Skip to content

Record your screen, ship a demo. Free and open-source, GPU-accelerated, no watermarks, no subscriptions. Windows, macOS, Linux. Actively maintained.

docs(mcp): list restoreCheckpoint among the destructive tools

EtienneLescotc3b2cad433ef
diff --git a/technical-documentation/architecture/mcp-server.md b/technical-documentation/architecture/mcp-server.md
index dd53df464..7b7a336b2 100644
--- a/technical-documentation/architecture/mcp-server.md
+++ b/technical-documentation/architecture/mcp-server.md
@@ -15,7 +15,7 @@ OpenScreen can offer the in-app agent's tools to MCP clients the user runs thems
Nothing about the tools is reimplemented. The server registers `TOOL_ARG_SCHEMAS` (names and zod schemas), `TOOL_DESCRIPTIONS`, and hands `buildSystemPrompt` to the client as the server `instructions` — all exported from [`deep-agent/service.ts`](../../electron/ai-edition/deep-agent/service.ts), where `buildTools` builds the in-app agent from the same table. Every call goes through `runDocumentTool`, the function the in-app agent's `documentTool` also calls: the cursor-telemetry read the zoom tools need, then `executeAgentTool`.
-So a tool added to the agent appears over MCP with no further work, and the MCP test asserts the listed tools equal `OPENSCREEN_TOOL_NAMES`, followed by the two checkpoint tools below, the only ones that exist over MCP alone. What the server adds is MCP metadata only: `readOnlyHint` for the reads (`!isMutatingTool`) and `destructiveHint` for `replaceTimeline` and the three `remove*` tools.
+So a tool added to the agent appears over MCP with no further work, and the MCP test asserts the listed tools equal `OPENSCREEN_TOOL_NAMES`, followed by the two checkpoint tools below, the only ones that exist over MCP alone. What the server adds is MCP metadata only: `readOnlyHint` for the reads (`!isMutatingTool`) and `destructiveHint` for `replaceTimeline`, the three `remove*` tools and `restoreCheckpoint`.
**Writes are a second opt-in.** MCP clients have their own "Project edits" switch, `allowEdits` in `mcp-server.json`, **off by default** and independent of the in-app agent's `allowAgentEdits`. Turning the server on therefore grants read access only. The value is read on every call and passed to the executor as `editsAllowed`, the same gate the in-app agent's switch drives: while it is off every mutating tool is refused with the executor's consent message, and the consent block of the system prompt is in the instructions. It is independent because `allowAgentEdits` defaults to allowed and lives in the provider form, so a user with no provider configured could never have turned MCP writes off.