Understand legacy systems
Map entry points and responsibilities in software with limited documentation before maintenance or modernization.
“Where does an incoming request reach the persistence layer?”
Private AI. Local by default.
Private intelligence for your codebase.
Designed to run locally.
Investigate unfamiliar repositories, trace behavior across files, and plan changes with evidence from your own source. Local inference keeps the investigation on your hardware.
Current: repository intelligence · Experimental: local coding agent and document knowledge
Example question
“How does a command-line option become a typed value?”
decorators.pycore.pytypes.py
Follow the flow. Inspect the source.
Illustrative workflow · not a live product response
Your repository. Your compute. Your control.
Built for code you need to keep private.
Proprietary systemsClient code under NDARestricted environmentsFrom investigation to a grounded change
Work through the code
behind your next task.
AndesCode combines repository indexing, source retrieval, and a local model to help your team investigate real engineering questions. You can inspect the source behind recommendations before deciding what to change.
Investigate dependencies
Distinguish declared dependencies from runtime usage, then inspect the files behind the explanation.
“Where is this library declared, and which modules use it?”
Trace behavior before a change
Follow a feature across files in an unfamiliar repository. Review likely change locations, related tests, and missing context.
“How does a failed upload reach the retry path?”
Evidence before confidence
Recommendations you can
check against your source.
Review the source behind an explanation and the gaps in its coverage. Answers and recommendations still need an engineer’s judgment; real-task accuracy and productivity gains have not been established.
Inspect the evidence
File, symbol, configuration, and semantic search signals help locate relevant code. Inspect retrieved files and compare the answer with the actual implementation.
Keep the task in context
The experimental agent works within a bounded context budget and can read additional source as it investigates. Large or unusual codebases may need more targeted questions.
Make changes reviewable
Experimental edits use stale-content checks and require diff inspection. When validation is enabled, the agent must attempt an allowed check. Passing these gates does not prove the change is correct.
Product stages · Current and planned
Start with code.
Build toward private knowledge.
Repository understanding is the starting point. A separate local coding agent and a single-user document-knowledge prototype extend the work. Shared enterprise access and workflow actions remain future capabilities.
Current: repository understanding
Desktop Q&A and source-grounded change recommendations. Start with a read-only evaluation against questions and code your team knows.
Experimental: agent and local documents
A separate coding-agent CLI can inspect, make controlled edits when enabled, and attempt allowlisted validation. A single-user local knowledge CLI supports scoped Markdown/TXT collections. Neither is a production multi-user enterprise service.
Roadmap: customer-controlled private AI
The next phases target permission-aware shared knowledge, authorized enterprise connectors, and explicitly approved workflow actions, with independent security and acceptance gates. No release date or enterprise certification is promised.
A clear data boundary
Local means
on your machine.
In LOCAL mode, the repository, index, retrieval, prompts, and model inference stay on the same machine.
Review deployment boundariesPrepare once. Run locally.
Initial setup downloads dependencies and model files. Once provisioned, LOCAL inference can operate offline.
Remote inference is an explicit choice.
Optional private-server inference sends your question, selected source chunks, and metadata to the server you configure. Review that data flow before enabling it.
Validate against your own requirements.
No formal compliance certification is claimed. Network restrictions, deployment controls, and security approval belong in the evaluation.
Deployment and access boundaries
LOCAL inference can operate offline after dependencies and model files are provisioned. Optional remote inference sends your query, selected source, and metadata to the configured endpoint. Agent test and build commands execute repository code: application policies are not an OS sandbox or proof of zero network egress. Review host isolation and network controls for restricted environments. A public technical overview describes the processing model and its limitations. The AndesCode source repository remains private; it is not part of a standard evaluation.
Start small. Measure what matters.
One repository.
A concrete evaluation.
Discuss a small, paid, read-only repository-understanding evaluation on a public or otherwise authorized codebase. Confirm hardware and product readiness first; agree on questions, success criteria, deliverables, scope, price and schedule before any work is committed.
Define the test
Choose one repository and representative engineering questions. Confirm hardware, privacy requirements, and success criteria.
Evaluate locally
Provision AndesCode in your environment and assess answers against source code your engineers know. Begin with read-only Q&A.
Make an informed decision
Review source-backed findings, answer quality, response time, and limitations against an agreed baseline, such as your team's current investigation workflow. Decide whether a wider pilot is worthwhile.
The deliverable: a report with source-backed findings, measured results, gaps, and a recommendation on whether to continue.
Before we talk
A few practical details.
What hardware should we plan for?
Apple Silicon with 32 GB or more of unified memory is the recommended starting point. Plan for 20–30 GB of free disk space and Python 3.10+. We confirm the model and hardware fit before agreeing to an evaluation. Windows and Linux with NVIDIA GPUs are documented paths; platform suitability should be checked for your setup.
Will it automatically change our code?
The initial evaluation is read-only. An experimental local coding-agent CLI already exists and is under active development. With editing and validation explicitly enabled, it can inspect source, apply controlled edits, attempt allowed checks, and inspect the final diff. It is separate from the desktop chat interface; reliable real-task performance remains to be established.
How can we evaluate it?
We can discuss a bounded, read-only evaluation using public or otherwise authorized code, subject to hardware and readiness checks. The AndesCode source repository is private and is not included in standard evaluations. No installation or usage rights are granted by an enquiry. Read the public technical overview.
Is this a hosted service or IDE plugin?
AndesCode is a standalone desktop app. There is no official hosted SaaS or current IDE plugin. Private remote inference is optional and must be explicitly configured and operated by your team.
Let’s see if it fits
Bring a codebase.
Start with a conversation.
Tell us the stack, available hardware and two or three investigation questions for a public or authorized repository. We will assess fit and readiness before agreeing to an evaluation. Please do not email source code, credentials or confidential information.
Discuss an evaluation andescode@datalogic.techOpens your email app. You can also write directly to the address above.
Built in the Andes.
Designed for your environment.
Andrés Garrido
Independent software engineer · Chile