Skip to content

Droid: MCP tools force approval at every autonomy level — no risk annotations #1097

Open

Description

@ahjota

User problem

I was really happy to see support for Factory Droid in #329 and #331 — that's what finally pushed me to evaluate agentmemory for my own use.

But the actual experience took me aback: Depending on the task, I usually run Droid sessions with a lower level of autonomy (Auto (Low) or Auto (Medium)) to keep destructive actions gated. I was expecting agentmemory to just work given its marketing in the README. What I actually found was that agentmemory MCP "read" calls, like memory_recall, memory_sessions, memory_slot_list -- would not happen, and "write" calls, like memory_save -- required my explicit approval.

This makes low / medium autonomy sessions in Droid unusable with agentmemory, and if I crank to high autonomy, I lose the otherwise restrictive gating I usually want.

Droid classifies tools missing safety metadata as high risk, so they prompt at every level below Auto (High). This makes low/medium autonomy unusable with agentmemory: I either approve every recall/search by hand, or I crank to high autonomy and lose the gating I want.

What I expected

In Auto (High) mode, Droid should keep auto-running everything related to agentmemory.
In Auto (Medium) mode, Droid should auto-run read-only tools (search, recall, list, get, query, export, audit, etc.) and state-changing tools, and still prompt me for use of destructive tools.
In Auto (Low) mode, Droid should auto-run read-only tools, and still prompt for use of state-changing and destructive tools.
in Manual Autonomy mode, I should be prompted before any tool use.

Why it happens

Droid uses the standard MCP annotations hints (readOnlyHint, destructiveHint) to classify tool risk, comparing each tool's risk level to the session's Autonomy Level and auto-running when risk is at or below that level (see Autonomy Level). Per the MCP autonomy docs, servers with no override use "Droid's built-in tool risk classification (read-only hints and curated defaults)," and the safety floor is one-directional: low/medium cannot pull a tool that isn't read-only (destructive, or missing safety metadata) below high.

Here, src/mcp/tools-registry.ts defines 53 tools but none carry annotations, so Droid falls back to high for all of them.

Proposed fix and scope

I believe the fix is to add annotations to each McpToolDef:

  • read-only tools: { readOnlyHint: true, destructiveHint: false }
  • non-destructive writes: { readOnlyHint: false, destructiveHint: false }
  • destructive tools: { readOnlyHint: false, destructiveHint: true }

The annotations propagate through the existing tools/list response in both stdio and http transports, so no handler or REST changes are needed. Tool count is unchanged.

  • Extend McpToolDef with an optional annotations field.
  • Annotate all 53 tools (23 read-only, 26 non-destructive writes, 4 destructive).
  • Add test/mcp-tool-annotations.test.ts asserting classification + wire-shape.
  • One-line note in AGENTS.md consistency checklist.

Happy to open a PR if this sounds right.

Activity

  1. ahjota commented on Jul 22, 2026

    @ahjota
    Author

    A little more context. Droid has a built-in workaround -- when an individual tool call exceeds the session's autonomy level, Droid shows an approval prompt with three "always allow" options:

    • Always allow this tool — persists permission for one specific tool
    • Always allow all "agentmemory" tools — persists permission for every tool from the server
    • Yes, allow / No, cancel — one-time decision

    The second option ("always allow all agentmemory tools") is the only way to quiet the prompts without escalating to high autonomy, but it's a blunt instrument: it auto-approves every agentmemory tool, including genuinely destructive ones like memory_governance_delete and memory_slot_delete.

    This option is also not available when Droid defers multiple mcp tool calls and presents them all to me for approval.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions