There is no remote (HTTP) MCP endpoint today. The server runs on your machine. In Claude Code, the plugin starts it for you.
Connect
1
Get the server
The server ships inside the Claude Code plugin repository. Clone it and note the absolute path of
plugins/agentic-id/server/mcp.mjs.2
Add it to your client
- Claude Code
- Generic JSON config
3
Check the connection
Ask your agent to call
agentic_id_status. It reports the server address, the workspace and whether contact values are masked.Authentication
The server sends one credential as a Bearer token. It picks the first that exists:
An API key, when set, is used instead of a sign-in.
Sign-in is OAuth 2.1 with PKCE. The browser opens the sign-in page, and the session is stored in
~/.agentic-id/session.json, readable by you only. The access token lasts 1 hour and is renewed with a refresh token that lasts 30 days. agentic_id_login with action: "logout" ends the session.
Environment variables
How answers look
Each tool result has a text block (a one-line summary, then the full answer as JSON) and the same answer as structured content. A failed call is markedisError and carries this body:
"sent": false means the server refused the call on your machine. Nothing reached Agentic ID and nothing was charged. See Errors.
Every paid call is sent with an idempotency key, so the server can retry a dropped connection without charging twice.
Available tools
Fifteen tools. An argument the tool does not have is refused with the list of allowed names.
The
cost block in each answer is the authority on what was charged. See Credits.
Shared parameters
The most the call may charge, 0 to 100000. A call that could cost more than 10 credits is refused without it.
Why you are asking. One of
b2b_outbound, ad_audience, own_customer_enrichment, fraud_check. Defaults to b2b_outbound. See Data and privacy.What you know about the person or company:
email, linkedin_url, name, company_name, company_domain. One strong identifier is enough. A name needs a company or another identifier with it.agentic_id_validate
agentic_id_resolve
agentic_id_resolve_many
Rows that resolve to the same ID are grouped in
same_id.
agentic_id_enrich
Phone fields are refused.
agentic_id_find
A search needs
roles or seniority.
agentic_id_discover
Keywords rank results. They filter only with
keywords_required.
agentic_id_lookalike
agentic_id_facts and agentic_id_ledger
agentic_id_workspace
agentic_id_report
agentic_id_list
The rows are read and written by the server and do not enter the conversation.
agentic_id_login
Returns a sign-in link at once. Call it again after signing in to confirm.
agentic_id_estimate
Worked out on your machine. Nothing is sent.
agentic_id_status
No parameters.What the server tells agents about spending
The server sends instructions to the agent when it connects. On spending they say:- These calls are not charged: status, workspace, facts, ledger, estimate, report, login, discover with
count_only, list withdry_run. Every other call costs credits. - Before any call that could cost more than 10 credits, the user must have stated a number of credits. “Go ahead” is not a budget. The agent shows the arithmetic (an estimate or a count) and asks for a number.
- The agent never picks
max_creditsabove what the user said. If the user gave a budget, it is used without asking again. max_creditsis set on every paid call. Afterwards the agent says what was charged, fromcost.charged_credits.- For “what did I spend”, the agent calls
agentic_id_workspace. That is the billing record.
max_credits is refused with cap_required and the estimate.
