Solution architecture outcome pattern: agent and its MCP servers
Some checks failed
check agents / check (push) Failing after 13s

The cp1-solution-architecture-agent and the sa-aws/azure/gcp-cloud MCP servers, a separate
stream from the individual and sub-agent patterns.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
2026-11-05 22:13:25 +00:00
parent 8bb62a4a42
commit bc84ee4df6
267 changed files with 44287 additions and 45 deletions

View File

@@ -0,0 +1,54 @@
# Solution Architecture MCP servers
One server per hyperscaler, used only by the Solution Architecture agent. Each carries the tool set of `gcp_solution_architecture_agent` (its `providers/gcp.yaml`), like for like, so the agent doesn't depend on the shared servers in the repository's root `mcp-servers/`.
| Server | Port | Tools | How it reaches the cloud |
|---|---|---|---|
| [`sa-gcp-cloud`](sa-gcp-cloud/) | 8115 | 44 | Forwards to the same 12 hosted Google MCP servers as the source, under the same allow lists |
| [`sa-azure-cloud`](sa-azure-cloud/) | 8116 | 40 | Microsoft Learn, Resource Graph and ARM GET |
| [`sa-aws-cloud`](sa-aws-cloud/) | 8117 | 44 | The AWS documentation server, and read-only API calls through the customer's role |
## How the Google tools work
Each tool is `gcp_<server>_<tool>`, such as `gcp_compute_list_instances`, and takes `tool_arguments`, passed through unchanged. `gcp_describe_tool` returns a hosted tool's live description and argument schema. A tool must be on the allow list and must declare `readOnlyHint` when it's called, or it's refused.
## Azure and AWS counterparts
The Azure and AWS tools keep the source's tool names where they can. The match is by capability, not by behaviour:
| Source | Azure | AWS |
|---|---|---|
| `answer_query` | None | None |
| `search_projects` | Subscriptions | Organization accounts |
| `analyze_iam_policy` | Role assignments | Account authorization details |
| Cloud Run | Container Apps, App Service | App Runner, plus Lambda |
| Cloud SQL | SQL, PostgreSQL, MySQL servers | RDS |
| Cloud Storage | Storage accounts | S3 |
| Pub/Sub | Service Bus, Event Hubs, Event Grid | SNS |
| Deny policies | Deny assignments | Service control policies |
| Connectivity tests | Connection monitors | Reachability Analyzer paths |
| Commitments | Reserved capacity orders | Savings Plans |
Azure has no `get_subscription` or `get_deny_policy`, because Event Grid subscriptions and deny assignments come through a subscription-wide list.
## Rules
- **Free-text documentation search is an accepted exception** to the root `mcp-servers/README.md` rule. Queries go to the vendor as written and aren't scrubbed. Each one is logged (`docs_query`). Replies are scrubbed as usual.
- **Read-only.** Don't add a tool that creates, changes or deletes anything.
- **Store each server's connection per project, never tenant-wide.** It holds the customer's credential.
- **Don't edit `<server>/server.py` or `*.server.yaml` by hand.** They're generated.
## Working on a server
The sources are `build_definitions.py` (the tool tables), `_support/`, and the shared `../../../mcp-servers/_common/`, which these servers reuse and never change.
```sh
python3 build_definitions.py
for n in sa-gcp-cloud sa-azure-cloud sa-aws-cloud; do
python3 ../../../scripts/render_mcp_server.py $n.server.yaml --out $n \
--platform ~/platform-engineering/kyndryl-agent-builder/modernization-platform
SERVER_DIR=$PWD/$n python3 -m pytest tests/test_servers.py
done
```
Rendering needs `pydantic`, `pyyaml`, `httpx`, `pymongo` and `jinja2`. Testing also needs the packages in each server's `requirements.txt`, and `pytest`.