🔌MCP Tools (28)

The 28 tools available to AI agents (read 12 / write 13 / approval 3) and how strict mode works.

The 28 tools available to AI agents via the MCP server, split into three namespaces: read / write / approval — enabling dashboard-equivalent configuration changes. The default transport is stdio (HTTP + Bearer also available).

Breakdown (read 12 / write 13 / approval 3 = 28 total)

read

Tool	What it does	Implementation
`read_list_clients`	List accessible clients	`backend/mcp/tools/clients.py`
`read_get_client_overview`	A client's basic settings + integration policy + protection_enabled	`backend/mcp/tools/clients.py`
`read_list_whitelist`	List whitelist (allowed numbers)	`backend/mcp/tools/whitelist.py`
`read_list_blocklist`	List blocklist (blocked numbers)	`backend/mcp/tools/blocklist.py`
`read_list_auth_keywords`	List auth keywords	`backend/mcp/tools/auth_keywords.py`
`read_list_dictionary`	List the STT correction dictionary	`backend/mcp/tools/dictionary.py`
`read_get_engine_config`	Verdict engine config (thresholds, keyword weights)	`backend/mcp/tools/engine_config.py`
`read_list_categories`	List global spam categories	`backend/mcp/tools/categories.py`
`read_list_client_category_configs`	List per-client category overrides	`backend/mcp/tools/categories.py`
`read_list_pending_approvals`	List pending change approvals	`backend/mcp/tools/approvals.py`
`read_get_change_approval`	A change approval's detail (with diff)	`backend/mcp/tools/approvals.py`
`read_list_integration_audit`	List integration audit events	`backend/mcp/tools/audit.py`

write

Tool	What it does	Implementation
`write_add_whitelist_entry`	Add a whitelist entry (loosening; approval required under strict)	`backend/mcp/tools/whitelist.py`
`write_remove_whitelist_entry`	Remove a whitelist entry	`backend/mcp/tools/whitelist.py`
`write_add_blocklist_entry`	Add a blocklist entry	`backend/mcp/tools/blocklist.py`
`write_remove_blocklist_entry`	Remove a blocklist entry (loosening; approval required under strict)	`backend/mcp/tools/blocklist.py`
`write_add_auth_keyword`	Add an auth keyword (loosening; approval required under strict)	`backend/mcp/tools/auth_keywords.py`
`write_remove_auth_keyword`	Remove an auth keyword	`backend/mcp/tools/auth_keywords.py`
`write_add_dictionary_entry`	Add a dictionary entry	`backend/mcp/tools/dictionary.py`
`write_remove_dictionary_entry`	Remove a dictionary entry	`backend/mcp/tools/dictionary.py`
`write_update_engine_config`	Update engine config (threshold-lowering patches are loosening; approval required under strict)	`backend/mcp/tools/engine_config.py`
`write_set_client_category_override`	Set a per-client category override (allowing is loosening)	`backend/mcp/tools/categories.py`
`write_remove_client_category_override`	Remove a per-client category override	`backend/mcp/tools/categories.py`
`write_set_protection_enabled`	Toggle protection on/off (turning off is loosening; approval required under strict)	`backend/mcp/tools/clients.py`
`write_update_client_overview`	Update a client's integration policy etc. (approval required under strict)	`backend/mcp/tools/clients.py`

approval

Tool	What it does	Implementation
`approval_propose_change`	Propose a change (the re-entry point for writes 409'd under strict)	`backend/mcp/tools/approvals.py`
`approval_approve_change`	Approve a proposal (done by an operator other than the proposer)	`backend/mcp/tools/approvals.py`
`approval_apply_change`	Apply an approved proposal (CAS prevents double application)	`backend/mcp/tools/approvals.py`

Strict mode (enforced two-person approval)

Strict mode (clients.strict_approval_mode, per client) is OFF by default. While off, writes pass through (legacy behavior). For a client with it ON, only loosening two-person-approval changes are rejected by the backend with a 409, making the three steps mandatory: propose → approval (by an operator other than the proposer) → apply. The MCP layer converts the 409 into "the next action to take + a ready-made propose payload", so an AI agent can call approval_propose_change directly.

ℹ️The gist

Off by default. For a client with it ON, loosening changes require the three steps: propose → approval (by a different operator) → apply. Every write is tracked by the AuditLogger, so even with strict mode off, who changed what and when remains auditable after the fact.