The engine handles "how", and the server decides "what is allowed and what is recorded" — AI coding delivery becomes trustworthy.
Reads code, edits files, runs tests, commits and opens MRs in a real repository workspace — producing real files, real diffs and real branches, not an AI narrating "I did it".
Code indexing, a scope graph and workspace discovery support cross-file tasks; a context bundle (role constraints + work-order spec/plan/AC + model routing) is materialized and injected per task.
Driven by the desktop and sandbox through a managed protocol: long-lived sessions, real-time event streaming, and permission and approval cards. Native built-in login is fully locked down, and the managed entry point binds to enterprise identity.
Sub-agents start in parallel or nested by task complexity, but every sub-agent must pass engine-level consent confirmation before its initial spawn — "whether to use sub-agent collaboration" becomes a decision the enterprise can confirm.
Every session is pinned to an identity subject (issuer / tenant / user) with cross-tenant access failing closed; engine actions run in a restricted sandbox, and tenants sharing an installation are invisible to each other.
The engine holds no model-provider keys — it receives short-lived session tickets issued by the server, forwarded to the real model through an engine gateway. The real key never leaves server-process memory.
A text-based primary model can still see and hear: images are turned into descriptions by an auxiliary vision model, and voice runs through the engine's built-in voice runtime. Model routing is always decided server-side.
Work orders, knowledge, attachments and messages enter the engine through a single standard MCP exit — business capabilities are not scattered across the engine side, and credentials are never passed down to clients.
Repositories with hooks or configuration are not loaded for execution by default (the Folder Trust gate). A patch ledger plus source-equivalence checks can answer at any time: where did this code come from, what changed, and can it be rolled back.
The AI dev team completes work orders in real repositories — no simulated progress, but real edits, real tests, and real commits.
Triage → spec → plan → AC; the context package is materialized and injected into the engine with each task.
Inside the Executor sandbox: clone → read code → edit files → run tests.
Real browser screenshots + a vision model judging the render — not jsdom guesswork.
Push branch → open MR → code review → merge gate → backfill the work order.
Book a demo to see a work order fully completed by AI in a real repository, returned as a mergeable MR.