orreth-agent

The SDK for building agents that live in an Orreth universe — a governed runtime where every agent has a permanent identity, every action is a signed record, and every model call is authorized and metered by a kernel that never sees the prompt.

pip install orreth-agent            # one dependency: cryptography
pip install "orreth-agent[governed]"  # + litellm, for real model calls

from orreth_agent import FieldClient, Chassis, RuleThink

client = FieldClient("http://127.0.0.1:4502", name="scout")
client.join()   # a governed request — a human approves your admission
Chassis(client, RuleThink()).run("say hello to the universe")
# …and the agent is now visible in the console: roster, diary, spend.

What the kit gives you:

• FieldClient — join a world's floor through its governed gate, keep a permanent identity (KeyPair seeds persist under ~/.orreth/agents/<name>/ — the same agent re-joins across runs), remember and recall signed records, and think through the kernel's metered model gateway.
• Chassis — the fixed sense→think→act→record loop; bring your own thinking (RuleThink for deterministic logic, GovernedThink for model calls through the gateway) or keep your existing framework and use the client alone — a LangGraph agent joins the same way.
• Content guardrails — pass a composed rule set to GovernedThink(client, rails={...}) and the lane checks content the same way the kernel's own lanes do: inputs before the model call (a refused exchange returns a plain refusal sentence, spends nothing), outputs before they return (masked where the rules say mask). Detectors for payment cards (Luhn-checked) and personal identifiers ship built in; orreth_agent.rails is parity-tested against the kernel reference.
• OrrethMind / @generation — declare model-backed generations whose prompts are acquired from the world's own governed shelf (acquire), so the words your agent runs on are versioned, human-editable assets, never hardcoded strings.
• Per-style attribution (new in 0.3.0) — ask(variant=...) chooses any of the world's eleven retrieval styles by name (the reply always says which served); a mind's generation takes _variant= per call, and the metered cost of that thought lands on the fuel ledger wearing the style's name — per-style cost is a query, never an estimate.
• FieldClient.ask() (new in 0.2.0) — ask a world a question through its governed retrieval and read the structured answer: the reply, its citations, which retrieval variant served it, and the whole decision record behind the choice. Asks are signed with your agent's own key — never a bearer token — and repeated questions may serve from the world's cache, always confessed in the envelope.
• python -m orreth_agent.joindoor (new in 0.2.0) — the publishable join door: a small service an operator runs beside their own kernel to admit agents through the governed gate. The operator's root key signs the door's credential once, offline, and never enters the serving process; mint · serve · pending · approve complete a stranger's admission from published artifacts alone.
• manifest / PANEL_KINDS — build and validate capability manifests: whole purposes the Orreth console renders from declarations, never from your code.
• canonical / content_hash / KeyPair — the exact signing and content-addressing bytes the kernel verifies, held to byte-for-byte parity with the reference implementation by the test suite.

Start at docs.orreth.ai — the quickstart gets a governed universe running in ten minutes; Build your first world stands one up from three files against the published kernel image (ghcr.io/iotlodge/orrethd).

Licensed Apache-2.0.