AstraCode pairs a provider-agnostic engine with a desktop window. The engine speaks to hosted and local model services through one interface, normalises streaming answers, tool calls and usage, and records which route it chose. The window gives that engine a place to work: sessions, workspaces, attachments, artifacts and a searchable history.
Nothing is proxied through AstraNeuron. Requests travel from your machine to the provider you selected with the key you supplied, while sessions and files stay in a local database you can open, export and delete. Where a cheaper route can do the job — a local model or a zero-cost tier — the router prefers it and tells you why.
Key capabilities
The following capabilities are planned for the Windows desktop milestones described in Product scope & deployment.
01
One engine for the providers you already use
Connect OpenAI-compatible services, Google Gemini, a local Ollama or llama.cpp runtime, or a custom base URL from a single window. Model metadata — context window, output limit, vision, tool calling, price and free-tier status — is normalised so the picker and router compare like with like.
02
Keys held by the operating system, not the application
Each key is stored once through the operating system's credential protection and used only by the engine on your machine. Keys stay out of the interface, logs and diagnostics, can be labelled, rotated or removed at any time, and a short probe checks a new key before you rely on it.
03
Streaming conversations you can branch
Answers arrive incrementally with Markdown, code highlighting and a per-turn token and cost badge. Editing a message or regenerating a turn creates a branch instead of overwriting history, so alternatives can be compared and promoted, and stopping a request keeps the partial answer while still recording its usage.
04
Workspaces with explicit file access
A workspace carries its own instructions, default model, memory and scoped tools. Files and folders are added through an explicit grant, the interface shows exactly what the model can see, and access can be revoked item by item.
05
Tools through MCP, behind a permission gate
Attach stdio or Streamable HTTP MCP servers and review the tools each one exposes. Every server and workspace carries an allow, ask or deny policy, so a tool that needs a decision pauses the turn rather than acting silently, with each call kept in a visible trace.
06
Spend caps enforced before dispatch
The usage ledger records input, output and cached tokens, estimated cost, provider, model and route reason for every turn, then aggregates by session, day and month. Soft and hard caps are checked before a metered request leaves the machine; a hard cap blocks paid calls and suggests a local or free route instead.
07
Artifacts, search and portable history
Structured outputs — documents, code, diagrams, slide decks and interactive pages — are stored as versioned artifacts beside the conversation, with clear export paths. Full-text search covers sessions, messages, artifacts and file names, and a session or workspace can be exported as a portable archive and imported elsewhere.
08
Quiet by default
No telemetry, crash upload or background reporting is enabled unless you turn it on, and any payload can be inspected first. A redacted support bundle can be produced on request, the renderer runs sandboxed behind a strict content security policy, and external links open only through an allowlist.
See how it could work for your organisation.
Discuss your workflows, configuration, deployment and integration requirements with AstraNeuron.