# Agent Test Backend / Portal PoC

Browser-based PoC tester and hardcoded Portal Registry screen for:

```text
Browser -> Agent Test Backend -> MCP Server -> Tool Server
```

## Ports

```text
Tool Server        http://localhost:9092
MCP Server         http://localhost:8080/mcp
Agent Test Backend http://localhost:7070
```

## STS Run

Import this folder as an existing Gradle project:

```text
C:\Users\hyo\Documents\Codex\agent-test-backend
```

Run `com.example.agenttest.AgentTestBackendApplication`.

Default environment:

```text
AGENT_TEST_PORT=7070
MCP_ENDPOINT_URL=http://localhost:8080/mcp
MCP_PROTOCOL_VERSION=2025-11-25
TOOL_MANIFEST_URL=http://localhost:9092/tool-manifest
TOOL_SERVER_API_KEY=tool-server-key
PORTAL_REGISTRY_REVISION=1
PORTAL_ROUTE_KEY=external
TOOL_SERVICE_DOMAIN=http://localhost:9092
```

Open:

```text
http://localhost:7070
```

## Recommended Verification Order

1. Start Tool Server in STS.
2. Start MCP Server with Tool Server environment values.
3. Start Agent Test Backend.
4. Open `http://localhost:7070`.
5. Click `Initialize`, `Tools/List`, then run `Agent Chat` or `Tools/Call`.

## Portal PoC Scope

The current screen intentionally uses hardcoded registry data instead of DB tables:

```text
MCP Route     external -> http://localhost:8080/mcp
Tool Service  external-tools -> http://localhost:9092/tool-manifest
Mapping       external -> external-tools
```

The MCP-facing Portal Registry API is:

```text
GET http://localhost:7070/api/portal/registry/external
```

For this PoC, MCP can use these IntelliJ environment variables after applying
`C:\Users\hyo\Documents\Codex\mcp-portal-registry.patch` to the MCP project:

```text
MCP_DISCOVERY_ENABLED=true
MCP_PORTAL_ENABLED=true
MCP_PORTAL_ROUTE_KEY=external
MCP_PORTAL_REGISTRY_URL=http://localhost:7070/api/portal/registry/external
MCP_REGISTRY_REFRESH_INTERVAL_SECONDS=60
```

`POST /api/agent/chat` is a first Agent Backend skeleton. It currently uses a simple rule-based
planner so it can run without an OpenAI API key:

```text
"서울 날씨 알려줘" -> external.weather_lookup
"달러 환율 알려줘" -> external.exchange_rate
```

Later, replace the rule-based planner with Codex/OpenAI model selection while keeping the same MCP
`tools/list` and `tools/call` boundary.

## Portal Tool Bundle Management

The Portal now manages both seeded pull-discovery bundles through one API and UI flow:

```text
external-tools -> http://localhost:9092/tool-manifest
business-tools -> http://localhost:9090/tool-manifest
```

Bundle APIs:

```text
GET  /api/portal/bundles
POST /api/portal/bundles
PUT  /api/portal/bundles/{bundleId}
POST /api/portal/bundles/{bundleId}/sync
```

The API key is write-only. Read responses expose only `apiKeyConfigured`, and logs never contain
the key value. Manifest synchronization validates the configured Bundle ID, Tool name prefix,
required metadata objects, and the configured execution endpoint for every mapped Tool.

Business Tool endpoint mappings are owned by the Portal/MCP configuration rather than the
manifest:

```text
business.customer_search -> POST http://localhost:9090/internal/tools/customer-search
business.order_status    -> POST http://localhost:9090/internal/tools/order-status
business.ticket_create   -> POST http://localhost:9090/internal/tools/ticket-create
```

The Bundle definitions and synchronized snapshots are currently held in memory because this PoC
does not include a database dependency or an existing schema/migration framework. Therefore no DB
migration is required. Restarting the Portal restores the two seeded definitions and clears cached
manifest snapshots. For persistent operation, the `PortalBundleService` store is the boundary to
replace with the project's chosen repository and migration framework.

Start the business Tool server with pull discovery only:

```text
TOOL_AUTO_REGISTER_ENABLED=false
TOOL_HEARTBEAT_ENABLED=false
TOOL_SERVER_PORT=9090
```

After starting the Tool server, open the Portal and click `Sync Manifest` for `business-tools`.