Privacy modes
Control which recognized AWS identifiers are tokenized before results reach your model.
How tokenization works
The local MCP replaces recognized AWS identifiers with stable local tokens such as SG_001 and S3_001. Later calls expand tokens locally, call the backend, and tokenize again. The mapping stays on your machine and is not sent to Emfirge or the model.
LLM sees: "SG_001 reaches S3_001"
local mapping: SG_001 = <local security-group ID>
S3_001 = <local bucket ID>
The three modes
strict (default) tokenizes recognized AWS IDs, ARNs, account IDs, names, and IPs. balanced tokenizes common identifiers while retaining selected context. off leaves identifiers raw.
Change the mode
Use the CLI or set EMFIRGE_PRIVACY in client configuration.
npx @emfirge/mcp privacy
npx @emfirge/mcp privacy strict
Backend and report URLs
The backend receives data required for analysis. Credentials in presigned report URLs are scrubbed. Derived graph labels may not be recognized identifiers, so tokenization is not a completeness guarantee.
The three modes
strict (default) tokenizes recognized AWS IDs, ARNs, account IDs, resource names, and IP addresses; it is the strongest default boundary for sensitive transcripts.
balanced tokenizes common identifiers such as ARNs, compute/network/IAM IDs, IPs, and account IDs while retaining selected infrastructure context; it balances context and minimization.
off leaves identifiers un-tokenized so they may reach the LLM as returned; use only when raw identifiers are acceptable.
Changing modes and client environment
Use npx @emfirge/mcp privacy to show the current mode or npx @emfirge/mcp privacy strict to set strict, balanced, or off. When configuring a client directly, set EMFIRGE_PRIVACY in the MCP client configuration.
npx @emfirge/mcp privacy # show current mode
npx @emfirge/mcp privacy strict # set strict, balanced, or off
"env": { "EMFIRGE_PRIVACY": "strict" }
Backend and report URLs
The backend receives the data required for the requested scan and analysis. Credentials embedded in presigned report URLs are scrubbed before URLs are returned or displayed. Do not share raw credentials or unreviewed transcripts. Derived graph labels or descriptive values may not be recognized identifiers, so tokenization is not a completeness guarantee; review outputs against organizational policy, especially in strict mode.