Open-source AI application platform

AI Agents are probabilistic.
Their authority should not be.

Symbia is a standalone platform for AI applications that act on other systems. It gives every AI Agent zero-trust identity, explicit capabilities, signed application definitions, and a cryptographic chain of custody for deployed activity.

Test the model Run Symbia
Open source · local-first · built for external effects

The product boundary

Three properties of an accountable AI application.

Symbia defines the actor, the application, and the account of what happened when the application acts.

01

Governed AI Agents

Each AI Agent is independently authenticated and operates through explicit capabilities—not a shared application credential.

  • Zero-trust identity
  • Explicit authority
  • Delegated coordination
02

Governed applications

Catalog makes the application itself reviewable: components, graphs, executors, integrations, artifacts, and approvals.

  • Signed definitions
  • Versioned objects
  • Declared contracts
03

Chain of custody

Every deployed activity carries its authority, application version, execution path, observed outcome, and disclosed gaps.

  • Tamper-evident lineage
  • Observed, not invented
  • Integrity can be checked

The deterministic boundary

Let the model reason. Make the system answerable.

LLM output is probabilistic. Symbia makes identity, permissions, application versions, execution contracts, routing, and activity capture explicit. It does not certify a decision as correct; it preserves the material needed to inspect the action.

Read the technical model →
Model
probabilistic intent + plan
↓ bound to a signed application and capability
System
identity → contract → external action
↓ sealed with source, outcome, and gaps
Record
activity receipt + chain of custody

Start where it is useful

See it. Connect it. Run it.

◉

In-browser model

Explorer

Test the deterministic boundary and compare it with an LLM-only answer.

→ ⌘

MCP and clients

Integrations

Connect Claude or another MCP client to a Symbia application you operate.

→ ▦

Open source

Stack

Run the platform that defines, executes, and records your AI application.

→

A necessary constraint

A sealed record is not a model verdict.

Cryptographic evidence can establish what the system received, allowed, executed, and observed. It cannot make an uncertain model decision correct. Symbia keeps that distinction intact so the next person or system can inspect, recompute, or require review.

Read the determinism experiment →