Levanto Sage for OpenClaw
Official external plugin for native typed decisions with Levanto Sage. It supports Boolean, Choice, Score, Sort, and Tags questions through OpenClaw's prepared Decision V2 API. It does not provide chat completions.
Setup
Released OpenClaw 2026.10.4 exposes V1 only. For development, use a matching
source checkout containing the prepared V2 runtime and this plugin. Packaged
use requires a supporting paired host/plugin release; the older installation
floor alone does not establish V2 compatibility.
Enable the plugin and connect a Levanto credential through the normal provider connection flow:
openclaw models auth login --provider levanto
The common credential owner supports agent-scoped profiles and SecretRefs.
LEVANTO_API_KEY is the advertised environment source. Select
levanto/levanto-sage as the agent's decisionModel. Connecting a credential
does not select a model or schedule background work.
Evaluations send explicitly supplied evidence to Levanto and consume the account's allowance. The adapter supports inline images for supported operations, but does not enable grounding or perform implicit web searches. It retains abstentions, partial batch errors, and reported usage without inventing cost.
See the Levanto Sage guide for configuration, operation limits, billing, and result semantics.
Live adapter test
From a matching source checkout, inject a Levanto-issued LEVANTO_API_KEY
through your secret manager, then run:
OPENCLAW_LIVE_TEST=1 node scripts/run-vitest.mjs --config test/vitest/vitest.live.config.ts extensions/levanto/registration.live.test.ts
The test makes five paid requests with reasoning off: four Boolean/Choice/Score/Tags batches using synthetic text, PNG, JPEG, and WebP evidence, followed by one Sort request. These requests consume the account's allowance and submit the synthetic images to Levanto. It skips without both the opt-in flag and key. It checks the registered V2 adapter and real transport, accepts native abstention, and validates reported usage when present. Prepared credentials are fixture input; the test does not exercise host credential discovery or change model selection.
