Build & ask online, governed by the runtime
Build & ask online.
Keep the data you own.
Describe what your business needs to the built-in AI Builder — a helpdesk, an approval flow, a CRM — and it is running: generated Console, REST APIs, SSO, permissions, and audit logs. Underneath, the whole app is compact ObjectStack metadata — a complete CRM is under 2,000 lines, roughly 16k tokens — small enough to review, governed on every call. Prefer your own coding agent? Build & ask with Claude Code on the open-source ObjectStack.
- AI Builder
- Describe it — it’s running
- Reviewable by design
- A whole CRM is ~16k tokens
- Your data, never ours
- Our cloud or your servers
For AI-written enterprise software
Keep the systems that work.
Add a governed runtime for agents.
Enterprise AI does not need another rebuild project or another pile of generated code. It needs a compact target format agents can write, humans can review, and a runtime that keeps each change governed across legacy systems, new applications, and AI agents. Your objects, permissions, and flows are your business ontology — open files you own, not an asset locked inside someone else’s platform.
Platform capabilities
Start with the business model,
not a blank codebase
- 01
Give agents a business model
Model customers, orders, equipment, cases, and approvals as objects agents can read, relate, and act on.
- 02
Extend systems without replacing them
Add APIs, permissions, workflows, and intelligence on top of databases, ERP, CRM, and custom systems.
- 03
Generate metadata, not app code
For typical CRUD and workflow software, agents write the compact ObjectStack definition while ObjectOS supplies tables, APIs, UI, tools, permissions, and audit. Less code to generate, less code to review.
- 04
Enforce governance at runtime
Reuse enterprise identity, permissions, approval queues, and audit logs so every agent action has a defined boundary.
AI build and agent operations
Let agents create the software.
Keep people in the review loop.
ObjectOS turns objects, fields, workflows, permissions, and actions into declarative metadata that agents can read and update through governed tools. On the open-source ObjectStack, you bring your own AI: a coding agent writes metadata as source files — the bundled CRM is 1,792 lines across 31 files, so the whole business system fits in the agent's context — you review the diff, and any MCP client can query your data. The runtime supplies everything repetitive underneath: tables, API, UI, permission checks, audit. The in-app Build and Ask assistants run on Cloud and Enterprise.
View the AI security model →AI Builder
Cloud & Enterprise: describe a change in natural language. The in-app Builder generates objects, fields, views, and workflows, then routes structural changes for approval. On the open-source ObjectStack, your coding agent writes the same compact metadata diff instead of a full app codebase.
AI Ask
Cloud & Enterprise: ask questions inside the product, analyze business context, and trigger approved actions within the signed-in user’s permissions. On the open-source ObjectStack, query the same objects through MCP with your own AI.
Tools / MCP
All editions: @objectstack/mcp exposes objects, queries, and actions as policy-aware tools for Claude, Cursor, any MCP client, or a local model.
How it works
Turn business operations into
a structure agents can use
ObjectOS describes objects, relationships, permissions, workflows, and actions in unified metadata. Agents change a compact definition layer instead of regenerating application code, so business iterations stay fast, reviewable, and governed.
Security and governance
Keep data in your network.
Let AI work inside permissions.
ObjectOS runs as a self-hosted runtime on your infrastructure. Business records, identities, audit logs, and files stay under your control; AI agents access objects through governed tools and inherit the signed-in user’s permissions.
Explore security and governance →Data residency
Connect your databases and storage. Unless you configure an external service, ObjectOS does not send telemetry, contact a license server, or transmit data back to ObjectStack.
User-scoped AI
Agents act as signed-in users and obey object, record, and field permissions, so they cannot see data the user cannot see.
Approval and audit
Structural changes go through a human approval queue. Reads, writes, tool calls, and permission changes can be written to audit logs.
Offline ready
Run in a VPC, on local servers, or in air-gapped networks with local models, internal identity, and your own secrets management.
Application templates
Start with working templates,
not a blank canvas
Helpdesk template
An AI-first customer support template for tickets, SLA, summaries, suggested replies, and knowledge retrieval.
View template source →Contracts template
Manage the contract lifecycle with metadata extraction, approval, renewal reminders, and audit trails.
View template source →Procurement template
Run purchase requests, suppliers, POs, receiving, and three-way matching as a governed application.
View template source →How it compares
Different from
the tools you know
vs Airtable
A real database with server-side logic and runtime governance — not a spreadsheet-style workspace.
Read the comparison →vs Retool
Business logic is reviewable metadata — not JavaScript scattered across screens.
Read the comparison →vs Lovable & Bolt
Agents generate governed metadata with schema and permissions — not a one-off codebase.
Read the comparison →Latest insights
Practical thinking on AI-native software
What Tools Do Forward-Deployed Engineers Use? An Ontology-First Open Stack
Five pains define forward-deployed work: plumbing eats week one, demos die in security review, requirements outrun code, patterns never compound, and the handover poisons trust. An ontology-first open stack removes each one.
How Many Tokens Is a Business App? A Complete CRM in 16k
Measured in the unit AI works in, a complete CRM is ~16k tokens of typed metadata — 8% of one context window. Software that fits whole in an agent's context is maintained differently. We call it context-sized software.
When an AI Agent Deletes Production Data: Runtime Guardrails Beat Prompts
The Replit database incident shows a structural lesson: an agent's blast radius must be controlled by runtime permissions, approvals, and audit logs, not only by a prompt.
Next step
Start with the business data you know best.
Connect one existing system, define its key business objects, and let your agent ship the first governed AI-written application as a small metadata diff.
Learn how to connect existing systems →