Agents

Claude Code

Claude Code speaks the Anthropic format. Set the base URL and the key as environment variables and point every model role at a TOKENWERK model.

curl -fsSL https://claude.ai/install.sh | bash

export ANTHROPIC_BASE_URL="https://api.tokenwerk.ai"
export ANTHROPIC_AUTH_TOKEN="$TOKENWERK_API_KEY"
export ANTHROPIC_MODEL="MODEL"
export ANTHROPIC_DEFAULT_OPUS_MODEL="MODEL"
export ANTHROPIC_DEFAULT_SONNET_MODEL="MODEL"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="MODEL"
export CLAUDE_CODE_MAX_CONTEXT_TOKENS="CONTEXT_WINDOW"
claude

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.tokenwerk.ai",
    "ANTHROPIC_AUTH_TOKEN": "tw-...",
    "ANTHROPIC_MODEL": "MODEL",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "MODEL",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "MODEL",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "MODEL",
    "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "CONTEXT_WINDOW"
  }
}

• CLAUDE_CODE_MAX_CONTEXT_TOKENS is the model's context window (see the model list). Claude Code doesn't know TOKENWERK's models and would otherwise assume 200,000 tokens, so it would compact the conversation too late or too early.
• Only put the key in ~/.claude/settings.json or .claude/settings.local.json, not in a project's shared .claude/settings.json.
• Anthropic does not officially support using models other than Claude through gateways. If a model answers beta features with 400 errors, CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 helps.

Official guide: Claude Code: LLM gateways ↗

opencode

opencode adds TOKENWERK as its own provider via the OpenAI format. Save the configuration as opencode.json in your project or globally as ~/.config/opencode/opencode.json.

curl -fsSL https://opencode.ai/install | bash

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "tokenwerk": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "TOKENWERK",
      "options": {
        "baseURL": "https://api.tokenwerk.ai/v1",
        "apiKey": "{env:TOKENWERK_API_KEY}"
      },
      "models": {
        "MODEL": {
          "name": "MODEL"
        }
      }
    }
  },
  "model": "tokenwerk/MODEL"
}

• Optionally add context and output under models.<name>.limit (values under Models), so opencode compacts long sessions in time.

Official guide: opencode: Providers ↗

OpenClaw

OpenClaw gets TOKENWERK as its own provider in ~/.openclaw/openclaw.json.

npm install -g openclaw@latest
openclaw onboard --install-daemon

{
  agents: { defaults: { model: { primary: "tokenwerk/MODEL" } } },
  models: {
    mode: "merge",
    providers: {
      tokenwerk: {
        baseUrl: "https://api.tokenwerk.ai/v1",
        apiKey: "${TOKENWERK_API_KEY}",
        api: "openai-completions",
        models: [{ id: "MODEL", name: "MODEL" }],
      },
    },
  },
}

• For the Anthropic format, set baseUrl to https://api.tokenwerk.ai and api to "anthropic-messages".
• Every model must be listed under models with its id; you can add contextWindow and maxTokens there.

Official guide: OpenClaw: Custom providers ↗

Hermes Agent

Hermes Agent by Nous Research connects through a “Custom endpoint”. The easiest way is hermes model, or directly in the configuration:

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

model:
  provider: custom
  base_url: https://api.tokenwerk.ai/v1
  default: MODEL
  key_env: TOKENWERK_API_KEY

• Hermes needs a model with tool calling and at least 64,000 tokens of context.

Official guide: Hermes Agent: Providers ↗

Codex CLI

Codex CLI only speaks the Responses API (/v1/responses). TOKENWERK passes it through; it works with models that support this API.

npm install -g @openai/codex
            

model = "MODEL"
model_provider = "tokenwerk"

[model_providers.tokenwerk]
name = "TOKENWERK"
base_url = "https://api.tokenwerk.ai/v1"
env_key = "TOKENWERK_API_KEY"
wire_api = "responses"

Official guide: Codex: Configuration ↗

Aider

Aider uses the OpenAI format; the model name gets the prefix openai/.

python -m pip install aider-install
aider-install

export OPENAI_API_BASE="https://api.tokenwerk.ai/v1"
export OPENAI_API_KEY="$TOKENWERK_API_KEY"
aider --model openai/MODEL

• You can ignore warnings about unknown models; Aider only knows the models of large providers.

Official guide: Aider: OpenAI compatible APIs ↗

Cline, Roo Code, Kilo Code

The three VS Code extensions have an “OpenAI Compatible” provider. Enter in their settings:

• In the model settings you can set the context window, maximum output and image input.
• Kilo Code also offers “Anthropic Messages”, then with the base URL https://api.tokenwerk.ai.
• Roo Code only works with models that support tool calling.

Official guides: Cline ↗ · Roo Code ↗ · Kilo Code ↗

Continue

In Continue's config.yaml (opened from the extension's settings), add a model with the provider openai:

models:
  - name: TOKENWERK MODEL
    provider: openai
    model: MODEL
    apiBase: https://api.tokenwerk.ai/v1
    apiKey: tw-...
    capabilities:
      - tool_use
    roles:
      - chat
      - edit
      - apply

Official guide: Continue: config.yaml ↗

Goose

Goose gets TOKENWERK as its own provider: create the file, then choose the provider in goose configure.

curl -fsSL https://github.com/aaif-goose/goose/releases/download/stable/download_cli.sh | bash

{
  "name": "tokenwerk",
  "engine": "openai",
  "display_name": "TOKENWERK",
  "description": "TOKENWERK API",
  "api_key_env": "TOKENWERK_API_KEY",
  "base_url": "https://api.tokenwerk.ai/v1/chat/completions",
  "models": [{ "name": "MODEL", "context_limit": 128000 }],
  "supports_streaming": true,
  "requires_auth": true
}

• Set context_limit to the model's context window (see Models).

Official guide: Goose: Providers ↗

Zed

Zed adds TOKENWERK as an OpenAI compatible provider in its settings.json. For the provider tokenwerk, Zed reads the key from TOKENWERK_API_KEY automatically; otherwise it asks for it on first use.

{
  "language_models": {
    "openai_compatible": {
      "tokenwerk": {
        "api_url": "https://api.tokenwerk.ai/v1",
        "available_models": [
          { "name": "MODEL", "display_name": "MODEL", "max_tokens": 128000 }
        ]
      }
    }
  }
}

• max_tokens is the model's context window (see Models).

Official guide: Zed: API access ↗