venv-manager

One Python environment control layer for developers and every AI coding agent.

Written in Go. One static binary, no runtime deps beyond python3 (or uv, if available).

The GIF above is real: venv-manager watch app.py --venv X monitors a file, scans its imports with a tiny AST-lite parser, and pip-installs whatever is missing — every time the file changes. Point it at a script an LLM is iterating on and the venv converges as the code does.

Why

Claude, Codex, Cursor and other coding agents can already run shell commands, create a .venv, and ask for approval before sensitive operations. What they do not share is durable Python environment state.

Sandboxing protects the machine. venv-manager protects the workflow: it gives every agent the same environments, metadata, package history and recovery path, independent of the client that happens to be running.

Two failure modes drove this tool:

1. Human sprawl. Venvs multiply across ~, cache directories eat GB, activation syntax varies by shell, and cloning "the env that worked" means copy-pasting pip freeze between terminals.

2. Agent sprawl. AI agents can install into the wrong interpreter, leave partial changes behind, and lose environment context when you switch client or start a new session.

venv-manager solves (1) with a clean CLI and (2) with a shared Model Context Protocol server, a persistent registry, typed snapshots and diffs, reversible package changes, ephemeral venvs with OS-level sandboxing, and a file watcher that keeps a venv in sync with evolving code.

What the agent sandbox does not solve

The two layers complement each other: agent permissions control what may happen now; venv-manager records what exists, what changed, and how to recover.

Related MCP server: Sympathy-MCP

Install

Homebrew (macOS, Linux):

brew install jacopobonomi/tap/venv-manager

One-line install script (macOS, Linux):

curl -sSL https://raw.githubusercontent.com/jacopobonomi/venv_manager/main/install.sh | bash

From source:

git clone https://github.com/jacopobonomi/venv_manager && cd venv_manager
make install

Requires Go 1.24+ to build, Python 3.x at runtime.

AI integration

MCP server

Exposes venv operations as native Model Context Protocol tools. Claude, Codex, Cursor, Zed and other MCP clients call the same typed tools and operate on the same persistent environment state instead of guessing shell invocations independently.

Wire it up in Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "venv-manager": {
      "command": "venv-manager",
      "args": ["mcp", "--policy", "safe"]
    }
  }
}

Tools exposed (JSON-RPC 2.0 over stdio):

The server defaults to safe policy. Installation, rollback, removal, and arbitrary execution require confirm: true. Use --policy read-only for inspection-only clients, --policy full for unrestricted compatibility, and repeat --allow-tool NAME to expose only an explicit subset. These policies are defense in depth: they remain consistent even when different clients have different approval settings.

Implementation uses zero third-party MCP dependencies. Newline-delimited JSON-RPC 2.0 on stdin/stdout.

Ephemeral execution (uvx-style, sandboxed)

# create → install → run → destroy, all in one call
venv-manager exec --with requests -- python -c "import requests; print(requests.__version__)"

# with an OS sandbox: no network, no writes outside /tmp + the ephemeral venv
venv-manager exec --sandbox --with pandas -- python untrusted.py

--sandbox uses sandbox-exec on macOS and bwrap on Linux. Deny-by-default profile with explicit allow-lists for the venv path, /tmp, and process management. Network is unshared.

File watcher

venv-manager watch app.py --venv myenv

fsnotify on the parent directory (survives editor atomic-rename writes), 500 ms debounce, then:

1. AST-lite regex scan of .py files (skips docstrings, relative imports, local modules/packages, and vendored dirs like .venv, .git, __pycache__, node_modules)

2. Filter against a stdlib module set

3. Resolve import-name → pip-package aliases (cv2 → opencv-python, sklearn → scikit-learn, PIL → Pillow, bs4 → beautifulsoup4, yaml → PyYAML, ...)

4. Diff against installed packages

5. pip install the delta

The venv is always a superset of the current file's requirements. This is the loop the demo GIF above exercises.

Persistent registry

Every environment is tracked in ~/.venvs/.venv-manager/registry.json with creation and last-used timestamps, an optional project path, and tags. Writes are atomic and the registry reconciles itself with live venv directories.

venv-manager registry
venv-manager registry set research --project ~/work/paper --tag data,ai
venv-manager registry research

prune uses registry last_used_at rather than directory modification time when metadata is available.

JSON snapshot as a single-call context primer

venv-manager describe myenv

{
  "name": "myenv",
  "path": "/Users/me/.venvs/myenv",
  "python_version": "3.12.6",
  "python_path": "/Users/me/.venvs/myenv/bin/python",
  "pip_path": "/Users/me/.venvs/myenv/bin/pip",
  "packages": ["requests==2.34.2", "rich==15.0.0", ...],
  "package_count": 12,
  "size_bytes": 45123456,
  "size_human": "43.03 MB",
  "modified_at": "2026-07-20T15:41:35Z",
  "freeze_hash": "sha256:2c58d830...",
  "activation": {
    "bash": "source '/Users/me/.venvs/myenv/bin/activate'",
    "zsh":  "source '/Users/me/.venvs/myenv/bin/activate'",
    "fish": "source '/Users/me/.venvs/myenv/bin/activate.fish'"
  }
}

One tool call, everything an agent needs to reason about the environment. freeze_hash lets an agent detect drift between two describe calls in O(1) instead of diffing package lists.

Commands

Most read commands also accept --json for stable, machine-parseable output.

Configuration

~/.config/venv-manager/config.json (respects $XDG_CONFIG_HOME and $VENV_MANAGER_CONFIG):

{
  "base_dir": "/custom/path/to/venvs",
  "default_python": "3.12",
  "use_uv": true,
  "prune_after_days": 90
}

Bootstrap: venv-manager config init.

uv backend

If uv is on PATH and use_uv: true, create runs uv venv. Typically 10–100× faster than python -m venv on cold cache.

Development

make build            # go build -o bin/venv-manager
make test             # unit tests
make demo             # regenerate scripts/demo/demo.gif via VHS
go test -tags=integration ./internal/manager/...   # integration tests (real pip, real PyPI)

CI runs go vet, go test -race on Ubuntu + macOS, and integration tests on Ubuntu with Python 3.12.

Architecture:

cmd/venv-manager/           cobra CLI
internal/manager/           core operations (create, install, snapshot, scan, watch, exec, describe, ...)
internal/config/            XDG-aware JSON config
internal/mcp/               JSON-RPC 2.0 MCP server (stdio)
internal/tui/               Bubble Tea browser
internal/utils/             platform helpers, size formatting

License

MIT.

Author

Jacopo Bonomi