Agent vendors
How do I put Inhold behind my agent?
Point your agent at Inhold over MCP or REST, run every call under one purpose, and your customer's reviewer gets one audit row per decision. The memory stays in their account.
Inhold is being built. No customers yet. The AWS install's Terraform passes CI checks but has not been run in a real AWS account. Repo goes public 9 Nov 2026.
What you get
Six tools and a policy check outside the model
Six MCP tools
context_search, context_write, context_get, context_whoami, context_request_access and skill_list. The purpose is fixed when the session starts, so a prompt cannot change it.
Deny by default
Every read, write and delete is checked on caller, purpose and data label in Inhold's code, not in the model. A refused id and a missing id look the same.
One audit row per decision
Allowed or denied, hash-chained and anchored. Your customer's auditor exports the evidence pack; you point their reviewer at it.
Corrections become skills
A nurse's correction is stored with a diff and becomes a draft skill; a different person signs it off before any agent uses it.
Connect
Your SDK, a laptop, about 15 minutes
Claude Agent SDK
Connect a Claude agent over MCP, with the sign-in options and their limits.
LangGraph
Connect a LangGraph agent: read one allowed and one refused note, then the two audit rows.
OpenAI Agents SDK
Connect an OpenAI Agents SDK agent the same way.
Anything with HTTP
The quickstart uses curl against the REST API and ends with an exported pack.
In your deal
What changes when the reviewer asks
Where does the agent's memory live?
In the hospital's own account, on their key. The data map answers it line by line.
Who read it?
The audit log, and the reviewer can check it without trusting you or us. The 20 reviewer answers carry a status each.
What is not built
Vector search, an Azure install and an outside security review. The AWS install's Terraform passes CI checks but has not been run in a real AWS account. Everything else is on the gap list.