Kirimana is the open-source, AI-native data contract and automation platform. Kiri assists you inside the product, every AI call is classification-gated, and any AI agent that speaks MCP can read your governance state without a bespoke integration. Enterprise-grade by design, equally at home in a 50-person growth team and a 50,000-person bank.
We borrowed kirimana — Māori for contract — because it captured the meaning English doesn't have a word for. A contract isn't paperwork. It's where your name carries weight. That's the brand. That's also the architecture.
Mandatory. The AI gateway refuses what doesn't name a human. When the regulator asks who decided this — there's an answer.
Every LLM call is gated by data classification. Restricted data never leaves your tenant. Every call is logged — prompt, response, cost, caller.
ODCS v3 canonical. The contract you write today runs on Databricks tomorrow and on Trino next year. No platform lock-in by design.
Most data tools bolt on a chat box. Kirimana started AI-first — the same governance state that powers contracts is what Kiri reads, what Skills extend, and what every external AI agent sees over Model Context Protocol. One trust ladder, one audit log, everywhere AI touches your data.
Every AI surface — UI, CLI, the docs chat above — speaks as Kiri. She knows your stack, your role, your contracts, and your audit history. She drafts contracts, explains lineage, triages failed applies, and never invents numbers she can't substantiate.
A skill is a folder with a SKILL.md and a few prompts. Versioned in git, reviewed in PR, runnable from Kiri or as a CLI step. Reference skills ship with the product — draft-gold-model, debug-apply, azure-deploy, incident-coach, explain-kpi, maturity-coach — and you write your own.
Classification gate first — restricted data never leaves your tenant. Anthropic prompt cache always on, so Kirimana's brand voice and contract context don't get re-tokenised every turn. Local Ollama provider for air-gapped. Every prompt, response, model, cost, caller — audit-logged.
Claude Code, Cursor, Cline, your in-house copilot, any agent that speaks MCP can read contracts, classifications, lineage, AI-policy decisions, release status — without a bespoke integration. Every MCP read and tool-call is itself audit-logged.
Mid-sized teams get the same governance, audit trail, and AI-policy enforcement that Fortune 500s spend millions on. Enterprises get more value, not less — federated libraries, hub-and-spoke domains, audit redaction, multi-environment CI/CD, OIDC RBAC, DORA + EU AI Act + GDPR reports — all of it compounds at scale. One product, one Apache-2.0 codebase, both audiences.
Same contracts, same governance, same Apache-2.0 core. The difference is the platform adapter wired underneath.
Whichever edition you choose, the architecture underneath is the same. Six things every edition ships.
One open standard contract format. Extended via documented kirimana.* customProperties. Same artefact runs on every edition.
Every LLM call (Anthropic, Azure OpenAI, Bedrock, Ollama, Databricks AI Assistants) is gated by per-contract classification. Restricted data never reaches the model.
Databricks, Fabric, Trino, DuckDB, Postgres — same contract, different runtime. Adapter pattern is ~400 lines per platform.
Compliance reports generate from contract metadata + audit log. The work was done all along; the report is the proof.
Unity Catalog, Microsoft Purview, Snowflake Horizon, Apache Polaris — we feed your catalog, we don't replace it.
GitHub-backed marketplace for contracts. Submit a pack via PR; install from any edition. Patterns travel across organisations.