Model-neutral
Cloud or local intelligence can change without moving the system's authority boundary.
YANA AI · CONTROL PLANE
Intelligence may change.
Authority must endure.
Yana AI is neither a model nor a chat window. It is the orchestration layer between intelligence, context and execution surfaces, so every action has boundaries and evidence.
01 / MODELIntelligence sourceModels can change
02 / AUTHORITYAuthority gatePeople decide
03 / RECEIPTEvidenceInspectable result
01 — THE BOUNDARY
A proposal is not execution.
A model creates a proposal. Yana checks authority, grants a bounded capability, executes through the appropriate surface, then writes a receipt. When policy requires it, a human approves the decision.
Human-governed
Final authority is not conflated with a model's ability to reason.
Inspectable
Every action connects to a decision, a capability and an inspectable receipt.
02 · ONE TURN OF WORK
Four steps, none inside the model.
The model only proposes an action, such as writing a file or running a command. That proposal changes nothing. YanaAuthorityChain checks it first, deterministically: allow, ask a person, or deny.
Only an allowed capability executes within its granted scope. The result becomes a receipt linked to the decision that permitted it. Studio and Runtime CLI follow the same path.
There is no gray area between proposal and action.
Model proposesProposal
Authority chain decidesallow · ask · deny
Capability executes within scopeExecution
Result is recordedReceipt
03 · INSIDE THE SYSTEM
Six layers, each with a job.
Yana separates intelligence, authority, execution and evidence into independent layers connected by clear contracts. Changing one layer need not move the boundaries of the others.
Where people work
Studio, Terminal, Web and Discord are interaction surfaces. Studio and Runtime CLI share a control plane.
Oversight outside the AI session
A supervisor runs independently at the operating-system layer. On detecting an unapproved change, it can HALT the entire session. Claude, Codex and Cursor hooks respect HALT; they do not grant themselves that authority.
Policy, tasks and memory
Policy, identity, capability and evidence live here. The task graph, missions, L1/L2 memory and project memory keep work alive beyond one chat.
Typed, cancellable turns
yana-rt handles typed turns, streaming, mid-run cancellation and the tool-call loop. Studio launches the same yana-rt binary used by the CLI, rather than a separate implementation.
Intelligence and scoped tools
Local and cloud providers use one authority model. Capabilities fall into access (file reads and git inspection), mutation (file writes, configuration and commits), and execution (shell commands, processes and system actions).
Receipts and audit trail
Each tool call joins a hash-chained audit record, and execution results carry an HMAC signature through Truth Gate. A history edit leaves a detectable trace.
04 · AUTHORITY IS ARCHITECTURE
Scoped, time-limited, revocable.
Policy lives outside the model. Even an emergency stop has a procedure: lifting HALT requires a recorded person and reason first, and no agent can unlock itself.
Every action has a formal name
Each action has its own risk level and approval rule. For example, file.write changes state and requires approval.
Authority by subject, scope and time
A grant applies to the right session and directory, for a limited period, and can be revoked at any time.
Missing governance must be visible
Yana does not silently skip authority checks. Denials are explicit, and HALT remains the highest priority.
05 · MODEL-NEUTRAL
Change the model. Keep the rules.
Yana supports Anthropic, OpenAI, Gemini, Groq, DeepSeek, OpenRouter and local Ollama. Choose a provider and Yana selects that provider's default model; you need not look up the catalog yourself.
The model changes, while the layers to the right remain. That is why Yana is a control plane, not just a wrapper around one model.
Intelligence may change. Authority must endure.
Missionunchanged
Tasksunchanged
Project memoryunchanged
Authority chainunchanged
Evidenceunchanged
06 · HONEST LIMITS
Yana makes unfinished work visible.
This page only claims what has supporting evidence. Yana's audit log is tamper-detectable, not tamper-proof. These areas remain incomplete.
Continuity Engine
The task graph, missions and L1/L2 memory are running. A complete Continuity Engine remains on the roadmap.
Wheelbot
Wheelbot has a separate WebSocket/MCP device bridge. There is no claim yet that this path has been verified end to end through the same TurnEngine.
MCP
MCP is optional and experimental; it does not create approval by itself.
Provider gateway
The newer provider gateway is not yet the dispatch path for every request.
Continuous approval
Only the Terminal chat path currently has approval continuation for state-changing capabilities. Packaged Studio and Web run headless and do not expose mutation capabilities yet.
External adapters
Harness adapters enforce at different strengths. Yana does not assume every harness is equally strict.
07 · TERMINAL TOOLS
You don't need Studio to use Yana.
Yana AI has two independent CLIs, each installed through its language's standard package manager — neither touches Yana Studio.
PIP · REPOSITORY GOVERNANCE
yana-ai — apply the governance layer to a project
Installs the Python CLI yana-ai, used to wire rules, hooks, agents, skills and commands into one specific project and check its integrity — everything stays scoped to that project, not your machine or other projects.
pip install yana-ai
cd your-project
yana-ai install # wires the hooks, ~60 seconds
yana-ai doctor . # integrity checkAfter yana-ai install, your AI agent (Claude Code, Cursor, Codex, Antigravity...) gets intercepted if it tries a dangerous command — hard-blocked on Claude Code/Cursor, advisory on Codex/Antigravity — and logged, entirely locally with no LLM call.
CARGO · RUNTIME & ORCHESTRATION
yana-rt — the canonical TurnEngine and terminal workspace
Installs the Rust CLI yana-rt (the same binary Yana Studio runs internally) — both a terminal workspace for running local models and a task-orchestration tool with direct command-line access to evidence, capability and memory, no Studio window needed.
cargo install yana-rt
yana-ai-rt --provider ollama # terminal workspace, local model
yana-rt route classify "fix auth"
yana-rt mission create "add-auth"The terminal workspace has streaming, mid-run cancellation, multiple tabs, multiple sessions, and live model switching. route/mission classify and create dependency-aware tasks through the same runtime Studio uses.
08 · KEEP EXPLORING
Want to see every layer at work?
Architecture shows the full stack, Runtime follows one turn of work, and Governance explains how authority is retained.