AI agent server with MCP tools for memory, session management, chat transcripts, dynamic tool registration, and email sending
Phantom has 7 tools with mixed quality. Naming is strong (verb_noun pattern, clear intent). Descriptions are present and detailed (100-300 chars), exceeding the 10-char minimum. However, input schemas lack proper JSON Schema type declarations, parameters are documented in descriptions but not formally typed in the schema objects. Output schemas are undocumented. Error handling is minimal; tools do not guide recovery or categorize errors. The send_email tool exposes implementation details (Resend, idempotency keys) that belong in internal logic, not the interface. Memory and session tools are read-only and well-scoped, but lack pagination guidance despite returning lists.
Search or list the durable transcript for a Phantom web chat session. Use this after Murph compaction when an older detail from the current chat is missing from provider context. - session_id: the current Phantom chat session id from the <phantom_chat_context> block. - query: optional search text. If omitted, returns recent committed transcript entries. - role: "user", "assistant", or "all". Default is "all". - before_seq / after_seq: optional transcript sequence window. - limit: max results. Default 10, max 50. Returns compact, cited snippets with seq, role, created_at, status, and attachment metadata. Output is redacted for credentials, magic-login tokens, and large base64 payloads. This is operational transcript recovery, not long-term user memory.
List all dynamically registered tools.
List recent conversation sessions the agent has had, with channel, start time, turn count, and total cost. Use this to anchor reflective observations to specific days and channels, to count interactions across a window, or to pick the most expensive recent session for a cost explanation. Returns the most recently active sessions first. - limit: max sessions to return. Default 20, max 200. - days_back: optional. Only sessions active within the last N days. - channel: optional. Substring filter on channel_id (e.g. "slack" to get all slack sessions). Each row has session_key, channel_id, conversation_id, status, total_cost_usd, turn_count, created_at, last_active_at.
Search the agent's persistent memory for past sessions, topics, and facts. Supports semantic search and temporal filtering. Use this when reflecting on past behavior, looking up prior conversations, finding patterns, or checking whether a question has already been resolved. Returns episodes (past conversation turns with outcomes), facts (learned knowledge about the user, team, codebase), and optional procedures. - query: the semantic search text. For temporal scans, use a broad query like "this week" or the topic keyword. - memory_type: "episodic" for past sessions, "semantic" for facts, "all" for both. Default is "all". - days_back: optional. Limits results to items from the last N days. - limit: max results per type. Default 10, max 50. Each episode includes summary, detail, outcome, started_at, tools_used, and lessons. Facts include natural_language, category, confidence, and valid_from.
Input schemas lack formal JSON Schema type declarations. Parameters are documented in descriptions but not typed in the schema objects (e.g., 'query' is described as string but schema shows only description, no 'type' field). This prevents LLM validation and forces reliance on description text alone.
Output schemas are not documented. Tools return complex objects (episodes, facts, sessions, transcripts) but LLMs have no formal schema to understand field types, nesting, or required fields. This forces LLMs to infer structure from descriptions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Register a new dynamic MCP tool. The tool is persisted and survives restarts. For shell handlers, provide handler_code with a bash command. For script handlers, provide handler_path with a path to a script file. Tool input is available via the TOOL_INPUT environment variable (JSON string).
Remove a previously registered dynamic tool. Built-in tools cannot be removed.
Send transactional email via Resend with cost-attribution tags, tenant-salted idempotency key, recipient-policy gate, and Prometheus counter. Supports daily rate limiting, recipient policy enforcement, and structured error kinds.
send_email exposes implementation details (Resend provider, idempotency keys, tenant-salting, Prometheus counters) in the description. These belong in internal logic, not the tool interface. The description should focus on what the user/agent needs: send transactional email with optional CC/BCC/reply-to.
List tools (phantom_list_sessions, phantom_list_dynamic_tools) lack pagination guidance. phantom_list_sessions accepts 'limit' (default 20, max 200) but does not document how to fetch the next page (offset, cursor, or sequential calls). phantom_list_dynamic_tools has no limit parameter at all.
Error handling is absent. Tools do not document error conditions, recovery steps, or error categorization (retryable vs. user-fixable vs. fatal). An LLM calling phantom_register_tool with an invalid name has no guidance on what went wrong or how to fix it.