VARIANT-1

Work with local files, Python, and AI in one Windows workspace.

VARIANT-1 brings a live Python environment, connected tools, and supported local or cloud models into a desktop workspace for coding, research, and desktop work. Keep variables, imports, and helper functions available across calls within a live session, combine tools through Python, and inspect intermediate results.

The current preview is for technically comfortable users who want to inspect and adapt multi-step workflows. Windows is the primary source-development path; preview packages are also available for Linux and Apple Silicon macOS. It uses an Electron/React interface and a CPython backend; see Architecture for the technical design.

Website · Setup · Architecture · Issues

Status

VARIANT-1 is an early public preview, under active development. Unsigned Windows x64, Linux x86_64, and macOS arm64 packages are available in GitHub Releases, with checksums and release-specific installation instructions. Packaged launch checks do not certify every provider, model, hardware combination, or desktop workflow.

Current source can contain fixes newer than the latest published installer. Use the release notes for the behavior and qualification of a downloaded build.

What it provides

• Continue work within a live session. Reuse Python variables, imports, and helper functions across calls instead of recreating that working state for each step. Live state is not a guarantee of recovery after a restart. Overview > Python shows every live kernel with its current cell, CPU, memory, and actions, including retained sessions and explicit release controls; closing a generation ends its live objects and background Python work.
• Combine code and tools. Work with files, processes, browser/desktop operations, and connectors through Python, with inspectable results. The ASTB tool-discovery layer exposes capabilities as needed through shared backend services. Search by method name to find its owning capability, category, and call signature.
• Inspect execution and changes. Activity distinguishes failures, stopped work, and uncertain outcomes. Bounded previews redact managed or recognized credentials and mark omitted content. Review Git changes in collapsible file cards with line numbers and bounded expansion of unchanged context. Browse a branch's complete delta against its detected or chosen base, browse history across branches without checkout, or inspect individually selected commits. Branch totals describe net changes; selected-commit totals sum those commits' edits. The Changed files sidebar opens on the right when needed; source-line links are limited to live changes. Request commits, pushes, or pull requests through the agent.
• Choose supported local or cloud inference. Use local models, your own API keys or custom endpoints, or supported provider-account connections. Available features depend on the model, provider, and configuration.
• Stay involved as work runs. Steer or cancel execution, with explicit state and resource lifetimes. Stopping work does not undo actions already completed. Cancelling kernel startup retires that generation without retrying or running the cancelled request; a later explicit request can start normally.
• Adapt tools when needed. Optional chat-local tool authoring lets the model register its own helpers as tools or replace a failing tool method. What it registers stays until it or you roll it back. Authoring is off by default and does not train model weights.
• Run a Goal in its owning session. The session agent retains its conversation, model route, peer identity, and live Python state while working toward the objective. It can explicitly delegate children or coordinate other sessions. Goal completion records the agent's claim and the available evidence. Cancellation cleans up Goal-attributed managed resources while keeping the parent Python kernel. Use managed process handles for services that should stop with the Goal; raw Python subprocesses need their own lifecycle management. Admitted Goal and peer work appears live in that same chat, with session, admission and run identities keeping separate executions isolated. Addressed peer requests wake idle sessions or reach an active turn at its next safe step. Request handles expose the recipient's settled answer and outcome. Explicit Goal budgets pause at model/tool boundaries; defaults remain unlimited. Unavailable peer work blocks an awaiting Goal for inspection. Provider usage is durably recorded off the event loop before terminal budget reconciliation. Restart recovery surfaces consumed peer work without a settlement as uncertain, preserving its evidence for inspection. Delivered steering appears between the work before and after it. Opening an active chat restores its admitted task and delivered inputs before the final transcript is committed. Older evidence is accessible from the activity trace.

Costs and model access

The app requires no VARIANT subscription. You supply the inference: local models use your hardware, while cloud providers set their own costs, quotas, and account requirements. OAuth authorizes access; it does not promise free inference. OpenCode Zen and Go use API keys and an explicitly selected model. Their public catalogs do not certify inference access; the retired anonymous Free connection is no longer offered.

Model weights and native inference binaries are not stored in this repository. Google subscription OAuth also requires an authorized client configuration that this source preview does not bundle. See Setup for requirements and limitations before choosing a connection.

Run from source

Use Windows x64, Python 3.13, Node.js, and Git. Read the setup guide for runtime prerequisites and optional components.

git clone https://github.com/Nocturn3529/VARIANT-1.git
cd VARIANT-1
npm ci
npm run setup:backend
npm start

After launch, configure a supported provider or local runtime in the app. Provider Settings distinguishes saved accounts from recently checked connections. Use Check to refresh a service's status; model listing access does not certify every model's inference availability. Nous Portal can be reconnected in Settings through the installed Hermes account. Its sign-out control also signs Hermes out. Mutation authoring is under the composer's Session tools. Start with sample files or recoverable copies while learning how a workflow behaves.

Execution and data

This is not a sandbox. Python and desktop tools run with your operating-system permissions. They can change files and operate applications; choosing a project folder does not restrict them to that folder. Inspect important results and keep recoverable copies. Cancellation cannot reverse completed effects.

Cloud model requests send selected messages and tool results to the chosen provider. Using a desktop app does not make cloud inference local.

Live Python state can be lost on restart, reset, eviction, or failure. Optional portable checkpoints, durable chat records, and on-demand session context have separate contracts, and not every live object is restorable. Agents retrieve earlier session evidence through session.context(); the harness does not inject stored profiles or inferred fact memories, or extract memories after turns. Settings includes a Session context browser for frozen views, search, source details, and JSONL or Markdown exports. Goal controls remain separately available. Cross-session memory is deferred. Portable checkpointing and restoration are disabled by default. See state lifetimes.

Model routing Settings can configure ordered backup models and separate routes for internal JSON, summary, and vision calls. Recovery is disabled by default; saving routes does not change the selected main model.

Chat navigation offers recovery when an opening acknowledgement is lost. Running indicators and retained tool rows reconcile with the native runtime; an unobserved tool result remains explicit rather than being shown as success. Archived project chats remain accessible in Archived.

Development

See CONTRIBUTING.md, validation, and release management. Report ordinary bugs through Issues with redacted reproduction steps. For sensitive reports, use SECURITY.md.

The experimental endurance controller in experiments/live-canary/run_endurance.py creates a disposable lab with one Goal coordinator and independent cloud collaborators. Initialize once, then run the same lab to retain its sessions and evidence:

backend\.venv\Scripts\python.exe experiments\live-canary\run_endurance.py init --root "$env:USERPROFILE\Desktop\VARIANT-1-Endurance-Lab" --duration 7200
backend\.venv\Scripts\python.exe experiments\live-canary\run_endurance.py preflight --root "$env:USERPROFILE\Desktop\VARIANT-1-Endurance-Lab"
backend\.venv\Scripts\python.exe experiments\live-canary\run_endurance.py admit --root "$env:USERPROFILE\Desktop\VARIANT-1-Endurance-Lab"
backend\.venv\Scripts\python.exe experiments\live-canary\run_endurance.py run --root "$env:USERPROFILE\Desktop\VARIANT-1-Endurance-Lab"

Commit the tested source before initializing a pinned live phase. The controller requires a disposable OPENROUTER_API_KEY for OpenRouter routes, or a platform-encrypted credential file selected by VARIANT1_ENDURANCE_CREDENTIAL_FILE. The default lineup uses Space Bunny and Qwen through OpenRouter and LongCat through Hermes's signed-in Nous proxy. --routes accepts openrouter::MODEL and hermes::MODEL; existing bare model IDs retain their OpenRouter meaning. OAuth tokens are not copied into the lab. Catalog checks verify zero prompt/completion prices for both providers. LongCat uses default reasoning without an invented effort scale. status, pause, resume, continue, stop, and export use the same --root; continue is an explicit operator action for blocked work. Reports distinguish artifact checks from full mission qualification, which still needs review. Catalog checks and short runs do not certify days of uptime or every browser/desktop stack. reconcile optionally refreshes missing native usage and cost measurements from OpenRouter generation metadata. It does not retrieve stored prompts or completions.

New unattended phases select managed headless Chromium in the isolated lab and check that it can render before admitting agents. admit --browser-check adds a local-page rendering and screenshot gate. Browser waits requiring user recovery are recorded and stop the unattended phase; a working dashboard alone does not prove that its coordinating agent remained active.

Original VARIANT-1 code is available under the MIT License. Third-party components retain their own licenses; see THIRD_PARTY_NOTICES.md.