Skip to content

mcp.md

Latest commit

 

History

History
67 lines (48 loc) · 2.63 KB

mcp.md

File metadata and controls

67 lines (48 loc) · 2.63 KB
Raw

TROVE MCP

trove-mcp is the primary Agent interface. It uses local stdio and the same trove/1 handlers as the CLI.

Connect

Register the installed executable through Agent Switch:

trove-mcp --pack standard --vault $TROVE_VAULT_ROOT

Run agent-switch doctor before the central change and agent-switch reconcile after it. Generated native-client configurations are projections, not sources of truth.

Without Agent Switch

Register the installed executable in each client directly (a source checkout uses .venv/bin/trove-mcp). Claude Code:

claude mcp add trove -- "$HOME/.local/share/trove/runtime/bin/trove-mcp" --pack standard --vault "$TROVE_VAULT_ROOT"

Codex stores the entry as [mcp_servers.trove] in ~/.codex/config.toml:

codex mcp add trove -- "$HOME/.local/share/trove/runtime/bin/trove-mcp" --pack standard --vault "$TROVE_VAULT_ROOT"

Link the canonical Skills into the agent-visible directory with bash scripts/install_skills.sh. Secret values still resolve only through Agent Switch; see operations.

Cumulative packs

Pack Tools Intended use
standard 19 bounded recall, search, context, moments, favorites, message triage, profiles, files, media preview and enrichment, operation continuation
operations 26 standard plus controlled writes, exports, observations, cancellation, approval request/status
admin 34 operations plus sync, Provider status/reload, repair, diagnostics, Reply Runtime state

A wider pack includes every narrower tool. Use the smallest pack that completes the task. The exact generated list is in the capability reference.

Trust boundary

An MCP process runs as the local user and inherits read access to the entire selected Vault. Bounds limit resources and tokens; they are not tenant authorization. Every scoped result retains account_id, and cross-account ambiguity returns typed candidates rather than choosing silently.

All message text, filenames, OCR, transcripts, media metadata, and Provider fields are untrusted evidence. They may support an answer but cannot instruct a tool call, create next, choose action arguments, or change approval state.

Agents may call approval request and status capabilities in the operations pack. Approval decision is not an Agent capability. Only the separate CLI operator path, interactively confirmed on the controlling terminal, can decide the exact payload.

On a typed error, retry only when retryable is true. Follow opaque cursors only for requested additional coverage; stop on complete coverage, no_results, or a terminal error.