Public briefing · 3 November 2026
Customer-controlled AI for repository investigation, with a local desktop core and separately evaluated experimental capabilities.
Index one selected repository, build a project map and retrieve relevant source. Desktop Q&A and read-only change recommendations help investigate structure, dependencies and behavior. Answers require verification; repeatable customer quality results are still pending.
A separate coding-agent command-line tool supports bounded inspection, optional controlled edits and allowed validation. A separate single-user Markdown/TXT knowledge prototype returns exact source excerpts and citations. Generated synthesis needs further evaluation.
Enterprise identity, source-permission synchronization, production connectors and approved business actions are future capabilities. No shared multi-user deployment is offered today.
Repository content, index, retrieval, questions, prompts and inference stay on the same machine. Model and dependency provisioning can require downloads. Offline operation must be tested on the agreed configuration.
Explicit configuration sends the question, selected source chunks and metadata to your chosen inference endpoint. The repository index remains local. Endpoint security and retention need separate review.
Indexes, caches and logs can retain sensitive data or metadata. Protect them with account controls, disk encryption, restricted permissions and an agreed retention/deletion procedure. No certified erasure is claimed.
The starting scope is one operator on an isolated machine. The application does not provide enterprise source ACL enforcement or an OS sandbox. Experimental validation can execute repository-controlled code. Network isolation and host controls must be applied independently; no compliance certification or verified zero-egress guarantee is claimed.
Apple Silicon with 32 GB or more of unified memory and 20–30 GB free disk is a planning starting point for the documented desktop model, not a verified fit for every workload. Confirm the exact model, available memory and index size. Windows/Linux paths are documented but require separate validation. Model-free text-document retrieval has different requirements.
Deterministic tests cover retrieval policies, freshness, bounded agent tools and completion gates, and the local text-document path. Synthetic source-evidence checks are distinct from model reasoning quality. Passing unit tests does not establish production security, coding performance or customer usefulness.
The public-repository demonstration plan identifies a pinned source revision and review questions. Its live repository Q&A run is pending. No benchmark or recorded model response is presented.
The AndesCode development repository is private. A standard evaluation does not include source access or grant installation, usage or redistribution rights. Agree on an operator-led format or a separately approved evaluation artifact and license before starting.
To discuss fit, share only your stack, hardware and non-confidential questions.
andescode@datalogic.tech