Recon exposes six read-only tools via its MCP server. Your AI tool calls these directly; you don't need to call them yourself. This is the reference for anyone writing custom prompts or building integrations on top of Recon.
get_account
Full memory profile for one customer.
When to use: any "tell me about X" question where X is a specific customer.
Input: account (identifier, UUID, or canonical name).
Returns: synthesized claims grouped by lens, recent observations with source links, derived memory_state (unknown / known / at_risk / archived), and risk_score (higher = more risk).
list_accounts
Ranked, slim list of accounts (not full detail).
When to use: triage. "Who's at risk?", "List my biggest customers", "Which accounts are most active this week?"
Input: optional memory_state filter, optional lens filter, sort (risk default, recent, claims), limit (default 20).
Returns: total, returned, and an array of rows with memory_state, active_claim_count, risk_score. Sorted by risk by default so the accounts that need attention come first.
search
Full-text search across observations and claims.
When to use: "Has anyone mentioned X?", "Which customers asked about Y?", "Find the conversations where Z came up."
Input: query (free text), optional account_id, lens, predicate (e.g. reported_by for "who reported X", admin_for for admin lookups), limit.
Returns: bounded matching rows. Not a whole-account picture; for that follow up with get_account.
impact
For a given entity (feature, code path, ticket id, Sentry issue), return the customer accounts affected.
When to use: "Who's affected by REC-447?", "Which customers use the export feature?", "Cross-customer impact of this change."
Input: entity_type (linear, sentry, code_path, feature, etc.), identifier (the ID or path).
Returns: accounts only (not the underlying observations), ranked by observation count. Follow up with get_account for any one account's detail.
themes
Cross-account claim clusters (LLM-backed, slower).
When to use: "What are customers asking for?", "Common pain points across the customer base", "Themes in feedback this month."
Input: optional lens.
Returns: themes spanning >= 2 accounts. Each theme has account_count, claim_count, top accounts, sample quotes, and a claim_ids array. Does NOT return every quote inline. Pass claim_ids to theme_evidence to hydrate.
theme_evidence
The detail companion to themes. Hydrates the full receipts behind a theme.
When to use: only after themes, when you need the actual quotes behind a specific theme.
Input: claim_ids (the array a themes call returned). Bounded to 100 IDs per call.
Returns: claim text, source observation IDs, citation counts, owning accounts.
Tool selection cheat sheet
| Question shape | Tool |
|---|---|
| "Tell me about Acme" | get_account |
| "Who's at risk?" | list_accounts (default risk sort) |
| "Which customers mentioned X?" | search |
| "Who's affected by bug Y?" | impact |
| "What are customers asking for?" | themes |
| "Show me the quotes for theme T" | theme_evidence |
What Recon does not do
The MCP tools are read-only. Recon does not modify any of your data. Recon does not call back out to your other systems on behalf of your AI tool. It returns its own memory. Recon does not run an AI loop to answer your AI tool's questions; it returns structured JSON and your AI tool synthesizes the answer.
Write actions (creating tickets, drafting emails, posting Slack messages) are a Team-tier feature and live outside the MCP surface, behind the Recon agent layer with explicit per-action approval.