Everything a coding agent should do, without leaving your network
iCode reads your code, edits it, runs your tests and reports what really happened. Your administrators decide what it may do, and every step is recorded.
An agent that proves its work
The engine decides when a job is finished, not the model. iCode runs your tests itself, checks they were not edited to pass, and catches any result the model invented.
- Ask, Plan and Agent modes, plus modes your teams define
- Undo a whole run, one step or one file, even without version control
- Several attempts at a hard fix, keeping the one that passes the most tests
- Hooks run your own checks before and after each step
Ask about code you have never cloned
Connect your repositories once. Answers point to the real files and lines, and each source is tied to the roles allowed to see it.
- Search across every connected repository
- Credential files, keys and environment files are never indexed
- People only see answers from repositories they may read
See what a change will touch before you make it
iCode maps how your code fits together: who calls what, which services talk to each other, and which tests cover a path. Answers and plans cover the whole system, not one file.
LedgerClient.postAny model you host, measured and tuned
iCode talks to any OpenAI-compatible endpoint. Connect a model and iCode measures what it can really do, then sets itself up to match. Drift checks catch a quality drop after a model or settings change.
On your GPUs
The bundled model server starts from weights you copy in, and never downloads anything at run time.
Through your AI gateway
Point iCode at the gateway you already run. Keys stay in the service, never on laptops.
Per role and team
Choose which models each role may use, with budgets and rate limits per team.
Everything on. Your admins decide.
Every feature ships switched on. Administrators switch off what they do not want, for one team or for everyone, from a page served by the service itself. Controls your regulator requires can be locked so nobody can change them.
- Around thirty switches, per tenant
- Every change recorded, with a reason when a guardrail is weakened
- Usage per team, model and day
Learns how each person works
Standing preferences and methods that worked become personal skills, reviewed on your own model. Everyone can see, edit, lock or delete what was learned, and nothing leaves your network.
Guardrails
Secrets are masked before anything is stored. Package installs follow your policy: public, internal mirror or blocked.
Security scanning
Built-in checks flag vulnerable dependencies and risky code in what the agent writes, before you keep it.
Audit trail
Every request, tool call and change lands in an audit log you own, ready for your monitoring tools.
See it on your own code
Book a demo, or run a pilot with an evaluation licence in your own network.