Isolation

MINT runs plugins with as little process overhead as possible while still tolerating dependency conflicts and crashes. Two strategies cover the spectrum: shared environment and per-plugin venv.

When each kicks in

The conflict check compares each declared requirement against the platform's resolved environment using PEP 440 specifier intersection. If every plugin requirement is satisfiable by a single resolution that also keeps existing plugins working, shared mode is used.

Shared mode

When a plugin's dependencies don't clash, MINT installs the wheel into the platform's environment and mounts the plugin's routers directly inside the FastAPI app. There's no extra process, no extra port, no proxy hop.

	Shared mode
Startup cost	None — plugin code is imported in-process
Per-request cost	Zero — direct function call
Crash blast radius	Wrapped by api/plugins/middleware.py — a route exception becomes a 500 for that route only
Visible to user	Identical to native platform routes

This is the default and the right choice for the vast majority of plugins. Only reach for isolation when you genuinely have conflicting deps.

Isolated mode

When conflict.py detects an unresolvable clash (most often: two plugins want different majors of the same library), MINT provisions a per-plugin venv via uv and runs the plugin in its own subprocess on a dedicated port.

	Isolated mode
Startup cost	One subprocess + one venv per isolated plugin
Per-request cost	One extra HTTP hop (loopback)
Crash blast radius	Subprocess crash; subprocess_manager.py restarts it
Visible to user	Identical URL — the proxy is transparent

The proxy forwards:

• Request method, path, query string, body
• Auth headers (the platform's JWT cookie / bearer token)
• The X-Request-Id for log correlation
• The plugin internal-token (issued by the platform, scoped per plugin) so the plugin can call back to the platform's internal API for things like reading experiments

Communication back to the platform

An isolated plugin doesn't share memory with the platform — it talks back over HTTP using the platform's /api/internal surface, authenticated by the per-plugin internal token. PlatformContext hides this from plugin code: when integrated and isolated, the context's get_experiment_repository() returns a wrapper that issues HTTP calls; when integrated and shared, the wrapper points at the in-process repository.

Plugin code is identical in both modes.

Dev mode proxy

In development, plugins are typically run as standalone subprocesses with mint dev --platform so the developer can hot-reload either side independently. api/plugins/dev_proxy.py consumes a config.dev.toml that maps route prefixes to localhost URLs:

# MINT/config.dev.toml
[proxy]
"/hello-mint"    = "http://localhost:8003"
"/peak-picking"  = "http://localhost:8004"

The dev proxy preserves the production URL shape and forwards normal request headers, including user identity headers when the request has a bearer token. It does not install the plugin into the platform process and does not set MINT_PLATFORM_URL / MINT_PLUGIN_TOKEN for the dev server. Code that needs a full PlatformContext should still be tested with the plugin installed in a disposable MINT instance or with explicit SDK test harnesses.

Trade-offs and guidance

Concern	Shared	Isolated
Startup time	Fastest	Adds 200–500 ms per plugin (venv creation amortized after first run)
Cold-call latency	~0 ms	~1–3 ms localhost overhead
Memory	Shared with platform	Each subprocess has its own Python runtime (~30–80 MB base)
Debugger attach	Attach once to platform	Attach to platform AND plugin process
Logs	Single stream	Per-process; subprocess_manager.py aggregates with a prefix

Default to shared. Reach for isolation when:

• A plugin pins to a major version of a library the platform also uses
• A plugin links a native binary that's incompatible with another plugin's
• A plugin is unstable enough that you want crashes contained as separate processes (rare — middleware already covers route-level errors)

Configuration

Isolation is decided automatically. To force a plugin into isolated mode (e.g., to test the proxy path), set:

{
  "plugins": {
    "forceIsolated": ["my-plugin-slug"]
  }
}

Conversely, to force shared mode and skip the conflict check (rarely a good idea):

{
  "plugins": {
    "forceShared": ["my-plugin-slug"]
  }
}

Next

→ PlatformContext — how plugins reach platform services from either mode → Operations → Deploying — production considerations when isolation kicks in