Vector Agent

The main session can inspect the repository, edit files, run commands, use repository tools, and report a reviewable result. Use Plan mode when you want analysis without edits, and switch back to an editing mode when the plan is approved.

When a big task splits into pieces that do not depend on each other, the agent hands them to subagents that run at the same time; a small task it does itself. It calls a subagent specialist when part of the work fits one. See Subagents and subagent specialists.

Modes

The agent's tools

The agent changes your repository through a fixed set of first-class tools rather than free-form shell commands, so its behaviour is the same in the desktop app, in the terminal and in GitHub Actions.

• Filesread with line offsets and limits, write, edit for an exact string replacement, and apply_patch for a patch that spans several files.
• Searchglob for filename patterns and grep for content, so it finds code without reading whole directories.
• Language serverslsp asks your repository's language servers for a definition, the references to a symbol, hover type information or a workspace symbol search — the same answers your editor gets.
• The webwebfetch returns one URL as markdown, text or HTML. In the desktop app the agent searches the web with Vector's own browser: browser opens a Google results page for its words and reads it, the way the address bar turns typed words into a search, after you approve the site. Deep research searches the same way in each research agent's own isolated browser, and falls back once to DuckDuckGo's HTML page when Google asks for a check; Vector never answers a check. There is no search service or search key to set up.
• The shellbash runs commands in the repository, behind the permission rules.
• Keeping it legibletodowrite maintains the checklist you can see, and question lets the agent stop and ask you a structured question instead of guessing.

Tools are built in. MCP servers and plugins add more beside them; they never replace these.

Context and compaction

A long session does not fail when it reaches the model's context limit. Vector compacts it: older turns are summarised and pruned while a protected window of recent work is kept intact, and the run continues from the compacted history. A dedicated internal agent writes that summary, so it is written for the model that has to keep working rather than for a reader.

Compaction can also be triggered deliberately, and the transcript records where it happened.

Repository rules

A rule is a standard written the way you would say it to a new teammate — "we don't put database query logic in the controller". Repository rules live in .vector/RULES.md, a plain Markdown file in the repository: write a rule there, commit the file, and your whole team gets the same rules. Rules for every repository go in the AGENTS.md in Vector's own config folder, which every session reads. The desktop app keeps that folder with its app data: ~/Library/Application Support/ai.vector.app/config/vector/AGENTS.md on macOS, %APPDATA%\ai.vector.app\config\vector\AGENTS.md on Windows and ~/.config/ai.vector.app/config/vector/AGENTS.md on Linux. The terminal reads ~/.config/vector/AGENTS.md (%USERPROFILE%\.config\vector\AGENTS.md on Windows) instead, so each has its own file.

1. 1
   Write it in plain English

   No rule language, no schema. One sentence describing the standard.

2. 2
   Scope it in words

   Say which files it governs, such as src/app/web/*.tsx, in the sentence itself.

3. 3
   Commit it

   The file travels with the clone, so everyone working in the repository reads the same rules.

In Vectorscope, Don't flag this again on a finding adds a rule to a section Vector manages in the same file, and Undo takes it out again. Each line in that section is one rule: delete a line to remove the rule, or edit it to change the rule, and Vector keeps your change the next time it saves. Vector rewrites everything else in the section, so keep notes of your own outside its markers.

• Vector agentsRead the rules file directly — the engine loads it as an instruction file on every prompt.
• Claude Code, Codex, CursorRead their own instruction files, so Vector copies the rules in its section of the file, as the file has them now, into the prompt it hands them, and tells them to read the rest of the file.
• Scope survives as wordsInstruction files load whole, so the file paths a rule covers are stated in the sentence the agent reads.

A rule that conflicts with the task you asked for is reported rather than silently resolved. Vector tells the agent to say so instead of quietly picking a side.