Undes
Get started
Documentation

Undes documentation

Learn how to install Undes Community, configure your AI provider keys, run engineering analyses, and understand the structured artifacts produced by the agent pipeline.

Overview

What is Undes?

Undes is an AI orchestration tool for engineering workflows. Instead of asking one model for a single polished answer, Undes coordinates multiple AI agents through a stricter process: hypothesis, implementation proposal, critique, evidence review, consensus and final synthesis.

The output is a structured engineering artifact that can be reviewed by humans before it is used in code review, architecture decisions, bug investigation or CI/CD checks.

Core idea: Undes is not optimized for chat speed. It is optimized for traceable engineering judgment.

Editions

Community, Pro and organization discussions

Undes is distributed in different editions. Community is public. Pro uses a separate licensed package. Team and Enterprise usage is discussed directly and does not have a public install path yet.

Edition Distribution Intended use
Community Public npm package First tests, local experiments, workflow evaluation.
Pro Licensed package Regular professional usage, session history, engineering memory, support.
Team / Enterprise Direct discussion Organization requirements; scope and access are not publicly packaged yet.

Capabilities

What each edition is meant to cover

The public boundary is intentionally conservative. Community should demonstrate the local generate-and-verify loop. Pro is the paid individual package for deeper usage. Team and Enterprise remain one direct-discussion organization track until architecture and functionality are decided.

Capability Community Pro Team / Enterprise
Local CLI generate-and-verify loop Yes Yes Not committed
First-party BYOK providers OpenAI, Anthropic, Google OpenAI, Anthropic, Google Not committed
OpenAI-compatible provider layer No Paid provider-expansion track Not committed
Local model servers No Paid local-model track with qualification guards Not committed
Cross-run history and engineering memory No Yes Not committed
Rich inspection and exports Limited Yes Not committed
Organization deployment, workspace or policy layer No No Direct discussion only
Provider boundary: OpenRouter, NVIDIA NIM, Ollama, LM Studio, llama.cpp, vLLM and generic OpenAI-compatible gateways are paid-edition provider-expansion surfaces, not Community features.

Installation

Install Community package

The Community edition is installed from the public npm package:

Terminal 
npm install -g @undes.ai/cli

Check the installation:

Terminal 
undes --version
Note: Pro installation commands are not public. They are provided in your account after purchase. Team and Enterprise install paths are not public yet.

Configuration

Configure AI provider keys

Undes is BYOK-first. You provide your own AI provider keys. This keeps provider billing, model settings, usage limits and data handling under your own provider account.

.ai.env 
OPENAI_API_KEY=your_openai_key
CLAUDE_API_KEY=your_claude_key
GEMINI_API_KEY=your_gemini_key

You can start with one provider and add more later. Community supports the three first-party cloud providers: OpenAI, Anthropic and Google.

Pro is the provider-expansion track. That is where OpenAI-compatible aggregators, hosted endpoints, local model servers and generic gateways belong when they are available in the licensed package.

Project initialization

Terminal 
undes init --project-path=/path/to/project
cp /path/to/project/.ai.env.example /path/to/project/.ai.env
# edit .ai.env and fill one or more provider keys

undes init creates .ai.env.example for provider keys and ai/context.json for project goals, context selection and pipeline policy. Treat generated project files as configuration owned by your repository.

Running analysis

Run an engineering analysis

Run Undes from the root of the repository you want to analyze:

Terminal 
cd your-project

undes run "Analyze the authentication flow and identify risky assumptions before refactoring."

A good Undes task should be concrete. It should point to an engineering concern, risk, decision or subsystem that requires verification.

Good Review the checkout flow for race conditions and missing idempotency checks.
Good Analyze this PR for unsafe assumptions, missing tests and architecture risks.
Weak Improve my code.
Weak Tell me if this project is good.

Pipeline flow

How Undes turns a prompt into a reviewable artifact

Undes uses a staged process rather than a single chat completion. The exact internal phases may evolve, but the public contract is stable: the answer should expose evidence, assumptions, rejected hypotheses, risks and open checks.

01 Context preparation

Collect repository material, project configuration and task-specific constraints.

02 Proposal and critique

Generate candidate reasoning, then challenge it with independent review passes.

03 Evidence expansion

Fetch missing anchors when claims depend on files, symbols or snippets that were not delivered yet.

04 Consensus and adversarial review

Reconcile disagreements, look for unsupported conclusions and mark diagnostic limits explicitly.

05 Final synthesis

Emit a structured result that separates the decision from evidence, assumptions and remaining validation work.

Artifacts

Understand the output

Undes produces artifacts intended for review. The final output should make the reasoning path visible: what was checked, what was assumed, which hypotheses were rejected and what still needs validation.

Decision

The final recommendation, conclusion or next action.

Evidence used

Files, logs, snippets, facts and observations used to support the conclusion.

Assumptions

Claims that influenced the result but still depend on incomplete context.

Rejected hypotheses

Alternative explanations considered and ruled out during critique.

Risks

Possible failure modes, unsafe assumptions or unresolved design concerns.

Open checks

Tests, manual review steps or missing evidence required before merge or release.

Prompting

Write prompts that lead to useful artifacts

Undes works best when the prompt includes a concrete engineering situation, expected risk, relevant subsystem and desired type of review.

Prompt template 
Analyze [subsystem/change] for [risk type].
Check [specific files/flow/behavior] if available.
Separate evidence, assumptions, rejected hypotheses and open checks.
Do not present unsupported conclusions as facts.

Examples

01

Analyze the payment retry flow for idempotency issues, duplicate charges, missing transaction boundaries and unsafe assumptions.

02

Review this pull request for missing tests, concurrency risks and architectural changes that could break existing integrations.

03

Investigate why direct file-read intent is selected during planning but does not reach the final synthesis artifact.

CI/CD

Use Undes in background workflows

Undes is suitable for tasks where a longer run is acceptable: scheduled analysis, PR review, architecture checks or release-readiness review.

CI command 
undes run "Review this pull request for risky assumptions, missing tests, and unsafe architectural changes."

For CI/CD usage, prefer narrow prompts and stable scopes. Ask Undes to review a specific change, subsystem or risk category rather than the entire codebase without boundaries.

Recommendation: For organization-wide CI/CD usage, contact us first so the scope can be discussed directly.

Security

Security and data handling notes

Undes may process prompts, code snippets, files, logs, repository metadata and generated artifacts. Before running analysis on sensitive repositories, check your provider settings, internal policies and data handling requirements.

  • Use BYOK provider accounts controlled by your organization.
  • Do not send secrets, private keys or credentials as prompts.
  • Review which files are included in analysis scope.
  • Check AI provider data retention and training settings.
  • Discuss sensitive environments directly before assuming an organization deployment model.

Troubleshooting

Common issues

npm cannot find the package

Confirm the package name and registry. Community uses the public package @undes.ai/cli.

The command is installed but not found

Check that your npm global bin directory is in your PATH. You can also try reinstalling the package or using a local npm script.

The AI provider rejects the request

Check API key validity, provider billing, model access, rate limits, region restrictions and provider-specific policy errors.

The output is too broad

Narrow the prompt. Include the subsystem, risk, expected behavior, relevant files and the decision you need reviewed.

The run takes too long

This can be expected. Undes is optimized for verification depth. Use smaller scopes for faster runs.