Preview iCode 0.23.1: pilots open for banks and government teams in air-gapped networks. Book a pilot →
Preview

What's new in iCode — November 2026

The iCode agent is now something you can hand real work to. It checks its own work against your tests, can undo anything it changed, and follows the permission rules your administrators set. Everything below runs inside your network, with no calls home.

An agent that proves its work

  • The engine decides when the job is finished, not the model. iCode runs your tests itself and checks that they were not edited to pass.
  • No invented results. If the model reports output that no command actually printed, iCode catches it and shows what really happened.
  • When a fix still fails, iCode can try several versions and keep the one that passes the most tests.

Undo anything the agent did

  • Before the agent changes, moves or deletes anything, iCode keeps a copy. Undo a whole run, everything from a given step, or a single file.
  • It works without version control, so changes you have not committed yet are safe.
  • Pick a run up again later with its full conversation, or branch from it to try a different approach.

Your rules, your modes, your checks

  • Permission rules in three tiers: your organisation, the project and each developer. An organisation's deny always wins.
  • Agent, Plan and Ask modes, plus modes your teams define themselves.
  • Hooks run your own programs before and after each step, and can block a step and tell the agent why.

Chat with your whole codebase

  • Connect your repositories once and ask about code you have never cloned. Answers point to the real files.
  • Each source is tied to roles, so people only search the repositories they are allowed to see.
  • Credential files, keys and environment files are never indexed.

iCode Code Graph

  • A map of how your code fits together: what calls what, what imports what, and which design notes explain which code.
  • See what a change could break before you merge, and which branches are likely to collide.
  • It needs no AI model and no network to build, so it works on air-gapped sites.

Measure any model before you trust it

  • Model fit measures what a model can really do and suggests the right settings.
  • Drift checks catch a quality drop after a model, prompt or settings change, before your users notice.

These features are in preview and are being rolled out to customers in stages. The product deck, with the full architecture, is available on request below.

The features on this page are in preview and are being rolled out to customers in stages. Contact IOTA to join the preview or to confirm what is included in your edition.