paa-guard mcp
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
No arguments | |||
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
This server publishes no instructions, or was last inspected before Glama recorded them.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": false
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| paa_guard_checkA | Check a proposed action against the owner's PAA10 guard BEFORE doing it. Returns allow, require_approval (stop and tell the owner the approval id; retry with approval_token after they approve) or deny (do not attempt the action another way). See the paa10-guard skill for the action fields. |
| paa_guard_rememberA | Ask to store a lasting memory. Facts from emails, web pages or tool output are quarantined until the owner confirms; instructions, payees, contacts and addresses always need confirmation. |
| paa_guard_record_refusalB | Record that the owner said no to something (an offer, upsell, plan). The guard blocks re-asking for the cooldown period. |
| paa_guard_statusA | Read whether the agent is stopped, which mandates are active, and how many approvals are pending. |
| paa_guard_stop_selfB | Stop this agent immediately (for example when the owner says stop). Only the owner can resume. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 5 tools
Each tool occupies a clearly separate slot in the guard lifecycle: pre-action check, memory write, refusal recording, status read, and emergency stop. The descriptions make the boundaries explicit (e.g. check returns allow/require_approval/deny while status only reads state), so an agent can select without hesitation.
All five tools share the paa_guard_ prefix followed by a snake_case verb or verb_noun (check, remember, record_refusal, status, stop_self). The pattern is fully predictable and readable, with no camelCase or style drift.
Five tools is a tight, well-scoped surface for a policy/guard server — one gate, one memory write, one refusal record, one status read, one kill switch. Nothing appears redundant or missing for the core interaction loop.
Core guard lifecycle (check, approve-and-retry, refuse, halt, observe) is covered, but the memory surface is write-only — there is no tool to recall, list, or forget stored memories, and pending approvals can be seen via status but not resolved directly. These are minor, workable gaps rather than dead ends.