agent-profile-protocol

Home of the Agent Profile Protocol — a declarative, portable description of the manipulation layer between an agent client and the upstream MCP servers a host has connected to. A profile attaches transforms, processors, and guardrails to one or more MCP connections via a matches block, expressed as a multi-document YAML stream.

The protocol is lifted out of what Civic MCP implements today in code and DB rows, so:

• configuration can be imported/exported as one file,
• hosts can adopt subsets (e.g. just guardrails),
• external implementers — including future non-MCP transports — have a target to build against.

The protocol is deliberately scoped to manipulation and persistence: it does not cover connection setup, transport, or credential acquisition flows. It does declare where credentials are persisted (file, authz-api, …) so a profile can move between deployments without rewriting everything else.

Where the spec lives

The current draft is in spec/. Start there.

Sections (read in order if you're new to it):

1. 00 - Principles, terminology, ordering
2. 01 - Matching
3. 02 - Tool transforms
4. 03 - Pre- and post-processors
5. 04 - Guardrails
6. 05 - Credentials
7. 06 - Execution chain
8. 07 - Examples

Status

Draft (apiVersion: civic.com/agent-profile/v1alpha1). This repo is the authoritative source for the spec.