Task
Prompt plus attached files enter one workspace.
08 / AI Laboratory · icuboid Local AI Studio
Live locally · v2.3.0
The operating environment is the product.
The chat surface is only the visible edge. Routing, memory pressure, tools, skills, terminals, artifacts and recovery determine what the system can actually do.
Studio is private by design: local models and shell-capable tools run inside a tailnet boundary with no public authentication.
01
One task inside the local frame
The path around the model
shapes the experience.
A useful local agent is an orchestration problem. The environment must choose capability, supply instructions, expose tools, preserve artifacts and recover when the small model stalls.
Model
Vision, coding and general heuristics consider what is already in memory.
Skill
Instruction sets shape how the model approaches the work.
Tool loop
Read, write and run stay visible inside the workspace.
Approval boundary
The full A1OS graduated-trust interceptor is planned, not shipped.
Artifact
Files and live previews become the durable result.
02
The operating environment
Capability lives
between components.
Studio v2.3 is more than its model catalog. The shipped value is how the parts work together under real local constraints.
Route around memory pressure.
Auto-routing considers task type and what is already loaded. Cold loads, unload controls and recovery states stay legible.
Give the model a method.
Sandboxed file tools, gated shell execution and SKILL.md instruction sets turn chat into repeatable work.
Keep the outcome visible.
Attachments bridge into the workspace; generated files become cards and live previews rather than disappearing in a transcript.
03
What v2.3 added
The laboratory became
an operations surface.
Browser terminals connect the local machines. Mobile layouts make Studio usable from a phone inside the private network. The Night Club tab manages a local simulation and can trigger the hosted environment. A device-control card proxies the round-display dashboard.
The Chat Playground now bridges attachments into the agent workspace, retries empty responses, shows routing decisions, auto-attaches prototyping guidance and lets the user stop generation without losing partial work.
These capabilities are shipped in v2.3.0. The full A1OS approval/timeline/handback port remains a separate pending session.
04
The honest limits
Local models change the design because they fail differently.
Small models are credible prototyping partners, not frontier-model equivalents. One coding model still returns empty output when tools are attached, so the router avoids it. The MCP client exists, but n8n is not verified as wired. Studio also has shell-capable surfaces and no public authentication; the private network is the security boundary, so there is intentionally no public link.
05
What carries forward
Build the system.
Design the boundary.
Studio closes the chapter by joining the themes from the other projects: agent-readable structure, bounded autonomy, persistent context, observable behavior and honest handback.
The environment is where those patterns can eventually be tested in real work rather than demonstrated in a standalone prototype.