Agent Context Manager

One panel that collects everything shaping your AI agents in the current project: skills, agents, rules, MCP servers and plugins. It marks the source of every entity, its scope (global or project), overrides and — most importantly — which applications pick it up (VS Code, Antigravity IDE, Antigravity/Gemini, Codex, opencode, Claude, Copilot).

Screenshots

Summary

Skills

Applicability

Features

• Summary — the winning entity per category after applying shadowing and precedence rules, so you can see what is actually connected to the project.
• Skills, Rules, Agents, MCP, Commands, Plugins — compact card lists grouped by scope (Project, Global (user), Third-party) with one toggle per entity, the description in a hover tooltip and open/reveal actions on hover.
• Enable / disable — disabling renames a managed file to *.disabled (SKILL.md.disabled, *.agent.md.disabled, …), which hides it from every application and is fully reversible; MCP servers flip enabled in servers.yaml. The panel warns when an agent still references a disabled skill.
• Applicability — a compact scope matrix (Project / User / Third-party, + defined, − shadowed, · absent) and an application matrix (entities against the applications that discover them).
• Health / Diagnostics — broken frontmatter, duplicate names, missing referenced skills, invalid YAML/JSONC/TOML, missing secret environment variables, generated Codex agents without a canonical source.
• Actions — enable/disable entities, open files, open generated Codex agents, and copy environment variable names.
• Live refresh — file watchers on ~/.agents/** and the workspace root.
• Secret safety — only environment variable names are read and shown, never values; the panel reports whether each variable is set in the process.

How it works

~/.agents is treated as the canonical source of truth. The extension is read-only except for toggles: it never runs scripts or regenerates files.

Scopes and overrides

• A project entity with the same name shadows the global one.
• Rules follow the usr-commit precedence: the first existing file wins (<repo>/.agents/rules/git-commits.md → other repo locations → ~/.agents/rules/git-commits.md).
• MCP servers merge by key; a project key overrides the global one.

Application mapping

The paths-per-application mapping is versioned in src/model/apps.ts (APP_MAPPING_VERSION) because discovery rules change between releases.

Requirements

• VS Code 1.85.0 or newer.
• A canonical ~/.agents directory (or a custom agentContext.agentsRoot).

Installation

Install Agent Context Manager from the Visual Studio Marketplace, or run npm run vsix and install the generated .vsix.

Getting started

1. Open the Agent Context Manager view in the Activity Bar.
2. Expand Summary to see what actually applies in this workspace.
3. Use the per-category lists to enable/disable entities and jump to their files.
4. Use Applicability to see which applications pick up each entity.
5. Check Health before trusting a skill or agent.

Settings

Commands

Security

• The extension is read-only except for enabling/disabling entities, and it never runs scripts or shell commands.
• It only modifies entities from the current scan and only inside managed roots (~/.agents, the workspace and the known global config directories), with symlinks resolved via realpath.
• servers.yaml is written atomically (temporary file + rename).
• In an untrusted workspace the panel is read-only: enabling and disabling is unavailable.
• It reads only environment variable names, never values.
• File reads are capped at 4 MB.

Architecture

src/
├─ extension.ts            composition root
├─ commands.ts             command registration
├─ config.ts               settings helpers
├─ statusBar.ts            status bar item
├─ model/                  types + versioned application registry
├─ parse/                  YAML frontmatter, JSONC and TOML parsers
├─ discovery/              scanners (skills, rules, agents, mcp, commands, plugins) + merge
├─ services/               context service, watcher, file/MCP editors
├─ views/                  generic list webview, builders, applicability webview
└─ test/                   node:test unit tests for the pure layers

Roadmap

The plan below is fixed here on purpose, so the scope stays explicit.

v1 — inspection and actions (shipped in this repository)

• Skills, Rules, Agents, MCP, Commands and Plugins across global and project scopes, with provenance, shadowing and a Summary view.
• Seven-application model and the applicability webview.
• Health diagnostics, status bar, live refresh.
• Actions: enable/disable entities, open/reveal, copy environment variable names, toggle MCP servers.

v2 — generated vs canonical

• Diff canonical entities against each application's generated config, with a stale marker based on mtime/hash.
• Settings-driven VS Code hint (chat.agentFilesLocations).
• Scaffolds for creating a new skill or agent.
• Richer dashboard in the webview.

v3 — workspace and remote

• Multi-root workspace context per folder.
• Remote / WSL path resolution.

Support the project

Agent Context Manager is free and open source. If it is useful to you:

• ⭐ Star the repository on GitHub — it helps other people find it.
• ☕ DonationAlerts — a one-time donation keeps the project alive.

License

MIT