You write the agent. LMbox provides the runtime.
The official SDK to package and deploy a business AI agent on your customer's LMbox appliance. YAML manifest, system prompt, golden eval suite, HMAC signature. Count 2 to 3 days of development, from manifest to a production agent.
Three commands, from skeleton to signed deployment.
The SDK encodes the whole agent workflow - scaffolding, validation, build, pack, deploy. No manual step, no CI pipeline to write.
Scaffold
Start from a template (legal-document, _base) or a blank agent. The CLI generates the manifest, system prompt, test suite, README.
lmbox agent new nda-reviewer --vendor magellan
Iterate
Edit the manifest and prompt. The CLI validates the manifest against the official schema (JSON Schema draft 2020-12) and runs the golden suite against a local or remote model.
lmbox agent test ./nda-reviewer
Deploy
The agent is packaged as a HMAC-signed .lmbox + JSON sidecar. The deploy command pushes it to the customer's LMbox. The heartbeat picks it up on the next cycle.
lmbox agent deploy ./nda-reviewer --box BOX-XXX
Four commands to make hallucination physically impossible.
The SDK ships the 4 anti-hallucination layers as CLI commands. The partner wires each layer on their agent via manifest flags and gets the same defence-in-depth as first-party LMbox agents. Understand the stack.
lmbox agent verify
Layer A
Check every citation against Légifrance + EUR-Lex
Reads the agent output, extracts references (11 families: cassation, court of appeal, council of state, constitutional council, Code articles, statutes, decrees, EU regulations…) and calls Légifrance + EUR-Lex live. CRITICAL on unfound or malformed reference. Exit code usable in partner CI.
lmbox agent run --guard strict
Layer B
Live run with real-time guard
Streams the agent and cancels generation in 200 ms when a hallucination is detected. Three modes: strict (cancel), annotate (inline [⚠] marker), warn (observability). JSON audit trail exportable for the chain of evidence.
lmbox agent lint-schema --strict
Layer C
Lint the output JSON Schema
7 rules: root_must_be_object, required_missing_in_properties, missing_description, missing_additional_properties, unbounded_string, oversized_enum, unspecified_array_items. Run as a pre-commit hook at the partner - saves a week of blind prompt-engineering.
lmbox agent check-grounding
Layer D
Verify every cited source was actually retrieved
Compares source_ids in the output to the tool-call trace. If the agent claims to cite interne-2019-453 but search_dossiers_internes NEVER returned that ID this turn → blocked. Invented sources made structurally impossible.
Everything fits in a 4-file folder.
No framework, no proprietary DSL. YAML, Markdown, JSON Lines. Readable by a lawyer or a Compliance Officer, executable by an AI agent.
-
manifest.yamlThe agent's contract: model, prompts, tools, required connectors, eval thresholds, deployment parameters (audit, PII redaction). Validated against `lmbox.ai/v1`.
-
prompts/system.mdThe system prompt encoding the business logic. This is what you, the partner, bring: your sector expertise translated into clear instructions.
-
evals/golden.jsonlGolden test suite - one JSON-Lines case per situation, with assertions on expected output. Blocking in CI: an agent doesn't ship below its threshold.
-
tools/Optional Python code for function-calling tools (RAG, actions, external API calls). Empty for pure-analyst agents.
apiVersion: lmbox.ai/v1
kind: Agent
metadata:
slug: nda-reviewer
version: 0.1.0
vendor: lmbox
vertical: legal
display_name: NDA Reviewer
description: >-
Lit chaque NDA, identifie les clauses non-standard,
rédige un mémo d'amendements.
spec:
model:
primary: mistral-large-2
fallback: qwen2.5-32b
temperature: 0.1
max_tokens: 1024
prompts:
system: prompts/system.md
tools:
- name: search_clause_library
type: rag
source: connectors.sharepoint.legal-templates
evals:
pass_threshold: 0.8
golden: evals/golden.jsonl
deployment:
audit: true
rgpd_redact: [nir, iban, phone]
Turn 15 days of prompt engineering into 2-3 days of packaging.
Iteration cycle
From client brief to agent in production. The SDK covers 100% of the cycle: scaffold, validate, test, package, sign, deploy.
Project revenue per agent
Discovery + prompt engineering + RAG + connectors + UAT + MEP. A clean deliverable you bill man-day to your end client.
Partner IP
The agent you write is yours. You keep the right to repackage it as a vertical offer and resell it on the Partner Marketplace.
Let's talk about your first agent.
If you're a service integrator, consulting firm, or vertical software vendor, we have a structured partner program: early-phase SDK access, 2-day Certified Agent Engineer training, hands-on support for your first deployment.