내 저장소에서 시작하기

Start with one question you want your codebase to answer: “How does an order get accepted, and where would I check before changing that rule?” A useful first map preserves that answer, its source evidence, and what is still unknown. You can grow it after reviewing this first slice.

Checked against v1.6.0 on October 5, 2026. In this version, repository scans produce review candidates. bootstrap does not write repository-derived nodes; analyze --apply writes nothing, and infer-imports --apply is disabled. Creating a starter folder, reviewing a proposal, and accepting its meaning are separate steps.

1. Choose your starting point

Atlas and the agent read the same Markdown folder. Opening it in Atlas does not upload your repository to an Atlas backend. A connected coding agent uses its own provider and data settings.

2. Set up the CLI once

The commands below need Git, Node.js 24, and pnpm. The CLI runs from the Atlas source checkout; there is no npx ontology-atlas package. Keep the tool checkout outside the repository you want to describe.

Run these commands in your terminal:

mkdir -p "$HOME/tools"
git clone --branch v1.6.0 --depth 1 https://github.com/wlsdks/ontology-atlas "$HOME/tools/ontology-atlas"
cd "$HOME/tools/ontology-atlas"
pnpm install --frozen-lockfile
pnpm --dir mcp install --frozen-lockfile
ATLAS_CLI="$HOME/tools/ontology-atlas/cli/src/index.mjs"

Both install commands are needed: mcp/ has its own lockfile. This example pins the checked release. If you already have an Atlas checkout, use its absolute CLI path instead of cloning over it; a newer checkout may have newer behavior.

ATLAS_CLI lasts for this terminal session. In a new terminal, set it again.

3. Create the vault inside your repository

Replace /absolute/path/to/your-repository with the folder containing your code, then run:

cd /absolute/path/to/your-repository
node "$ATLAS_CLI" init ./atlas
node "$ATLAS_CLI" mcp-verify ./atlas

init writes starter Markdown, agent configuration, and starter skills. The starter examples give you a shape to work with; they are not an analysis of your product. Review the resulting changes in git diff before committing them. If you already have a vault, keep it and use its path instead of creating another.

mcp-verify starts a server and checks that it can answer real queries. It does not prove that a running agent has loaded the configuration. Restart your coding agent from your repository, then ask:

Call connection_info. Show the resolved vault and repository paths, and confirm which folder contains the code we are about to read. Do not write yet.

The vault should resolve to your repository's atlas/ folder; the repository should resolve to your code root. They normally differ. If either is wrong, repair the connection before continuing.

4. Ask the agent for one useful slice

Use a coding agent that can read your repository and call Atlas MCP. The app's internal vault conversation does not have the repository-analysis tools.

Copy this prompt and replace the bracketed goal with a real behavior in your product, such as “accept an order” or “publish a document”:

Help me build the first useful Atlas map of this repository.
Focus on [one product behavior I want to understand].

First confirm the vault and repository with connection_info. Read list_kinds,
narrow list_concepts, and the full bodies of any relevant existing concepts so
we do not duplicate them. Treat starter examples as examples, not product facts.

Read the actual entry point, important conditions, failure paths, and relevant
callees. Use source and current documentation as evidence. Do not turn folders,
README headings, or import counts into accepted product concepts.

Propose a small batch in plain language before writing. For each concept, show
what it means, what it includes and excludes, its exact implementation evidence,
and what you could not verify. For each relation, explain why it holds and any
condition that limits it. Keep unsupported conclusions unknown.

Stop so I can correct, reject, or approve the exact proposal. After approval,
write only the selected concepts and relations, validate and compile the vault,
preview the project source binding, and finalize the accepted project meaning.
Report remaining gaps separately from successful writes.

For a first pass, a handful of concepts is enough. Ask for the shortest map that answers your question. More nodes can wait until you need another answer.

5. Review meaning before saving it

A good proposal lets you disagree with a specific claim. Here is an illustrative order example, not output from your repository:

Before approving, check these points:

• Can you explain it? The name describes a product ability or implementation role. A folder name or a technology name by itself tells you too little.
• Can you challenge it? The body states a definition, included and excluded responsibilities, and uncertainty. “We did not read this” is an uncertainty, not a product exclusion.
• Can you verify it? Implementation claims lead to an actual file and the relevant behavior, including its conditions. A path's existence alone does not prove the explanation.
• Does the relation hold? Ask whether the ability can still succeed when the target is skipped, disabled, or replaced. Keep that condition in the rationale.

You can reply, for example: “Keep order acceptance. Rename stock reservation to match its actual role. Leave the payment dependency unknown until you read that branch. Show the revised proposal before writing.” Approve the revised content you actually saw.

Small reviewed batches use the ordinary MCP writers. A bulk analyze_repo_structure write plan has a separate qualification lifecycle: independent evaluation, the exact review plan and gaps, human acceptance, and an unchanged released writePlan. If that evaluation is unavailable, continue with small explicitly approved batches; the builder must not invent an evaluator.

6. Open the map and check the answer

Open your repository's atlas/ folder in Atlas. The app is another reader of the files the agent wrote; it does not need a separate import of the graph.

After the approved writes, ask the agent to:

1. Run validate_vault and compile_ontology and report errors and warnings.
2. Preview connect_project_source for the project. Review the proposed code folder before confirming the binding.
3. Run finalize_project_meaning and report the returned assessment and gaps.

Then try answering your original question from the map and full node bodies. You should be able to explain the behavior, find its implementation starting point, and distinguish declared dependencies from unknown impact. If you cannot, ask for the missing explanation before expanding the map.

Validation success means the documents and graph passed those checks. A saved receipt or a large graph does not establish complete or accepted meaning.

7. Preview repository evidence from the terminal

These are optional inspection commands. Run them from your code root, using the CLI path from Section 2:

node "$ATLAS_CLI" analyze . --vault ./atlas
node "$ATLAS_CLI" infer-imports . --vault ./atlas
node "$ATLAS_CLI" bootstrap . --vault ./atlas

analyze proposes structure candidates. infer-imports reports observed code imports. bootstrap combines them into a review-only plan; exit code 3 means approval is required, and no repository-derived meaning was written.

The number of imports is evidence of code use, not a score of business importance or proof of a depends_on relation. --threshold N filters candidates for review; it does not approve them.

init ./atlas --quick-start also exists, but it creates the scaffold before running this review-only bootstrap. It can therefore leave starter files and configuration while returning exit code 3. It is not a one-command completed map.

8. Reuse documents you already have

A folder of Markdown documents: open it in Atlas. When untyped documents are found, the map offers Build the map from my documents. Review the candidates: folder and document names are starting suggestions. Approval adds frontmatter to selected documents while retaining their bodies, and creates or updates the project document. This does not inspect your repository's implementation. Agent/runtime files and the Library's sources/ and wiki/ are excluded.

An existing AGENTS.md or CLAUDE.md: preview absorption from your code root:

node "$ATLAS_CLI" absorb AGENTS.md --vault ./atlas

The default is dry-run. Review the plan before repeating it with --write. That write backs up the original as <file>.pre-absorb.bak, creates document nodes for the policy sections in the plan, and replaces the instruction file with a thin pointer preserving unabsorbed sections. Architecture sections remain proposals; suspected injection content is not absorbed. This rewrites an agent instruction file, so inspect both the backup and the resulting diff.

9. If you get stuck

Next, use What Becomes a Node to sharpen the concepts, Relations to judge the links, and After the Vault Grows to keep the map useful as code changes.