Skip to content

v3.6.2 — Richer tool descriptions for better AI tool-calling

Choose a tag to compare

View all tags
@bvisible bvisible released this 09 Jun 10:05
· 55 commits to main since this release
v3.6.2
333e26f

Richer tool descriptions — better AI tool-calling

All 37 MCP tool descriptions have been rewritten to document what each tool actually does, so an AI agent knows the consequences before invoking it.

What changed

Each description now states:

  • Side effects and whether the tool mutates remote/local state or is read-only
  • Destructive behavior — e.g. ssh_db_import and ssh_backup_restore DROP/overwrite data; ssh_db_query is strictly SELECT-only
  • Idempotency, sudo/auth requirements, and security-mode gating (which tools are blocked on readonly/restricted servers)
  • Parameter semantics — mutually-exclusive params (e.g. tunnelId vs server in ssh_tunnel_close), defaults, and edge cases

Average description length went from ~6 words to ~70 (shortest is now 55 words, no weak outliers).

Why

  • Better real tool-calling: agents make fewer mistakes when the contract is explicit.
  • Raises the Glama Tool Definition Quality score (was C — avg 3.1/5, lowest 2.1/5).

Safety

No behavioral change. Only description strings were edited — all logic, parameters, and tool names are unchanged. Every test suite (test:policy, test:config, test:lifecycle, …) passes unmodified, and tools/list was verified end-to-end to expose all 37 enriched descriptions.

Full changelog: https://github.com/bvisible/mcp-ssh-manager/blob/main/CHANGELOG.md#362---2026-06-09

Assets 2
Loading