From intent to execution
A bounded agent loop connects reasoning to typed tools, observations, approvals, and durable task traces.
AVAILABLE FOUNDATION INDEPENDENT INTELLIGENCE. OWNED BY YOU.
Intelligence,
under your
control.
The Autonomous AI That Gets Things Done.
Your models. Your infrastructure. Your rules.
An independent agent system built to turn intent into action—with you in command.
DRAKULIS CORE 01
REASONAny capable model
ACTPermissioned tools
VERIFYEvidence over claims
LOCAL SYSTEM / HUMAN CONTROLLED
A SYSTEM. NOT JUST A CHAT WINDOW.
Reason → Act → Observe → Verify
Built in the open · Development preview01 / THE PRINCIPLE
The model is a component.
DRAKULIS is the system.
Intelligence needs more than a good answer. It needs tools, memory, boundaries, and a way to tell whether the work actually got done.
Evidence before success
Explicit success criteria can require a real calculator result or a source that was read and cited. Model confidence is not verification.
INITIAL VERIFICATION HOOKSControl that stays yours
Read permissions are granted per call. Cancel a task, inspect its trace, or keep every component on your own machine.
AVAILABLE NOW02 / BUILT AROUND YOUR BOUNDARIES
Power, without
giving up ownership.
Start small. Run privately. Choose the intelligence that fits your work—not someone else’s infrastructure.
Pull the plug.
Keep the intelligence.
Once dependencies and a model are provisioned, local inference, memory, tools, and verification can run without a proprietary cloud service.
$ drakulis "Calculate (128+64)*3"
↳ permissioned tool execution
↳ persisted task trace
Illustrative command · requires a local modelActual network-disabled inference smoke passed.
Model tool-selection reliability is still being evaluated.
Model tool-selection reliability is still being evaluated.
Your server.
Your command center.
A prepared Compose stack separates the dashboard, control plane, database, and agent workers. Move execution to another machine through configuration.
STUDIO↔CONTROL PLANE↔WORKERS
DEPLOYMENT PREPARATION · NOT YET A PUBLIC SERVICEChoose a model.
Keep your system.
Local Ollama and OpenAI-compatible endpoints are supported. Capability-aware interfaces and optional cloud codecs keep provider details outside the execution loop.
* Provider-dependent capabilities. New cloud codecs are fixture-tested; live qualification is pending. No automatic cloud routing.
03 / REAL WORK, HONEST PROGRESS
Build toward work
that matters.
Useful workflows come first. We distinguish what you can inspect today from what still needs to be built.
01
Private knowledge assistance
EARLY WORKING SLICEUpload a small text document, approve a workspace read, and follow the result and its source through a persistent trace.
Semantic indexing, broad document formats, and robust claim verification are planned. Small-model answer quality varies.02
AI software engineering
IN DEVELOPMENTRepository inspection is the foundation. Approved code changes, isolated testing, and independently verified change reports are the next steps.
No arbitrary shell or code-execution tool is currently exposed.03
Durable workflow automation
RUNTIME FOUNDATIONQueue work, follow approvals, cancel execution, and recover read-only tasks after a worker crash using leases and fencing.
Scheduling, reusable workflow authoring, and broader integrations are planned.04 / ONE COHERENT SYSTEM
Separate the intelligence.
Connect the evidence.
Read the architecture ↗INTERFACEStudio + CLI
↔Your intent. Your controls.
CONTROL PLANEAPI + durable state
↔Tasks, permissions, history.
EXECUTION PLANEIndependent workers
Models, tools, verification.
SQLite or PostgreSQLAuthenticated worker contractRenewable leases + fencingExplicit artifact transport
Single-machine today. Separate workers when you need them. No mandatory hosted account.
05 / MAKE IT YOURS
Start on your
own terms.
Bring Python 3.12+, Node 22.12+, and a local Ollama model. Provision dependencies while connected, then keep the core local.
Installation guide ↗Development preview. No public hosted dashboard or download release is being advertised.
python3 -m venv .venv
.venv/bin/pip install -e '.[dev]'
npm --prefix apps/studio ci
ollama pull qwen3:0.6b
# With Ollama running locally:
DRAKULIS_MODEL=qwen3:0.6b \
.venv/bin/python scripts/local_stack.pyOpen localhost:5173 · use your private operator token
06 / BUILT WITH INTENT
A roadmap.
Not a list of promises.
No invented release dates. Progress is measured in working workflows, evidence, and limitations we can explain.
FOUNDATION · AVAILABLE
A working local core
Studio, CLI, persistent history, read approvals, bounded execution, independent workers, and offline smoke tests.
CURRENT FOCUS
Verified intelligence
Task success criteria, model capabilities, provider qualification, secure launch preparation, and reproducible evaluation.
PLANNED
Deeper autonomy
Isolated coding workflows, semantic memory, approved skills, scheduled work, and evaluated multi-agent orchestration.
INDEPENDENT BY DESIGN. BETTER THROUGH SCRUTINY.
Help build a system
you can trust.
Inspect the architecture. Reproduce a workflow. Challenge a claim.
Contributions start with useful evidence.
Public repository and release links will be added when available.
No license grant or community channel is implied.