mcp

MCP Adapter Boundary

internal/mcp implements the read-only MCP request handling over SQLite. Stdio is the default transport; HTTP mode is a minimal private-pilot transport wrapper around the same request handler.

Rules:

• read from the SQLite store surfaces; do not call Gong directly
• keep tools read-only
• require --db at the cmd/gongmcp boundary
• support MCP tool presets and allowlisting through gongmcp --tool-preset, GONGMCP_TOOL_PRESET, --tool-allowlist, or GONGMCP_TOOL_ALLOWLIST; when unset, stdio serves the full read-only catalog
• require an explicit preset or allowlist for all HTTP deployments
• generate tool schemas after server options are applied, so tools/list reflects the active allowlist and configured LimitPolicy
• keep row caps configurable through LimitPolicy, GONGMCP_MAX_* env vars, and gongmcp --max-* flags, but hard-cap them in code
• treat schema maximum as the configured ceiling; omitted limit values use the lower request defaults in limits.go unless the configured ceiling is lower
• return cap-feedback envelopes only for capped high-volume row tools that hit their effective limit; other capped aggregate/config tools may keep their normal result shape
• support private AI governance suppression through --ai-governance-config or GONGMCP_AI_GOVERNANCE_CONFIG; do not expose configured restricted names in MCP output, do not expose filtered-match counts, and require an explicit governance-compatible tool preset or allowlist
• keep browser/session auth separate from agent-client auth
• do not expose raw Gong API passthrough, arbitrary SQL, profile import, raw cached call JSON, or full transcript dumps
• return transcript snippets only, not full transcript bodies
• use search_transcripts_by_call_facts for scoped transcript evidence by date, lifecycle, scope, system, or direction; it may return bounded neighboring-segment excerpts, but must not return call IDs, titles, speaker IDs, or full transcript text
• use search_transcript_quotes_with_attribution when business users need bounded quote evidence with available Account/Opportunity attribution; call IDs, call titles, Account names/websites, and Opportunity names/close dates/probabilities require explicit opt-in flags, transcript_status=missing must fail closed because quote search starts from cached transcript segments, and person/contact titles must be reported as unavailable when not present in the cache
• control transcript evidence provenance with GONGMCP_TRANSCRIPT_EVIDENCE_PROVENANCE=redacted|alias|raw; redacted is default, alias returns stable call/speaker refs, and raw IDs still require explicit opt-in flags
• keep profile-aware tools tied to imported SQLite profile state
• return lifecycle source and profile provenance when profile-aware behavior is used
• keep unmapped CRM field output redacted by default
• treat search_crm_field_values as an explicit, bounded value lookup exception; call IDs are redacted unless include_call_ids=true, object IDs and names are always redacted, and value snippets plus call titles require include_value_snippets=true
• keep summarize_scorecard_activity aggregate-only by default; do not return call IDs, scorecard IDs, user IDs, answer text, call titles, transcript snippets, emails, raw JSON, or raw activity payloads
• serve profile-aware fact/lifecycle queries from the SQLite profile cache keyed by active profile and canonical hash; writable CLI commands warm it, and read-only MCP reports stale cache state instead of writing
• keep summarize_call_facts on safe business grouping dimensions; directed CRM value lookup belongs in search_crm_field_values

Practical checks:

go run ./cmd/gongctl mcp tools
GONGMCP_MAX_SEARCH_RESULTS=250 go run ./cmd/gongctl mcp tool-info search_transcript_segments
go test -count=1 ./internal/mcp ./cmd/gongmcp