Configuration

All Toolpack SDK configuration is passed directly to Toolpack.init(). There is no file-based config discovery — settings are explicit code.

Toolpack.init() Options

Option	Type	Default	Description
provider	string	—	Single provider shorthand ('openai', 'anthropic', 'gemini', etc.)
apiKey	string	—	API key override (defaults to env var)
model	string	—	Default model name
tools	boolean	false	Enable built-in tools
toolsConfig	Partial<ToolsConfig>	{}	Tool behavior overrides
customModes	ModeConfig[]	[]	Additional modes to register
defaultMode	string	'default'	Mode to activate on init
logging	LoggingConfig	—	File logging settings
hitl	HitlConfig	—	Human-in-the-loop confirmation
onToolConfirm	callback	—	Confirmation handler (also enables HITL)
contextWindow	ContextWindowConfig	—	Automatic conversation pruning/summarization

Tools Configuration

Pass toolsConfig to Toolpack.init():

const toolpack = await Toolpack.init({
    provider: 'openai',
    tools: true,
    toolsConfig: {
        enabled: true,
        autoExecute: true,
        maxToolRounds: 10,
        toolChoicePolicy: 'auto',
        enabledTools: [],
        enabledToolCategories: [],
        toolSearch: {
            enabled: false,
            alwaysLoadedTools: [],
            alwaysLoadedCategories: [],
            searchResultLimit: 5,
            cacheDiscoveredTools: true,
        },
        additionalConfigurations: {
            MY_CUSTOM_API_KEY: process.env.MY_CUSTOM_API_KEY,
        },
    },
});

Tools Options

Option	Type	Default	Description
enabled	boolean	true	Enable tool system
autoExecute	boolean	true	Automatically execute tool calls
maxToolRounds	number	5	Max tool execution rounds per request
toolChoicePolicy	string	"auto"	"auto", "required", or "required_for_actions"
enabledTools	string[]	[]	Specific tools to enable (empty = all)
enabledToolCategories	string[]	[]	Categories to enable (empty = all)
additionalConfigurations	object		Key-value config passed dynamically to custom tools via ToolContext

Tool Categories

Category	Description
filesystem	File system operations
execution	Command execution
system	System information
network	HTTP and web tools
coding	Code analysis tools
version-control	Git operations
diff	Diff and patch tools
database	Database operations
cloud	Cloud deployment

Tool Search

For large tool sets, enable on-demand tool discovery:

toolsConfig: {
    toolSearch: {
        enabled: true,
        alwaysLoadedTools: ['fs.read_file', 'exec.run'],
        alwaysLoadedCategories: ['filesystem'],
        searchResultLimit: 5,
        cacheDiscoveredTools: true,
    },
}

Logging Configuration

Pass logging to Toolpack.init():

const toolpack = await Toolpack.init({
    provider: 'openai',
    logging: {
        enabled: true,
        filePath: './toolpack-sdk.log',
        level: 'debug',
        console: false,
    },
});

Option	Type	Default	Description
enabled	boolean	false	Enable file logging
filePath	string	toolpack-sdk.log	Log file path (relative to CWD)
level	string	info	Log level (error, warn, info, debug, trace)
console	boolean	false	Mirror log output to console

Environment variables override programmatic config (highest precedence):

export TOOLPACK_SDK_LOG_FILE="./toolpack-sdk.log"   # also enables logging
export TOOLPACK_SDK_LOG_LEVEL="debug"
export TOOLPACK_SDK_LOG_ENABLED="true"
export TOOLPACK_SDK_LOG_CONSOLE="true"

Environment Variables

Variable	Description
OPENAI_API_KEY	OpenAI API key
ANTHROPIC_API_KEY	Anthropic API key
GEMINI_API_KEY	Google Gemini API key
TOOLPACK_OPENAI_KEY	Alternative OpenAI key
TOOLPACK_ANTHROPIC_KEY	Alternative Anthropic key
TOOLPACK_GEMINI_KEY	Alternative Gemini key
TOOLPACK_SDK_LOG_FILE	Log file path (enables logging)
TOOLPACK_SDK_LOG_LEVEL	Log level override (error, warn, info, debug, trace)
NETLIFY_AUTH_TOKEN	Netlify deployment token

Context Window Management

The contextWindow init option controls automatic conversation pruning and summarization. When the accumulated message history approaches the model's context limit, the SDK either prunes old messages or summarizes them before the next request.

const toolpack = await Toolpack.init({
    provider: 'openai',
    contextWindow: {
        enabled: true,                  // default: true
        strategy: 'prune',              // 'prune' | 'summarize' | 'fail'
        pruneThreshold: 85,             // trigger at 85% of context window
        maxMessageHistoryLength: 100,   // optional hard cap on message count
        summarizerModel: 'gpt-4.1-mini', // only used when strategy = 'summarize'
        retainSystemMessages: true,     // never prune system messages (default: true)
        outputTokenBuffer: 1.15,        // 15% safety buffer above maxOutputTokens
    },
});

ContextWindowConfig fields

Field	Type	Default	Description
enabled	boolean	true	Master switch for context window management
strategy	'prune' | 'summarize' | 'fail'	'prune'	What to do when the threshold is reached
pruneThreshold	number	85	Percentage of the context window that triggers cleanup
maxMessageHistoryLength	number	—	Optional cap on total message count, independent of token counting
summarizerModel	string	(current model)	Model to use for summarization — set to a faster/cheaper model to reduce cost
retainSystemMessages	boolean	true	Whether system messages are exempt from pruning
outputTokenBuffer	number	1.15	Safety multiplier applied to maxOutputTokens before computing available input space

Strategies:

• 'prune' — removes the oldest non-system messages until the history fits.
• 'summarize' — calls the LLM to produce a summary of removed messages and inserts it as a system message before continuing.
• 'fail' — throws an error instead of modifying the history.

---

maxToolRounds in AgentRunOptions

AgentRunOptions.maxToolRounds sets a per-run hard cap on the number of tool-call rounds the agent may execute. It overrides the toolsConfig.maxToolRounds value from Toolpack.init() and bypasses the query-classifier adjustment for that specific run.

// In your BaseAgent subclass:
const result = await this.run(prompt, {
    maxToolRounds: 1,   // single-shot: the LLM may call at most one tool round
});

interface AgentRunOptions {
    /** One-off workflow override for this specific run */
    workflow?: Record<string, unknown>;

    /**
     * Hard cap on tool-call rounds for this specific run.
     * Overrides ToolsConfig.maxToolRounds and bypasses the query-classifier
     * adjustment. Use for agents that should only make one tool call per
     * invocation (e.g. single-shot routers using delegate_to_agent).
     */
    maxToolRounds?: number;
}

maxToolRounds can also be set at the CompletionRequest level when calling toolpack.generate() directly:

await toolpack.generate({
    messages: [{ role: 'user', content: 'Which agent should handle this?' }],
    model: 'gpt-4o',
    maxToolRounds: 1,
});

---

Full Example

const toolpack = await Toolpack.init({
    provider: 'openai',
    apiKey: process.env.OPENAI_API_KEY,
    model: 'gpt-4o',
    tools: true,
    toolsConfig: {
        maxToolRounds: 10,
        additionalConfigurations: { MY_API_KEY: process.env.MY_API_KEY },
    },
    logging: { enabled: true, filePath: './toolpack.log', level: 'debug' },
    hitl: { enabled: true, confirmationMode: 'all' },
    onToolConfirm: async (tool) => askUser(`Allow ${tool.displayName}?`),
    customModes: [...],
    defaultMode: 'agent',
});

See ToolpackInitConfig for all options.