An open-source .NET framework for durable agents on Orleans and Aspire. Its paradigm is neurons and
synapses: neurons are durable Orleans-journaled agents, synapses are typed facts with full lineage,
and method-scoped TestBrain fixtures fire real multi-silo traffic and assert on committed journals.
A brain you program by writing ordinary C#, and that can program itself.
// production: builder.AddDigitalBrainClient(owner); inject IDigitalBrain
await brain.SendAsync<IAnalyst>(
"incident-42",
new SummaryRequested("Summarize the incident."));The owner-bound IDigitalBrain facade enters through a session; neurons call typed capabilities such
as IGemma4 inside the brain. The same vocabulary will later carry approved C# behaviors generated
from natural language.
- A synapse is a fact — a thin record, broadcast, no reply. An interface method is a request — directed at a capability, and it replies. Both are journaled; neither is privileged.
- Modules own vocabulary — synapse records and neuron interfaces, resolved at compile time.
- Namespaces and type names are architecture —
DigitalBrain.AI.Ollama.IGemma4is identity, not a lookup result. - Journals are the audit source, recording causal facts only — never arguments, prompts or secrets. Telemetry is a projection and never replaces them.
- Every install is a human-approved proposal, journaled and reversible.
The plan of record. Nothing is shipped unless it says Built.
| Area | State |
|---|---|
| Neuron/synapse foundation, owner-scoped client, module activation, AppHost composition, testing path | Built |
| Typed AI, Tasks, Google, Salesforce, Chat, Flutter, Quickstart families | Built |
Flutter vertical — shell/scene vocabulary, UI HTTP/SSE edge (WithUiEdge/WithHeadlessHost/WithWindowHost), headless Dart host, Windows chrome |
Built |
| Product shell — responsive Chat, content-safe Activity, live 3D-projected Brain topology, pulses and inspector | Built |
| Product MCP surface — durable chat send/read, neuron journal observation, active-neuron discovery | Built |
Time — durable one-shot ICountdown and its recovery tests |
Built |
| Time — reminders, recurring interval/calendar scheduling, DST | Designed |
Multi-principal IdP edge, journal observation on IDigitalBrain |
Designed |
| Behavior rail — proposal, compile, BDD gate, approval, activation, rollback, signed BehaviorHost deploy/execute (L1/L2) | Built |
| Observability spine — host OpenTelemetry, structured logs, causal kernel spans, GenAI spans and metrics | Built |
DigitalBrain.Behaviors is a packable SDK foundation (authoring interfaces, constrained context,
manifests, artifact identities) and holds the canonical artifact codec. It is not a compiler, builder,
broker or execution rail. Nothing in the product runs on it today: chat owns its whole turn in
ChatNeuron, and no shipped neuron implements IIntentProgram.
One assumption is load-bearing and unmeasured: that a model can reliably emit behaviour scripts.
src/ published packages: core/ (framework) and modules/ (IModule domains),
plus the publish gate that polices them
os/ the product: silo, northbound MCP host (OS.McpHost), Flutter UI edge
(OS.Ui), assistant neuron, AppHost
clients/ flutter/core (pure Dart edge) and flutter/shell (Material chrome)
samples/ product-shaped compositions and process neurons (not packable product)
tests/ fixtures/apphosts — shared L2 AppHost scaffolding
Southbound MCP transport lives in package DigitalBrain.Mcp (integrations mechanics shared by
Google/Salesforce). Northbound agent tools live in DigitalBrain.OS.McpHost. Older docs that say
DigitalBrain.Integrations.Mcp mean the southbound package.
Retired prototype generations live in git history — git log --diff-filter=D --summary, then
git show <sha>^:<path>.
git clean -fdx
aspire runThe explicit product suite performs a self-cleaning live Aspire proof across resource health, Gemma4 chat, command retry, durable journals, owner-scoped introspection, and OpenTelemetry:
dotnet test os/tests/DigitalBrain.OS.Product.Tests -c Release -- -explicit onlyCLAUDE.md is the working discipline for every agent and contributor: the gates, the verification ladder, and the traps. A green test suite is necessary, not sufficient — it proves the code holds, not that a behaviour works.
The public site lives in intochat/digitalbrain.docs and publishes https://digitalbrain.tech.