Project map: agents know where to look before they read

When an agent starts a new task it usually greps first, then reads files one by one. The project map cuts that search short: AgentVera indexes the project’s symbols, imports and files that change together on your computer, and agents ask the agentvera-map MCP tools for the relevant files, their users and their tests.

What gets indexed

The map is built with tree-sitter. In TypeScript, TSX, JavaScript, Python and Go files, functions, classes, methods, interfaces, types, enums and constants are extracted with their line ranges, signatures and the first sentence of their docs. About 30 more languages, such as Rust, Java, PHP and Ruby, are on the map by file path.

  • Imports are resolved to project files; tests are linked to the files they import and to source files with matching names.
  • Change together: files that changed together in at least two of the last 500 commits are paired; commits touching more than 40 files are ignored.
  • Files are found with git ls-files and .gitignore is respected; folders like node_modules, dist and build, files over 1 MB, generated and minified files are skipped.
  • In repositories with more than 20,000 files, only source folders (src, lib, app, packages/*/src…) are indexed.

The agentvera-map tools

The map is added to the agent as an MCP server. Its instructions tell the agent to start a task with map_related, check map_outline before reading a file and map_usages before changing code. A new agent’s instructions also get a short summary of the project’s modules, entry points and test command.

  • map_related: the files a task most likely involves, ranked with reasons. Turkish task words (sepet, kupon, ödeme…) also match their English names in code.
  • map_outline: a file’s symbols with line ranges, how many files import it, and its tests.
  • map_usages: where a symbol is defined and called, the files importing it and its tests; for a file, its importers and their importers.
  • map_find: fuzzy search over symbol and file names, with kind, file:line and signature.
  • map_overview: modules, entry points and the test command. map_api reads the real API of the installed version of a package in node_modules from its .d.ts files.

Fresh and local

There is one map per folder (a project or a worktree), shared by every agent working there. File changes are watched, commits trigger a recount of files that change together, and a file an agent asks about is re-checked before the answer. The map is built on your computer and costs no tokens; each agent can read only its own folder’s map.

Graph and impact

The Project map tab on a project’s Details page draws files as a graph coloured by module: imports in grey, files that change together as dashed orange lines. Type a task and press “Show” to see the files an agent would look at, with reasons. The “Project map impact” card in the AI ROI tab compares conversations that used the map with those that didn’t, by success, first-time pass, tokens per task and follow-ups.

How to turn it on

  1. 1

    Check that the “Project map” switch under Settings → Token savings is on; it is on by default.

  2. 2

    Open the project or start an agent: the map is built once, then follows changes.

  3. 3

    Look at the graph under Details → Project map, and type a task to see which files an agent would look at.

  4. 4

    To leave it out for one agent, clear the “Project map” box in that agent’s edit dialog.

Questions

Which agents use the map?

Agents that take MCP servers: Claude Code, Codex, opencode, Copilot, Qwen Code, Goose and Amp. It isn’t added to agents on remote (SSH) hosts.

Does the map send my code anywhere?

No. The map is built on your computer and stored compressed in the app’s data folder. What an agent reads from a tool answer goes to its own model provider, like anything else it reads.

Which languages get symbols?

TypeScript, TSX, JavaScript, Python and Go. Files in other languages are on the map by path and with their change-together links.

What if the map isn’t ready yet?

Answers say how far along the map is. The agent waits a few seconds at most, then is told to carry on with its normal tools.

How is this different from the code map?

The code map window shows you module dependencies and the files agents touch most. The project map is a live index that agents query themselves.

Related features

Related terms:MCP (Model Context Protocol)Context window

All featuresSee the 18 supported coding agents

Bring your agents to one desk.

Start on the free plan. Download it, add a project folder and open your first agent.