Setup AI

Spectre Scan integrates with two AI providers — pick whichever account you have credit on:

• OpenAI (openai.com) — openai plugin / provider.
• Anthropic Claude (anthropic.com) — claude plugin / provider.

Both expose the same surface: per-issue analysis (description, remediation, exploit, dissect, insights, patch, report) at scan time, and the Djin! side-dock chat assistant in the Pro web UI.

CLI

OpenAI:

bin/spectre --plugin=openai:apikey=YOUR_OPENAI_KEY [URL]

Claude:

bin/spectre --plugin=claude:apikey=YOUR_ANTHROPIC_KEY [URL]

Optional plugin parameters (both providers accept them):

• model — override the default model (gpt-4o for OpenAI, claude-opus-4-5 for Claude).
• min_severity — restrict per-issue analysis to issues at or above this severity (informational, low, medium, high). Default: medium. Lower-severity findings are skipped, saving API tokens.

Example with overrides:

bin/spectre \
    --plugin=claude:apikey=YOUR_KEY,model=claude-haiku-4-5,min_severity=high \
    [URL]

REST API

Drive the same per-issue plugin from POST /instances by adding the provider to the plugins hash on the options body. Same apikey / model / min_severity parameters the CLI accepts.

OpenAI:

{
  "url":    "http://example.com/",
  "checks": ["*"],
  "plugins": {
    "openai": { "apikey": "YOUR_OPENAI_KEY" }
  }
}

Claude:

{
  "url":    "http://example.com/",
  "checks": ["*"],
  "plugins": {
    "claude": { "apikey": "YOUR_ANTHROPIC_KEY" }
  }
}

Plugin parameters layer in as siblings of apikey:

{
  "plugins": {
    "claude": {
      "apikey":       "YOUR_KEY",
      "model":        "claude-haiku-4-5",
      "min_severity": "high"
    }
  }
}

Only one of openai / claude should be present per scan — they both annotate the same issues, so loading both wastes API tokens on duplicate work.

MCP

The MCP server forwards its spawn_instance.options straight to the engine, so the plugins shape is identical to REST — just nested under the tools/call envelope.

OpenAI:

{
  "jsonrpc": "2.0", "id": 1, "method": "tools/call",
  "params": {
    "name": "spawn_instance",
    "arguments": {
      "options": {
        "url":    "http://example.com/",
        "checks": ["*"],
        "plugins": {
          "openai": { "apikey": "YOUR_OPENAI_KEY" }
        }
      }
    }
  }
}

Claude:

{
  "jsonrpc": "2.0", "id": 1, "method": "tools/call",
  "params": {
    "name": "spawn_instance",
    "arguments": {
      "options": {
        "url":    "http://example.com/",
        "checks": ["*"],
        "plugins": {
          "claude": { "apikey": "YOUR_ANTHROPIC_KEY" }
        }
      }
    }
  }
}

model and min_severity slot in alongside apikey exactly as they do over REST.

The plugins key is documented in full at spectre://options/reference (or inlined under the Options reference on the MCP page).

Web UI

In Settings → Djin! AI assistant:

1. Pick a provider (OpenAI or Claude) from the dropdown.
2. Paste your API key. Spectre Pro pings the provider on save to verify the key — invalid keys / billing issues / no model access surface as a precise error rather than a silent failure.
3. Save.

That alone enables Djin! — the in-app side-dock AI assistant. Two extra toggles in the same section control the rest:

• Auto-analyze issues during scan — off by default. When on, the scanner attaches the same provider’s per-issue plugin to every scan, expanding each finding’s description / remediation / exploit fields automatically. Costs roughly one provider call per issue.
• Djin! daily token budget — per-user 24-hour cap on prompt
  • completion tokens summed across every Djin! conversation. 0 disables the cap.

Apex Recon ships with the same Djin! chat assistant, but no per-issue plugin (Apex’s domain is input-vector discovery, not security findings — there’s nothing per-record for the AI to auto-expand). Configuring the AI key in Apex Settings only enables Djin!.