MCP server for outbound phone calls and SMS. Accepts contact names or E.164 numbers. Provider: TELEPHONY_PROVIDER=asterisk (default, uses Asterisk ARI + speechops TTS) or TELEPHONY_PROVIDER=twilio. All calls logged to SQLite audit DB.
Telephony Bridge has clear, action-oriented tool names (make_call, send_sms, add_contact) and reasonable descriptions (avg ~150 chars). However, schemas are incomplete: most tools lack explicit JSON Schema definitions in the visible code. Parameter descriptions exist but are sparse (e.g., 'Contact name or E.164 number' lacks format/validation hints). Error handling is minimal, no recovery guidance or categorization. Output schemas are undocumented. The server follows a sensible composition pattern (separate tools for calls, SMS, contacts) but lacks the rigor expected of production tools. No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels.
Add or update a contact in the phone book.
List all contacts in the phone book.
Place an outbound phone call and speak a message via TTS. Accepts a contact name from the contacts store ("Steve", "Marion") or a raw E.164 number (+43664XXXXXXX). The message is synthesised via speechops (Gemini TTS) with a pre-recorded WAV fallback.
Trigger an automated emergency call (backward-compatible wrapper for make_call).
Remove a contact from the phone book.
Send an SMS to a contact or E.164 number. Note: Requires TELEPHONY_PROVIDER=twilio. Asterisk SIP SMS is trunk-dependent.
No explicit JSON Schema definitions visible for tool inputs. Schemas are inferred from function signatures but not formally declared in MCP tool registration. This prevents clients from validating inputs before sending and makes parameter constraints opaque to LLMs.
Output schemas are completely undocumented. Tools return dicts with fields like 'success', 'call_id', 'error', 'contact_name', 'to_number', but LLMs have no formal specification of what to expect. This forces LLMs to infer structure from examples.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | <=2025-11-25 | v2 |
Dry-run verification call to a local SIP URI (e.g. Linphone). Sends a fixed German test message. Does not require a real SIP trunk.
Return current telephony configuration, provider readiness, and audit stats.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. remove_contact is DESTRUCTIVE, list_contacts and telephony_status are READ_ONLY, but these are not declared in the tool metadata. Agents cannot reason about safety without explicit hints.
Error handling lacks recovery guidance. When contact resolution fails, the tool returns {'success': False, 'error': '...'} but does not suggest next steps (e.g., 'Try list_contacts() to see available names'). Errors are not categorized as retryable, user-fixable, or fatal.
Parameter descriptions lack format/validation hints. 'Contact name or E.164 number' does not specify E.164 format ('+' prefix, country code, etc.). 'Text to speak' does not mention language support or length limits. LLMs cannot validate inputs without explicit constraints.