Configuration

Wallbreaker uses config.toml. Copy the example and edit:

[profiles.*] — Provider Endpoints

Profiles are reusable provider endpoints. protocol selects the wire format (openai, anthropic, xai, or claude-code); api_key_env names the environment variable that holds the key:

default_profile = "openrouter"
 
[profiles.openrouter]
protocol = "openai"
base_url = "https://openrouter.ai/api/v1"
api_key_env = "OPENROUTER_API_KEY"
model = "anthropic/claude-sonnet-4.5"
 
[profiles.zai]
protocol = "anthropic"
base_url = "https://api.z.ai/api/anthropic"
api_key_env = "ZAI_API_KEY"
model = "glm-4.6"
 
[profiles.claude-code]
protocol = "claude-code"
model = "sonnet"   # CLI alias: sonnet / opus / haiku, or a full model id

claude-code uses the local claude CLI as the brain — no base URL or API key needed (the CLI authenticates itself). xai is OpenAI-wire-compatible: set protocol = "xai" and model, and the key env defaults to XAI_API_KEY.

[agents.*] — Role Assignment

Each role (attacker, target, judge) is assigned a provider/model directly, or a named profile:

[agents.attacker]
provider = "openrouter"
model = "anthropic/claude-sonnet-4.5"
 
[agents.target]
provider = "openrouter"
model = "openai/gpt-4o-mini"
 
[agents.judge]
provider = "openrouter"
model = "openai/gpt-4o-mini"

[agent_profiles.*] — Named Role Profiles

Role-specific profiles add prompt behavior on top of an endpoint. prompt_source is "none", "inline", or "file" (inline text and a file are mutually exclusive):

[agent_profiles.target."default target"]
provider = "openrouter"
model = "openai/gpt-4o-mini"
prompt_source = "inline"
system_prompt = "You are the model under an authorized red-team evaluation."

[target] — The Model Under Test

[target]
protocol = "openai"
base_url = "https://openrouter.ai/api/v1"
api_key_env = "OPENROUTER_API_KEY"
model = "openai/gpt-4o-mini"

Toggle system mode live in the TUI with /target sysmode merge.

[judge] — Verdict Evaluator

The judge grades text targets and vision-grades generated images — for image targets it MUST be a vision-capable model:

[judge]
protocol = "openai"
base_url = "https://openrouter.ai/api/v1"
api_key_env = "OPENROUTER_API_KEY"
model = "openai/gpt-4o-mini"

[[mcp.servers]] — MCP Tool Servers (client side)

Each enabled server is spawned over stdio at startup and its tools are proxied into the agent’s tool registry:

[[mcp.servers]]
name = "custom-tools"
command = "python"
args = ["-m", "my_tool_server"]
# env = { MY_VAR = "value" }
# enabled = true
# tool_prefix = ""

Schema fields: name, command, args, env, enabled, tool_prefix. The P4RS3LT0NGV3 engine is wired in natively as parsel_* tools (vendored via wallbreaker parsel update, Node.js required) — the example server block in config.example.toml is only for running that same engine out-of-process.

Wallbreaker also runs as an MCP server exposing wb_* tools to other agents — see MCP Server.

State Files

JSONL is the source of truth; the SQLite index may be deleted and rebuilt at any time. All of these are gitignored.