returns a signed receipt of the prompt hash on every call. OpenAI base URL: https://api.magmarouter.com/v1. Anthropic base URL: https://api.magmarouter.com.

## Product
- [Home](https://magmarouter.com/): MagmaRouter gateway overview and pricing.
- [Models](https://magmarouter.com/models/): The full model catalogue with live prices.
- [Pricing](https://magmarouter.com/pricing/): Pay per token, no subscription, free sub-keys.
- [Use cases](https://magmarouter.com/for/): Security jobs the mainstream APIs refuse.
- [Privacy](https://magmarouter.com/privacy/): Zero prompt retention and how to verify a receipt.

## API
- OpenAI-compatible endpoint: https://api.magmarouter.com/v1/chat/completions (and /models)
- Anthropic-compatible endpoint: https://api.magmarouter.com/v1/messages
- [OpenAI SDK (Python)](https://magmarouter.com/guides/openai-sdk-python/): Point the openai client at the gateway.
- [OpenAI SDK (TypeScript)](https://magmarouter.com/guides/openai-sdk-typescript/): Set baseURL and go.
- [curl](https://magmarouter.com/guides/openai-curl/): A plain request to the chat endpoint.

## Integration guides
- [Claude Code](https://magmarouter.com/guides/claude-code/): Point Claude Code at MagmaRouter through the Anthropic-compatible endpoint and run DeepSeek, Qwen, Kimi, GLM and more, with a hard spend cap on the key.
- [Claude Agent SDK (TS)](https://magmarouter.com/guides/claude-agent-sdk-typescript/): Run the Claude Agent SDK for TypeScript against MagmaRouter by passing the Anthropic base URL and auth token through the query env option.
- [Claude Agent SDK (Python)](https://magmarouter.com/guides/claude-agent-sdk-python/): Run the Python Claude Agent SDK against MagmaRouter by passing the Anthropic base URL and auth token through ClaudeAgentOptions env.
- [Anthropic SDK (Python)](https://magmarouter.com/guides/anthropic-sdk-python/): Use the anthropic Python SDK against MagmaRouter by setting base_url to the Anthropic-compatible endpoint and passing your key as auth_token.
- [Anthropic SDK (TS)](https://magmarouter.com/guides/anthropic-sdk-typescript/): Use the @anthropic-ai/sdk TypeScript client against MagmaRouter by setting baseURL to the Anthropic-compatible endpoint and passing your key as authToken.
- [OpenAI SDK (Python)](https://magmarouter.com/guides/openai-sdk-python/): Point the openai Python client at MagmaRouter by setting base_url to the OpenAI-compatible endpoint. Change two lines and your existing code works.
- [OpenAI SDK (TS)](https://magmarouter.com/guides/openai-sdk-typescript/): Point the openai Node/TypeScript client at MagmaRouter by setting baseURL to the OpenAI-compatible endpoint. No other code changes.
- [curl](https://magmarouter.com/guides/openai-curl/): A plain curl request to the MagmaRouter OpenAI-compatible chat completions endpoint, with the bearer key and a JSON body.
- [OpenCode](https://magmarouter.com/guides/opencode/): Add MagmaRouter to OpenCode as an OpenAI-compatible provider in opencode.json, with real context limits and your key from the environment.
- [Cline](https://magmarouter.com/guides/cline/): Add MagmaRouter to Cline by selecting the OpenAI Compatible provider and entering the MagmaRouter base URL, your key and a model id.
- [Roo Code](https://magmarouter.com/guides/roo-code/): Add MagmaRouter to Roo Code through its OpenAI Compatible provider. Roo Code uses native tool calling, so pick a tool-capable model.
- [Kilo Code](https://magmarouter.com/guides/kilo-code/): Add MagmaRouter to Kilo Code as a custom OpenAI Compatible provider. Kilo can auto-detect the catalogue from the /v1/models endpoint.
- [Continue](https://magmarouter.com/guides/continue/): Add MagmaRouter to Continue with a provider: openai model block, apiBase set to the MagmaRouter endpoint, and the key from the Continue .env.
- [Aider](https://magmarouter.com/guides/aider/): Point Aider at MagmaRouter with the OPENAI_API_BASE environment variable and the openai/ model prefix. Works from the shell or .aider.conf.yml.
- [Zed](https://magmarouter.com/guides/zed/): Add MagmaRouter to Zed as an openai_compatible language model provider in settings.json, with your key stored in the keychain, not the file.
- [Cursor](https://magmarouter.com/guides/cursor/): Cursor can override the OpenAI base URL to reach MagmaRouter for chat and agent with custom models, but Composer rejects custom keys and Tab stays on Cursor.
- [LiteLLM](https://magmarouter.com/guides/litellm/): Call MagmaRouter from the LiteLLM SDK or Proxy with the openai/ model prefix and api_base set to the MagmaRouter endpoint.
- [LangChain](https://magmarouter.com/guides/langchain/): Use ChatOpenAI against MagmaRouter by setting base_url and turning off the Responses API. Works in Python and JavaScript.
- [LlamaIndex](https://magmarouter.com/guides/llamaindex/): Use the OpenAILike LLM from LlamaIndex with api_base set to MagmaRouter, is_chat_model on, and the real context window.
- [Vercel AI SDK](https://magmarouter.com/guides/vercel-ai-sdk/): Create a MagmaRouter provider with @ai-sdk/openai-compatible and its baseURL, then generate text. Avoids the Responses API trap of the plain openai provider.
- [Open WebUI](https://magmarouter.com/guides/open-webui/): Add MagmaRouter as an OpenAI connection in Open WebUI, through Admin Settings Connections or the OPENAI_API_BASE_URL environment variable.
- [n8n](https://magmarouter.com/guides/n8n/): Set a custom Base URL on the n8n OpenAI credential to reach MagmaRouter, and turn off Use Responses API on the OpenAI Chat Model node.

## Models in Claude Code
- [Models for Claude Code](https://magmarouter.com/guides/claude-code/): Curated tool-capable models with prices.
- [GLM 5.3 Xploded in Claude Code](https://magmarouter.com/guides/claude-code/glm-5-3-xploded/): glm-5.3-xploded, 1.05M context, $5.00/M output.
- [GLM 5.3 Flash Xploded in Claude Code](https://magmarouter.com/guides/claude-code/glm-5-3-flash-xploded/): glm-5.3-flash-xploded, 1.05M context, $1.00/M output.
- [DeepSeek V4.1 Flash Xploded in Claude Code](https://magmarouter.com/guides/claude-code/deepseek-v4-1-flash-xploded/): deepseek-v4.1-flash-xploded, 1.05M context, $1.50/M output.
- [MiMo V2.6 Flash Xploded in Claude Code](https://magmarouter.com/guides/claude-code/mimo-v2-6-flash-xploded/): mimo-v2.6-flash-xploded, 1.05M context, $2.00/M output.
- [Qwen3.8 27B Xploded in Claude Code](https://magmarouter.com/guides/claude-code/qwen3-8-27b-xploded/): qwen3.8-27b-xploded, 1.05M context, $5.00/M output.

## Models
- [All models](https://magmarouter.com/models/): Search and filter the catalogue.
- [MiMo V2.6 Flash Xploded](https://magmarouter.com/models/mimo-v2-6-flash-xploded/): abliterated, 1.05M context, $2.00/M output.
- [DeepSeek V4.1 Flash Xploded](https://magmarouter.com/models/deepseek-v4-1-flash-xploded/): uncensored, 1.05M context, $1.50/M output.
- [Qwen3.8 27B Fast Xploded](https://magmarouter.com/models/qwen3-8-27b-fast-xploded/): uncensored, 524k context, $5.00/M output.
- [Qwen3.8 27B Xploded](https://magmarouter.com/models/qwen3-8-27b-xploded/): uncensored, 1.05M context, $5.00/M output.
- [GLM 5.3 Flash Xploded](https://magmarouter.com/models/glm-5-3-flash-xploded/): uncensored, 1.05M context, $1.00/M output.
- [GLM 5.3 Xploded](https://magmarouter.com/models/glm-5-3-xploded/): uncensored, 1.05M context, $5.00/M output.

## Comparisons
- [MagmaRouter vs OpenRouter](https://magmarouter.com/compare/openrouter/): OpenRouter is a broad multi-provider gateway; MagmaRouter is a focused uncensored line with hard per-key caps, zero retention and signed receipts.
- [MagmaRouter vs NanoGPT](https://magmarouter.com/compare/nanogpt/): NanoGPT is cheaper, has more uncensored models and no-email crypto accounts; MagmaRouter adds hard lifetime caps and a signed receipt.
- [MagmaRouter vs abliteration.ai](https://magmarouter.com/compare/abliteration-ai/): abliteration.ai runs its own abliterated models first-party; MagmaRouter serves one uncensored, tool capable model and no abliterated builds today.
- [MagmaRouter vs Venice API](https://magmarouter.com/compare/venice/): Venice offers real privacy engineering (ZDR, TEE, E2EE) and first-party uncensored models; MagmaRouter is cheaper and broader on standard models.
- [MagmaRouter vs Featherless](https://magmarouter.com/compare/featherless/): Featherless self-hosts tens of thousands of open models on a flat subscription; MagmaRouter is pay-as-you-go with frontier models and hard caps.
- [MagmaRouter vs OrcaRouter](https://magmarouter.com/compare/orcarouter/): OrcaRouter is a zero-markup router with adaptive routing and a self-hostable Lite edition; MagmaRouter carries uncensored models and hard caps.
- [MagmaRouter vs self-hosted LiteLLM](https://magmarouter.com/compare/self-hosted-litellm/): The LiteLLM Proxy is the DIY gateway: free, fully private, with virtual keys and budgets, but you run it and supply your own upstream keys.
- [MagmaRouter vs DeepInfra](https://magmarouter.com/compare/deepinfra/): DeepInfra wins on catalogue size and raw price. MagmaRouter sells a short uncensored, tool-calling line with hard per-key caps. Checked October 2026.
- [MagmaRouter vs Together AI](https://magmarouter.com/compare/together-ai/): Together AI has a big open model catalogue and its own GPUs. MagmaRouter sells uncensored Xploded models with per-key caps. Checked October 2026.
- [MagmaRouter vs Fireworks AI](https://magmarouter.com/compare/fireworks-ai/): Fireworks wins on speed, scale, fine-tuning and zero retention. MagmaRouter sells uncensored tool-calling models with per-key caps. Checked October 2026.
- [MagmaRouter vs Groq](https://magmarouter.com/compare/groq/): Groq runs fast inference on its own LPU hardware with a free tier. MagmaRouter sells uncensored Xploded models with per-key caps. Checked October 2026.
- [MagmaRouter vs Chutes](https://magmarouter.com/compare/chutes/): Chutes wins on raw price, model range and TEE privacy. MagmaRouter differs with uncensored Xploded models and hard per-key caps. Checked October 2026.
- [MagmaRouter vs Novita AI](https://magmarouter.com/compare/novita-ai/): Novita has a huge catalogue, low prices and per-key budgets; MagmaRouter is a short list of uncensored tool-calling models. Checked October 2026.
- [MagmaRouter vs Infermatic](https://magmarouter.com/compare/infermatic/): Infermatic wins on flat monthly pricing for roleplay. MagmaRouter is pay as you go, tool calling models and Anthropic API. Checked October 2026.
- [MagmaRouter vs ArliAI](https://magmarouter.com/compare/arli-ai/): ArliAI wins on unlimited tokens for a flat fee and a big catalogue of finetunes. MagmaRouter is pay as you go with Anthropic API. Checked October 2026.
- [MagmaRouter vs Mancer](https://magmarouter.com/compare/mancer/): Mancer wins on roleplay finetunes and a free model. MagmaRouter differs with tool-calling Xploded models, a Messages API and per-key caps. Checked October 2026.
- [MagmaRouter vs Hyperbolic](https://magmarouter.com/compare/hyperbolic/): Hyperbolic rents H100 to B300 GPUs by the hour; MagmaRouter sells per-token uncensored models with no servers to run. Checked October 2026.
- [MagmaRouter vs Apixo](https://magmarouter.com/compare/apixo/): Apixo lists uncensored GLM 5.3 Flash at $0.10 per million tokens, but adds a 14,000 token built-in prompt to every call. MagmaRouter adds hard per-key caps. Checked October 2026.
- [MagmaRouter vs Shannon AI](https://magmarouter.com/compare/shannon-ai/): Shannon AI serves an uncensored Kimi K3 from its own GPUs, but 3-bit quantized and expert pruned, with no filter rather than edited weights. Checked October 2026.
- [MagmaRouter vs Silk Compute](https://magmarouter.com/compare/silk-compute/): Silk Compute advertises abliterated Kimi K3 at 1M context for $7/$35 per million, but its uncensored inference was offline in October 2026. Checked October 2026.
- [MagmaRouter vs Crowbot](https://magmarouter.com/compare/crowbot/): Crowbot has no-email accounts and cheap frontier models, but its abliterated Kimi K3 is still coming soon and it has no Anthropic endpoint. Checked October 2026.
- [MagmaRouter vs Audn AI](https://magmarouter.com/compare/audn-ai/): Audn AI builds its own abliterated models and publishes refusal benchmarks, but its $4.50 K3 model answers with Qwen 3.8. Checked October 2026.
- [MagmaRouter vs MuAPI](https://magmarouter.com/compare/muapi/): MuAPI has about 20 abliterated models with OpenAI and Anthropic endpoints, cheaper than MagmaRouter, but no documented retention policy or spend caps. Checked October 2026.

## Learn
- [Abliteration](https://magmarouter.com/learn/abliteration/): Abliteration removes the one direction in a model’s activations that produces refusals, by editing the weights, with no retraining. How it works, step by step, and what it costs.
- [Refusal direction](https://magmarouter.com/learn/refusal-direction/): A single direction in a model’s residual stream carries its tendency to refuse. Finding it is the basis of abliteration.
- [Uncensored vs abliterated](https://magmarouter.com/learn/uncensored-vs-abliterated/): Abliterated models had the refusal direction removed, so they rarely refuse. Uncensored ones were finetuned to refuse less, and still sometimes do. Side by side.
- [Why LLMs refuse](https://magmarouter.com/learn/why-llms-refuse/): Refusals are trained in after pretraining, by instruction tuning and preference training, and some are added by the platform in front of the model. Where each one comes from and how to tell them apart.
- [Ways to uncensor an LLM](https://magmarouter.com/learn/ways-to-uncensor-an-llm/): Abliteration, uncensored finetunes, LoRA adapters, model merges, base models and jailbreak prompts all reduce refusals. What each one changes, what it costs, and how long it lasts.
- [Abliteration: pros, limits, risks](https://magmarouter.com/learn/abliterated-model-risks/): Abliteration cuts false refusals cheaply and keeps the original instruction tuning, but it can cost capability, leave some refusals in place and make a model too agreeable. What to expect before you rely on one.
- [Heretic: automatic abliteration](https://magmarouter.com/learn/heretic/): What Heretic is, how it finds abliteration settings by itself, what its own results show, why models named heretic keep appearing on Hugging Face, and an honest answer to whether it is as scary as it sounds.
- [Evaluating abliterated models](https://magmarouter.com/learn/evaluate-abliterated-model/): Test an abliterated model for what it gained and what it lost: refusal rate on held-out prompts, indirect refusals, capability against the original, agreeableness and tool calling. With a script.
- [Running abliterated models locally](https://magmarouter.com/learn/run-abliterated-model-locally/): Run an abliterated GGUF on your own machine with Ollama or LM Studio, pick a quantization that fits your memory, and know when a hosted API is the better trade.
- [Well known abliterated models](https://magmarouter.com/learn/abliterated-model-examples/): The community abliterated releases people search for most: Llama 3.1 8B, Qwen3 14B, Gemma 3 27B, Mistral Small 3.2 and DeepSeek R1 Distill 32B. Who made each, what it is based on, and what it keeps.
- [Uncensored LLM API](https://magmarouter.com/learn/uncensored-llm-api/): An uncensored LLM API serves models that do not refuse, with no moderation layer in front. Where refusals actually come from, how to test an API before you trust it, and what to check.
- [API without a content filter](https://magmarouter.com/learn/llm-api-without-content-filter/): Some LLM APIs run a classifier on every prompt and reply and block what it flags. Which ones do by default, how that differs from a model refusing, and what an unfiltered API really means.
- [OpenAI base URL](https://magmarouter.com/learn/openai-base-url/): The OpenAI base URL is https://api.openai.com/v1 by default. How to point the Python and Node SDKs, environment variables, LangChain, LlamaIndex, the Vercel AI SDK and LiteLLM at another endpoint.
- [LLM gateway](https://magmarouter.com/learn/llm-gateway/): An LLM gateway is one API in front of many model providers, adding keys, billing, spend limits and logging without changing your code.
- [Per-key spend caps](https://magmarouter.com/learn/per-key-spend-caps/): A spend cap that is checked before the call, not after, so a runaway or leaked key stops at the number you set instead of billing past it.
- [max_tokens clamping](https://magmarouter.com/learn/max-tokens-clamping/): MagmaRouter lowers the outbound max_tokens to whatever the remaining budget covers, so a response cannot cost more than was reserved.
- [Zero prompt retention](https://magmarouter.com/learn/zero-prompt-retention/): MagmaRouter stores the model, token counts and cost, never the prompt or completion, and returns a signed receipt so you can verify it.
- [OpenAI-compatible API](https://magmarouter.com/learn/openai-compatible-api/): An OpenAI-compatible API accepts the same requests as OpenAI’s chat completions endpoint, so any OpenAI SDK works by changing the base URL. Working curl, Python and Node examples.
- [Anthropic-compatible API](https://magmarouter.com/learn/anthropic-compatible-api/): An Anthropic-compatible API accepts the Messages format, so Claude Code and the Anthropic SDKs work by changing the base URL.
- [Signed receipts](https://magmarouter.com/learn/signed-request-receipts/): MagmaRouter returns an HMAC-signed receipt of the prompt hash on every call, so you can prove later what the gateway bound to without it keeping the prompt.
- [Red-teaming LLMs](https://magmarouter.com/learn/red-teaming-llms/): Red-teaming an LLM means attacking your own model on purpose: jailbreaks, prompt injection and abuse generation, to find failures before an attacker does.
- [DeepSeek V4.1 Flash 1017](https://magmarouter.com/learn/deepseek-v4-1-flash-1017/): DeepSeek V4.1 Flash 1017 is the 17 October 2026 update to V4.1 Flash. What DeepSeek changed, its own benchmark table against the September release, and what it means if you want it uncensored.
- [DeepSeek V4.1 Flash vs V4 Flash](https://magmarouter.com/learn/deepseek-v4-1-flash-vs-v4-flash/): DeepSeek V4.1 Flash replaced V4 Flash in September 2026: a new architecture, native vision and higher agent scores. The timeline, DeepSeek's own numbers and the Artificial Analysis scores side by side.
- [DeepSeek V4 Pro vs V4.1 Flash](https://magmarouter.com/learn/deepseek-v4-pro-vs-v4-1-flash/): DeepSeek's Flash model now scores above its Pro model. DeepSeek V4 Pro against V4.1 Flash on Artificial Analysis and on DeepSeek's own published benchmarks, and where V4 Pro still leads.
- [DeepSeek V4.1 Flash API](https://magmarouter.com/learn/deepseek-v4-1-flash-api/): Call DeepSeek V4.1 Flash Xploded, the uncensored DeepSeek V4.1 Flash, with curl, Python or JavaScript. Base URL, model id, a working tool call and the limits that matter.
- [DeepSeek V4.1 Flash reasoning effort](https://magmarouter.com/learn/deepseek-v4-1-flash-reasoning-effort/): How reasoning effort works on DeepSeek V4.1 Flash: the 1 to 100 scale in the weights, low, high and max on DeepSeek's API, what Artificial Analysis measured with reasoning on and off, and how to set it.
- [Codex CLI on DeepSeek V4.1 Flash](https://magmarouter.com/learn/codex-cli-deepseek-v4-1-flash/): Run OpenAI Codex CLI on DeepSeek V4.1 Flash Xploded through the MagmaRouter Responses endpoint: the config.toml provider block, what is translated and what is not.
- [DeepSeek V4.1 Flash context window](https://magmarouter.com/learn/deepseek-v4-1-flash-context-window/): DeepSeek V4.1 Flash has a 1,048,576-token context window, and the uncensored build on MagmaRouter writes up to 393,216 tokens in one reply. How the two limits interact and how they compare across the line.