mantismantis
selfhost

Getting provider access

Every hosted provider mantis ships in the catalog — the four first-party vendor APIs (Claude, OpenAI, Gemini, Grok) and the open-model hosts: where to get the key, the env var it reads, and the one-liner to turn it on. A provider is enabled the moment mantis can find its key — env var or saved via /enable (stored chmod 600 in ~/.mantis-agent/models.json).

Two ways to enable anything:

export DEEPSEEK_API_KEY=sk-...       # env — survives via your shell profile
/enable deepseek                     # in-app — prompts for the key, validates, saves

…or just pick a locked 🔒 model in /models and paste the key when asked.

Open-model hosts

provider get a key env var notes
DeepSeek platform.deepseek.com DEEPSEEK_API_KEY dirt-cheap V3/R1, official host
Moonshot (Kimi) platform.moonshot.ai MOONSHOT_API_KEY kimi-k2 — top-tier tool calling
Z.ai (GLM) z.ai/model-api ZHIPUAI_API_KEY (or ZAI_API_KEY/ZHIPU_API_KEY) glm-4.7 official host
Alibaba (Qwen) Model Studio DASHSCOPE_API_KEY (or QWEN_API_KEY) qwen-max / qwen3 international endpoint
Groq console.groq.com/keys GROQ_API_KEY free tier; absurdly fast gpt-oss + kimi
OpenRouter openrouter.ai/keys OPENROUTER_API_KEY one key, ~every model; :free variants
Together api.together.xyz TOGETHER_API_KEY broad OSS menu
Fireworks fireworks.ai FIREWORKS_API_KEY fast OSS serving
Cerebras cloud.cerebras.ai CEREBRAS_API_KEY free tier; fastest tokens/s anywhere

First-party vendor APIs

A bare model name is enough for each of these — claude-opus-5, gpt-5.4, gemini-2.5-pro, grok-4 route to their vendor on their own.

provider get a key env var notes
Claude (Anthropic) console.anthropic.com — or your Claude subscription ANTHROPIC_API_KEY · ANTHROPIC_AUTH_TOKEN first-class, native Messages API; paste an API key (sk-ant-api…) or a subscription OAuth token (sk-ant-oat…) and mantis tells them apart
Claude via a gateway Bedrock Access Gateway, Azure Foundry, LiteLLM ANTHROPIC_AUTH_TOKEN + a /anthropic/v1 backend URL Authorization: Bearer instead of x-api-key
OpenAI platform.openai.com OPENAI_API_KEY gpt-5.x — mantis handles the max_completion_tokens/temperature quirks
Google (Gemini) aistudio.google.com/apikey GEMINI_API_KEY (or GOOGLE_API_KEY) free tier via AI Studio
Grok (xAI) console.x.ai XAI_API_KEY (or GROK_API_KEY) grok-4 / grok-4-fast / grok-3; reasoning_effort mapped where the model takes it

Free ways to start (no card)

  1. Ollama — local, free forever: mantis setup pulls a model for you.
  2. Groq / Cerebras free tiers — hosted OSS with generous limits.
  3. OpenRouter :free models — e.g. meta-llama/llama-3.3-70b-instruct:free.
  4. Gemini via AI Studio — free quota on 2.5-flash.

How enablement actually works

  • Env vars win over saved keys; alias vars (GOOGLE_API_KEY, GROK_API_KEY, ZAI_API_KEY, QWEN_API_KEY) are honored, and the vendor's own key wins by host — a stale OPENAI_API_KEY is never sent to xAI or Google.
  • Keys/URLs are whitespace-stripped and validated on save — /enable does a live /models probe and refuses to store a bad key. A locked provider prints the exact fix: /enable xai · xai-… · get one at console.x.ai.
  • /models groups by family — Local, OpenAI, Claude, Gemini, Grok, Hosted OSS, Self-host — with each group's auth state on its header; locked 🔒 rows sit below the enabled ones so you can see the whole menu and enable inline. Bare /enable or /disable print the same table as text.
  • /disable <provider> forgets a saved key.
  • Switching models mid-session (/model kimi, /model claude-opus-5, /model grok-4) re-wires the backend + key automatically and confirms the route in one line; your session context carries over.
  • mantis serve shows the same five families as cards with a one-click connection test — see the dashboard.

Self-hosting instead? See Self-hosting models — or let mantis deploy the model for you.

← previous
Other platforms
next →
OpenRouter