SystemDULUS v3.10.19
Runtimepip install dulus · GPLv3
ExecutionMOD · Mesa Redonda
Clouddulus.online
dulus.ai

An open-source AI agent you can read, fork, and run on any model — frontier cloud or free local weights, no API key required to start. A Mesa Redonda where many models debate until the answer holds. And when one agent isn't enough, dulus.online flies whole fleets in the cloud.

01 The brief

Dulus is an open-source AI agent born in Santo Domingo — named after the Cigua Palmera, the national bird of the Dominican Republic. One pip install gives you the whole loop: REPL, tools, sub-agents, voice, memory, and a Mesa Redonda where many models debate until the answer holds. It runs frontier clouds or free local weights — no API key required to start. And when one agent is not enough, dulus.online deploys, coordinates, and governs whole fleets in the cloud, with the $DULUS token as fuel. Open code. Real work. Zero gatekeeping.

System
DULUS v3.10.19 · open source · GPLv3
Cloud
dulus.online · synthetic operations
Origin
Santo Domingo · Cigua Palmera
Native providers
26
Via LiteLLM
100+
Built-in tools
30+
Tests passing
263+
Languages
34
License
GPLv3
02 6-Plane architecture stack

Six planes.
One unbreakable path.

The architecture is organized as six stacked planes — Edge, Identity, Control, Execution, Memory, Observability — each isolating a concern and enforcing guarantees independently. Every plane carries its own technology stack, its own failure mode, and its own SLO. No plane trusts the plane above it.

P01
Plane 01 · outermost Edge
Cloudflare DNS CDN WAF
DNS resolution, content delivery, edge compute, and web application firewalling. First contact for every request.
P02
Plane 02 Identity
Auth0 RBAC M2M Auth Multi-tenant
Role-based access control, machine-to-machine authentication, multi-tenant identity isolation. Every actor verified, every action attributed.
P03
Plane 03 Control
Dulus Hub Command Router State Machine Approval Gates
The Dulus Hub orchestrates commands across distributed nodes. Mission graphs lock in state. Human approval gates at every critical transition.
P04
Plane 04 Execution
MOD / Mesa Redonda Model Router Swarm Agents Verification
Where the Mesa Redonda runs. Multiple LLMs debate, correct, and verify in parallel. Swarm execution with independent quality gates before any artifact ships.
P05
Plane 05 Memory
Canonical Data Model Event Envelope Evidence Store Audit Ledger
Every event wrapped in a canonical envelope. Artifacts, evidence, attempts, and tool calls persisted immutably. The system remembers everything it has done.
P06
Plane 06 · innermost Observability
Unified Telemetry Analytics SLO Monitoring Rollback Metadata
Unified observability across all planes. SLOs measured continuously. 100% rollback metadata captured. If it cannot be measured, it does not ship.
03 Autonomy maturation curve

L0 to L5.
From answer to intelligence.

Six levels of autonomy, each a strict superset of the previous. The open-source CLI ships L3 — Verified Delivery today; L4 — Operated System is what dulus.online fleets are built for. L5 — Network Intelligence — is the horizon.

Level 00
L0
Answer
Produces text or artifacts in response to a prompt. No execution, no verification, no persistence.
Status surpassed
Level 01
L1
Tool User
Executes a single bounded action. Calls a tool, returns a result. No multi-step planning.
Status surpassed
Level 02
L2
Multi-Step
Plans and executes a sequence of tool calls. No independent verification of outcomes.
Status surpassed
Level 03 · current
L3
Verified Delivery
Tests its own artifacts. Quality gates fire. Evidence is captured. Output is verified before delivery.
Status shipping today
Level 04
L4
Operated System
Deploys, observes, and repairs. The system runs in production and maintains itself within bounded authority.
Where dulus.online fleets
Level 05 · horizon
L5
Network Intelligence
Marketplace of nodes. Specialized agents negotiate, delegate, and compose across organizational boundaries.
Status the horizon
Autonomy →
L0 · Answer L1 · Tool User L2 · Multi-Step L3 · Verified Delivery L4 · Operated System L5 · Network Intelligence
04 The organizational engine

Ten modules.
Twelve steps.

The studio runs as ten specialized modules — each with its own domain, its own authority, its own accountability. Work flows through a Mission Execution Protocol of twelve sequential steps, from receipt to release. No step is skipped. No step is parallelized outside its gate.

01
Command
Mission intake, prioritization, resource allocation.
02
Product
Requirements, roadmap, user-facing surface definition.
03
Architecture
System design, plane boundaries, technology selection.
04
Platform
Infrastructure, deployment, runtime environment.
05
Security
Access control, isolation, threat modeling, audit.
06
Quality
Testing, verification, evidence capture, gate enforcement.
07
Reliability
SLO definition, monitoring, incident response, rollback.
08
Integrations
External systems, vendor management, API contracts.
09
Growth
Acquisition, conversion, pricing, market expansion.
10
Customer
Onboarding, support, feedback loop, retention.
Mission Execution Protocol · 12 Steps
01
Receive
02
Plan
03
Assign
04
Build
05
Integrate
06
Verify
07
Challenge
08
Approve
09
Release
10
Learn
05 The golden path pipeline

Three pipelines.
One unbreakable path.

Every mission traverses three sequential pipelines: Intake & Orchestration locks in intent and state; Swarm & Verification builds, verifies, and gates; Deployment & Telemetry ships, observes, and learns. Each node carries an SLO. Each SLO is measured.

Golden Path · Pipeline 01 — Intake & Orchestration
5 nodes · 2 gates
01
Customer Intent
Human submits a request — natural language, structured brief, or API call. The entry point for every mission.
No SLO · entry
02
Persist Request ID
Request is assigned a durable identifier and written to the canonical event store. State begins.
99.9% initial / 99.99% mature
SLO · Request Persisted
03
Scope + Estimate
System scopes the mission, estimates resource cost, and proposes a plan. Human reviews before commitment.
< 5 min mission plan time
04
Human Approval
Runtime gate. Human approves, rejects, or modifies the plan. No mission proceeds without explicit sign-off.
100% human-approved
Gate · Approval runtime gate
05
Mission Graph
Plan is locked into a mission graph — a state machine that cannot be altered mid-flight. Durable. Auditable.
100% durable task state
Lock · State machine lock-in
Golden Path · Pipeline 02 — Swarm & Verification
3 nodes · quality gate
06
Swarm Execution
MOD — Mesa Redonda dispatches the mission to specialized agents. Multiple models collaborate, debate, and correct in parallel.
> 95% mission recovery
07
Independent Verify
A separate verification agent — not the one that built — tests the artifact. Independence is structural, not procedural.
100% approval evidence
08
Quality Gate
If verification fails, the mission returns to swarm execution with correction context. If it passes, deployment is authorized.
100% gate evidence captured
Gate · Quality gate — hard stop
Golden Path · Pipeline 03 — Deployment & Telemetry
3 nodes · learn loop
09
Deploy Approval
Final runtime gate. Human or automated approver authorizes production deployment based on captured evidence.
100% deploy evidence
10
Production Deploy
Controlled rollout to production. Rollback metadata captured for every deployment — no exceptions.
100% rollback metadata
11
Telemetry + Learn
Production telemetry feeds back into the system. The mission is archived. Lessons are extracted and applied to the next mission.
Continuous · learn loop
06 Model router & MOD — Mesa Redonda

Any model.
One round table.

The MOD — Mixture of Dulus — is shipping today in the open-source CLI. In a Mesa Redonda, several models take a seat, answer independently, then debate — defending or conceding — until a moderator synthesizes the verdict. Twenty-six native providers, over a hundred more through LiteLLM, and free local weights can all sit at the table.

Core
Provider Router
Routes every call across 26 native providers — plus 100+ through the LiteLLM gateway. Auto-fallback on rate limits and quota errors. Local models mean a $0 floor: the agent always has somewhere to think.
MOD · Mesa Redonda active
Anthropic
cloud · claude
OpenAI
cloud · gpt
Google Gemini
cloud · gemini
DeepSeek
cloud · deepseek
Moonshot Kimi
cloud · kimi
Alibaba Qwen
cloud · qwen
Zhipu GLM
cloud · glm
xAI Grok
cloud · grok
Azure OpenAI
cloud · enterprise
Ollama
local · free weights
LM Studio
local · free weights
LiteLLM Gateway
+100 providers
Rule 01
Any mix of seats — frontier cloud models or free local weights, same table.
Rule 02
Every seat answers independently before the debate. No copying homework.
Rule 03
Debate rounds: models read each other, defend or concede, a moderator synthesizes.
Rule 04
Auto-fallback across providers on quota or rate errors — the mission continues.
07 Canonical data model

Organization
to tool call.

A single hierarchical data model governs every entity in the system — from Organization down to individual ToolCall. Every state change is wrapped in a canonical event envelope with a unique ID, version, and name. The model is the contract.

Entity Hierarchy
Organization
tenant root
Workspace
team boundary
Project
product surface
Mission
unit of intent
·
Task
executable step
·
Artifact
produced output
·
Evidence
verification proof
·
Attempt
execution record
·
ToolCall
model invocation
Canonical Event Envelope
Every state change, wrapped.
event_id
UUID v7 · string
event_version
semantic · string
event_name
dotted.path · string
timestamp
UTC iso-8601 · datetime
actor
agent or human · ref
mission_id
parent mission · ref
payload
typed body · json
evidence_hash
sha-256 · string
08 Security & governance — the moat

Three pillars.
Bare-metal + cloud.

Security is structural, not procedural. Three pillars — Access Control, Execution Safety, System Integrity — enforced across a dual infrastructure: bare-metal core nodes for sensitive workloads, cloud backbone for elastic scale. The two are bridged by an encrypted private fabric.

Pillar 01
Access Control
  • RBAC with principle of least privilege enforced at the gateway
  • Multi-tenant Auth0 isolation — no cross-tenant data access
  • M2M authentication on every service-to-service call
  • Secrets rotated automatically; no long-lived credentials
  • Audit log on every access decision, retained 18 months
Pillar 02
Execution Safety
  • Sandboxed execution environments for every agent
  • Resource quotas — CPU, memory, network, time — per mission
  • Human approval gates at every state transition
  • Rollback metadata on 100% of deployments, no exceptions
  • Challenge step in protocol — independent agent tests every artifact
Pillar 03
System Integrity
  • Immutable evidence store — cryptographic hashes on every event
  • Canonical event envelope — versioned, typed, auditable
  • State machine lock-in — mission graphs cannot be altered mid-flight
  • Independent verification agent — not the builder, always separate
  • 100% approval evidence captured before any release
Layer 01 · bare-metal core
Owned hardware
AMD Beast Node
192-core · 512GB RAM
Founder Workstation
local inference · dev
NVIDIA NIM cluster
on-prem GPU serving
Ollama self-hosted
open-weight models
Layer 02 · cloud backbone
Elastic scale
Cloudflare
edge · WAF · CDN
Azure AI Foundry
managed model serving
AWS
compute · storage · lambda
Auth0
identity · multi-tenant
09 The Dulus constellation

One bird.
Three domains.

Dulus is not one website — it's a constellation. dulus.ai is the front door you're standing in. dulus.online is where fleets fly: deploy, coordinate and govern AI workers in the cloud. GitHub is where the agent itself lives, open source, forever. The $DULUS token fuels all of it.

Node 01 · You are here
dulus.ai
The Front Door
The product surface. Talk to Dulus in the webchat, run your business from the dashboard, embed the widget on your own site. One Auth0 account works across the whole constellation.
  • AI webchat — ask Dulus anything, free, in your language
  • Business dashboard — CRM, orders, moderation, automations
  • One sign-in for dulus.ai and dulus.online alike
  • Embeddable widget for any site
Node 02 · The cloud
Synthetic Operations
Deploy, coordinate and govern fleets of AI workers across any model, tool and runtime. Free sandbox to prove the loop; Operator and Company plans when you're ready to scale.
  • Free sandbox — 1 workspace, 1 managed agent, starter credits
  • Operator $29/mo — multiple agents, MOD missions, Memory Palace
  • Company $79/mo — roles, approvals, shared memory, analytics
  • Pay in $DULUS and take 20% off any plan
Node 03 · The source
The Open Agent
The whole agent — REPL, tool dispatch, streaming, sub-agents, voice, Telegram bridge, MCP, plugins — in Python you can actually read and fork. GPLv3. No build step. No gatekeeping.
  • pip install dulus — free local models, no API key needed
  • 26 native providers + 100 more via LiteLLM
  • 263+ tests · 30+ tools · 34 languages
  • MemPalace semantic memory · embedded sandbox OS
Node 04 · The fuel
$DULUS · Solana
Token as Fuel
The $DULUS token on Solana is how the constellation pays for itself. Hold it, spend it on plans at a discount, and back the long game of an open agent from Santo Domingo.
  • 20% off any dulus.online plan paid in $DULUS
  • Community on Telegram — t.me/dulusx
  • Built for the long game, not the flip
  • Follow the flight — x.com/KevRojo
10 Pricing · lives on dulus.online

Four tiers.
Zero gatekeeping.

The CLI is free forever — GPLv3, local models, no API key. Cloud plans live on dulus.online and every paid plan can be paid in $DULUS for 20% off. Same one-account sign-in across the constellation.

Sandbox
$0
prove the loop
  • 1 workspace
  • 1 managed agent
  • Starter mission credits
  • Community support
  • The open-source CLI, always free
Start sandbox
Company
$79/mo
govern team operations
  • Roles and approvals
  • Shared company memory
  • Fleet analytics
  • Audit history
  • 20% off paid in $DULUS
Build a company
Enterprise
Custom
own the runtime
  • SSO and SCIM
  • Private deployment
  • Custom retention
  • SLA and support
  • Bare-metal option
Talk to Dulus
11 Take flight

Own the agent.
Rent the fleet.

pip install dulus and the agent is yours forever — GPLv3, local models, no API key required. When you want managed fleets, shared memory and missions in the cloud, dulus.online has a free sandbox waiting.