Standard Library and Defaults

Orchard ships with a small standard library of tool packs, a set of model providers, and a table of default constants and limits that wrap every run. The defaults are chosen so an agent is safe and useful with no policy block at all, and the host can only tighten them. This page lists what comes built in.

Built-in tool packs

You do not write every tool yourself. use <pack> at the top of an agent pulls in a set of ready-made tools the model can call. A pack tool is dispatched exactly like one you declared yourself, with the same validation, the same redaction, and the same accounting against the budget.

use calculator                 // a built-in pack: gives you calculate

calculator: adds a single calculate tool that does exact arithmetic, which is worth having because models are unreliable at math and this is not.
http: outbound HTTP requests, subject to the policy's domain allowlist and the egress guard.
web: keyless web search and readable page fetching.
browser: drives a real headless Chrome over the DevTools Protocol: browser_open, browser_read (the rendered page, after JavaScript runs), browser_click, browser_type, browser_eval, and browser_screenshot.
files: reading, writing, and listing files under a root directory you choose with use files { root: "." }.
shell: running shell commands, gated by the policy's allow_shell setting, which is off by default.
time: the current date and time.
mcp: connecting to a Model Context Protocol server and exposing its tools to the loop.

Providers

The provider enum selects which model backend a model call routes to. The mock and ollama providers are free and offline-friendly, so an agent can be developed and tested with no API key and no network.

anthropic: Claude models.
openai: OpenAI models.
ollama: local models, free and offline-friendly.
groq: models served on Groq.
together: models served on Together.
openrouter: models routed through OpenRouter.
mock: the offline test provider, free and deterministic, available to every binding.

The default model set is the latest Claude models, claude-opus-4-8 and claude-sonnet-4-6.

Defaults and limits

These are the constants the runtime ships with. Unless a policy block or a host override tightens them, these are the values in force.

max_tokens: 4096
conversation window: 40
top_k: 5
max_steps: 25
max_tool_calls: 100
max_requests_per_run: 50
allow_shell: never
on_violation: stop
max_gen_retries: 2
max_delegate_depth: 4
unattended-serve spend cap: $5.00
tool-result cap: 48 KiB
shell-stream cap: 24 KiB
HTTP body cap: 8 MiB
MAX_REDIRECTS: 5
circuit breaker: 3
embed batch: 64
chunk size / overlap: 1200 / 150
MCP protocol: 2025-06-18

Safety mechanisms

Several mechanisms wrap every run so that untrusted content, secrets, and runaway loops are handled before they cause harm.

sentinels: external content is wrapped in <<<external>>> and <<<end-external>>> so the model can tell host instructions from data it fetched.
redaction: secret values are replaced with «VAR» in output and in event payloads.
egress guard: outbound HTTP is checked against the policy's allowed domains before a request leaves.
shell gating: the shell pack is off unless allow_shell permits it.
circuit breaker: repeated tool failures trip a breaker rather than spinning.
unattended spend cap: an unattended serve run stops at the spend cap so a loop cannot run up an unbounded bill.

Note.Every one of these defaults composes with a host's policy by min: the host can only tighten them, never loosen them. See the policy section in the API Reference for how a host supplies its own budgets and gates.