Using OpenClaw with abliteration.ai

Before you start, make sure the basics are in place.

• An abliteration.ai account with credits. Purchase credits here.
• An API key (ak_...) from your abliteration.ai dashboard.
• OpenClaw installed and the openclaw CLI available in your shell. See the OpenClaw repo for platform-specific install steps.
• A working config directory at ~/.openclaw/. Running openclaw onboard once creates it if it does not exist.

OpenClaw reads its primary config from ~/.openclaw/openclaw.json (legacy path ~/.clawdbot/clawdbot.json is still symlinked on upgraded installs). The file is JSON5, so comments, unquoted keys, and trailing commas are allowed.

Export your API key first so the ${ABLITERATION_API_KEY} interpolation resolves at runtime.

# Put this in ~/.bashrc, ~/.zshrc, or ~/.profile
export ABLITERATION_API_KEY="ak_YOUR_API_KEY"

# Reload your shell
source ~/.zshrc   # or source ~/.bashrc

# Open the config in your editor
${EDITOR:-vi} ~/.openclaw/openclaw.json

Paste this provider block into ~/.openclaw/openclaw.json. The merge mode keeps OpenClaw's built-in providers intact so you can still switch to other models from openclaw onboard.

{
  models: {
    mode: "merge",
    providers: {
      abliteration: {
        baseUrl: "https://api.abliteration.ai/v1",
        apiKey: "${ABLITERATION_API_KEY}",
        api: "anthropic-messages",
        models: [
          {
            id: "abliterated-model",
            name: "Abliterated Model",
            reasoning: false,
            input: ["text"],
            contextWindow: 128000,
            maxTokens: 8192,
            cost: {
              input: 0.005,
              output: 0.005,
            },
          },
        ],
      },
    },
  },
  agents: {
    defaults: {
      model: {
        primary: "abliteration/abliterated-model",
        fallbacks: [],
      },
    },
  },
}

OpenClaw can dump its live JSON schema and validate the current config before you launch an agent.

# Print the schema OpenClaw uses to validate the config
openclaw config schema | less

# Show the resolved (merged) config that OpenClaw will use
openclaw config show

# Quick connectivity test against /v1/messages
curl -s https://api.abliteration.ai/v1/messages \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $ABLITERATION_API_KEY" \
  -d '{
    "model": "abliterated-model",
    "max_tokens": 64,
    "messages": [{"role": "user", "content": "ping"}]
  }' | python3 -m json.tool

Once the provider is in place and agents.defaults.model.primary is set, OpenClaw's agent uses abliteration.ai automatically. The agent command does not accept a positional prompt or a per-run --model flag — pass the prompt with --message and select an agent profile with --agent. The reserved default agent id is main.

To switch models, edit agents.defaults.model.primary in ~/.openclaw/openclaw.json, or create a dedicated agent profile with openclaw agents add that pins its own model.

# List available agents (main is the reserved default id)
openclaw agents list

# Send a one-off message to the default agent using the primary model
openclaw agent --agent main --message "summarize this directory"

# Raise the reasoning effort for a single run
openclaw agent --agent main --thinking high --message "refactor foo.py for readability"

# Emit structured JSON instead of the default formatted output
openclaw agent --agent main --json --message "list the files in this repo"

# Bypass the gateway and run the embedded agent directly
openclaw agent --agent main --local --message "ping"

OpenClaw's anthropic-messages adapter is a drop-in client for Anthropic-compatible gateways. The custom provider you defined has the same contract as OpenClaw's built-in Anthropic support.

• Inference path: POST /v1/messages.
• Token counting: POST /v1/messages/count_tokens.
• Auth: Authorization: Bearer ak_... sourced from ${ABLITERATION_API_KEY}.
• Streaming: Anthropic-style SSE, which is the default for agent loops and interactive turns.
• Tool calling: native Messages API tool use, including multi-turn agentic tool call chains.
• Rate limits: the normal abliteration.ai limits apply (120 req/min for API key callers).

If you prefer to route OpenClaw through the OpenAI Chat Completions surface, abliteration.ai also exposes /v1/chat/completions. Use api: "openai-completions" instead. This path is useful if another part of your stack already standardizes on OpenAI-style payloads.

Avoid api: "openai-responses" for custom providers on current OpenClaw releases — it has known issues for user-defined providers (see openclaw/openclaw#43018). Prefer the anthropic-messages config at the top of this page.

{
  models: {
    mode: "merge",
    providers: {
      "abliteration-openai": {
        baseUrl: "https://api.abliteration.ai/v1",
        apiKey: "${ABLITERATION_API_KEY}",
        api: "openai-completions",
        models: [
          {
            id: "abliterated-model",
            name: "Abliterated Model (OpenAI wire)",
            input: ["text"],
            contextWindow: 128000,
            maxTokens: 8192,
            cost: { input: 0.005, output: 0.005 },
          },
        ],
      },
    },
  },
  agents: {
    defaults: {
      model: {
        primary: "abliteration-openai/abliterated-model",
      },
    },
  },
}

Most setup issues are config resolution, env var interpolation, or model id mismatches.

• error: unknown option '--model' — openclaw agent has no --model flag. Model selection is config-driven: change agents.defaults.model.primary in ~/.openclaw/openclaw.json, or use openclaw agents add to pin a model on a specific agent profile. The supported flags are --agent, --message, --to, --session-id, --thinking, --channel, --reply-to, --reply-channel, --reply-account, --local, --deliver, --timeout, --json, and --verbose.
• 401 Unauthorized — verify ABLITERATION_API_KEY is exported in the same shell that launched OpenClaw, and that it starts with ak_. Run echo $ABLITERATION_API_KEY to confirm.
• Requests going to api.anthropic.com instead of abliteration.ai — the api field is probably missing, which causes OpenClaw to silently fall back to the built-in Anthropic provider. Make sure api: "anthropic-messages" is set on your provider (see openclaw/openclaw#23332).
• Model not found — the id inside models[] must be exactly abliterated-model, and agents.defaults.model.primary must use the <provider>/<model-id> form (e.g. abliteration/abliterated-model).
• Connection refused / wrong base URL — baseUrl must be exactly https://api.abliteration.ai/v1. Do not omit /v1 and do not append /messages yourself.
• Built-in provider wins over your custom one — naming your provider anthropic can be ignored on some OpenClaw releases because the built-in provider has priority (see openclaw/openclaw#56679). Use a different name such as abliteration.
• 402 Insufficient Credits — top up from the pricing page. OpenClaw agent loops can spend credits quickly.
• 429 Rate limited — back off using the Retry-After header, or reduce parallel sub-agents.
• Schema validation errors — run openclaw config schema to inspect the live schema, then run openclaw config show to see how OpenClaw merged your file with its defaults.