omnibase

What it is in one line

Turning agents from "one-off chat personas" into digital employees with a job description, version, project context, and auditable run records — and the whole thing is self-hosted, so your data never leaves your server.

Who built it

GitHub user lss100200, an open-source project (Apache-2.0, Python/FastAPI + Next.js + pgvector), currently in Public Preview. The site omnibase.chat carries a strikingly honest long-form page: "we only show what is actually done and actually being built," and the production-grade multi-agent Runtime is explicitly kept gated.

Read: this is a typical build-in-public self-hosted infrastructure project. The most notable thing is not the features — it is that the site openly says "production Runtime is still off and multi-agent is still on the roadmap." That honesty is rare in open source, where the usual move is to overclaim a half-finished demo.

What it actually does

• Personal model gateway → configure your own OpenAI-compatible Provider, store API keys, and test the connection before and after saving
• Workspace isolation → one independent context per project: members, agents, knowledge, tasks, and run records are bound to the Workspace instead of scattered across chat history
• Agent lifecycle management → AgentDefinition (role/responsibility) → AgentVersion (manual and Skills) → WorkspaceBinding (appointment to a project) → AgentRun (one concrete work shift), with a draft → trial → sealed → appointed state machine
• Task/Run audit ledger → every real model call leaves a persistent record; rollback and audit are built in
• Self-hosted RAG → knowledge retrieval stays inside Workspace boundaries, with citation backlinks, and read-only knowledge_search is governed by tenant, budget, audit, and capability boundaries
• Role language → six roles are defined — Workspace Steward, Explorer, Builder, Verifier, Knowledge Curator, Operator — but today only a reliable single-agent loop is guaranteed

Current real capability boundary: personal Provider + connection test + Workspace

• Agent Builder + "tool-less single-agent model calls" + persistent Task/Run records + read-only knowledge retrieval. Tool execution (Typed Executor), the multi-agent Runtime, and the production Runtime are all gated off.

What old behavior it replaces

Getting agents into real business use previously meant two unsatisfying paths:

SaaS agent platforms (Dify, Coze and peers). Knowledge bases, model keys, and run records all live on someone else's servers. For data-sensitive companies, that is a stopper.

Rolling your own. Wire the model API, build a knowledge base, write agent scheduling, keep the books — every piece exists as a tool, but nothing puts "knowledge + RAG + model providers + agents + run records" into one maintainable workbench. The result is context scattered across chat history: an agent starts work with no idea where it left off, which project it works for, or what it cost.

OmniBase replaces both: a self-hosted base that keeps data local, plus a management model that promotes agents from chat personas to "digital employees with a role, a version, a project, and an audit trail."

Business model

Not disclosed. Apache-2.0, Public Preview, no pricing page.

Read: the money in self-hosted infrastructure has always been enterprise edition plus support — multi-tenancy, SSO, a stable production Runtime — and only after that is worth pricing. With the production Runtime still gated, there is no business model to discuss; the discussion is "prove one agent reliably closes the loop first."

Hard numbers

• 166 stars / 4 forks / 4 open issues. Apache-2.0, Python + Next.js + pgvector
• Public Preview; production Agent Runtime remains gated
• Roadmap as self-reported: done = tool-less single-agent real calls + Task/Run ledger
  • self-hosted RAG; building = Planner / Typed Executor / Capability Gateway / Desktop; later = multi-agent Runtime, Self-Development, Hardened Production Runtime
• Team size, enterprise users: not disclosed

Four-way read

The call

The positioning and the honesty are both worth watching, but the functionality is still very early. "Self-hosted AI workbench + digital-employee agent management" is a differentiated combination: others race on how many agents run; this races on whether one agent can be managed properly — role, version, appointment, audit. It is a framework for treating agents as staff rather than toys, and the direction is right.

Three things to note: First, current capability is far from a closed loop. The site itself writes "real tool-less single-agent model calls" — without tool calling, an agent cannot actually do anything, which is the smallest possible loop. Second, multi-agent and tool execution are both roadmap items, so reading this project now means reading a manifesto; the basis for judgment is "honest roadmap + open code." Third, the self-hosted market is not short of entrants. Dify already open-sourced self-hosting; model gateways, RAG, and workbenches all exist. OmniBase's differentiation ultimately depends on whether the management model goes deep — a genuinely auditable agent run ledger, version rollback, and so on.

Worth watching: 166 stars show some attention, but what is worth studying is the positioning and the honesty; the product itself needs tool execution and multi-agent before a revisit.

What to watch next

① Whether Typed Executor (tool calling) lands on the main branch within three months — a single agent without tools is not a closed loop ② When the production Agent Runtime unlocks, and whether a real deployment appears after it does ③ Star growth and the quality of community issues — self-hosted infrastructure lives or dies on community validation, and slow is death

What you can take from it

Product logic: any agent-platform product should borrow its lifecycle model — Definition (role) → Version (manual) → Binding (appointment) → Run (shift). Managing an agent like an employee (job description, versioned changes, project ownership, an audit record per shift) is a management dimension above "give it a chat window."

Pricing structure: none. Not disclosed.

Verdict

Worth watching. Differentiated positioning (agents as managed digital employees + self-hosting), an honest roadmap (production Runtime explicitly gated), and open, verifiable code. But today's capability stops at tool-less bare calls; judging whether it works requires at least tool execution and multi-agent to land. Come back in three months against the three checks above.