diff --git a/.changeset/unified-mcp-layer.md b/.changeset/unified-mcp-layer.md
new file mode 100644
index 0000000..66fba53
--- /dev/null
+++ b/.changeset/unified-mcp-layer.md
@@ -0,0 +1,14 @@
+---
+"@timothycrooker/ai-context-core": minor
+"@timothycrooker/ai-context-cli": minor
+"@timothycrooker/ai-context-templates": patch
+---
+
+Add a unified MCP layer — MCP is now a third generated primitive alongside context and skills.
+
+Declare MCP servers once in `.ai/mcp.json`; `ai-context build` fans them out to each agent client's native config (Claude `.mcp.json`, Codex `.codex/config.toml`). Servers can carry backing: a linked skill (auto by co-name or explicit) and a one-line catalog entry in `AGENTS.md`/`CLAUDE.md` so an agent gets the tool and the knowledge to use it.
+
+- `project`-scope servers are committed; `user`-scope servers install per-machine via `ai-context mcp install <name> --user`.
+- Secrets stay as `${VAR}` references (resolved from `.ai/secrets.local.env`); `ai-context verify` fails on a credential literal in a generated config.
+- New CLI: `ai-context mcp list | install --user | setup`.
+- v1 ships `claude` + `codex` adapters behind a pluggable registry.