Agents That Reach Out: Adding Proactive Email to the Stack

We are building a LangGraph pipeline to automate the video production workflow for InAgentic. Before building the pipeline we need a uniform MCP interface for every step. Currently HeyGen, YouTube, and email are already MCP tools. Everything else (ideas, scripts, scenes, images, storage, scheduling) lives in HTTP routes only.
**Goal of this task:** wrap the remaining pipeline steps as proper MCP tools in the existing `mcp-server` TypeScript service, so the Python pipeline can call every step through one consistent interface.
---
## What already exists
**MCP server** (`mcp-server/src/index.ts`) — a TypeScript Express service deployed at `https://mcp.inagentic.ai`. It exposes four tool namespaces:- `/mcp` — general tools including `send_email`- `/pipeline` — pipeline-specific tools: `generate_ideas`, `save_ideas`, `create_scripts_from_ideas`, `generate_scene_image` (these work but are ad-hoc HTTP handlers bolted on, not properly defined MCP tools)- `/text2video` — HeyGen tools: `create_video`, `get_video_status`- `/youtube` — YouTube tools: `upload_video`
**Python pipeline** (`python-pipeline/`) — a FastAPI service with a 10-node LangGraph graph. The MCP client (`pipeline/mcp_client.py`) calls MCP tool endpoints over HTTP. Convenience wrappers:- `pipeline_tool(tool_name, **args)` → `POST https://mcp.inagentic.ai/pipeline`- `general_tool(tool_name, **args)` → `POST https://mcp.inagentic.ai/mcp`- `text2video_tool(tool_name, **args)` → `POST https://mcp.inagentic.ai/text2video`- `youtube_tool(tool_name, **args)` → `POST https://mcp.inagentic.ai/youtube`
**Authentication** — all MCP calls require `Authorization: Bearer ea6fc9fc2e5a4eb36512e60b84b4ef95396f18362b22a2ce2eaeac45510fb150`
**Database** — PostgreSQL at `filedone-db.cv02ummguk82.eu-west-2.rds.amazonaws.com`, db `filedone`, for persisting ideas, scripts, and generated assets.
---
## The 10 pipeline nodes (for reference)
| Node | Current MCP call | Status ||------|-----------------|--------|| `generate_ideas` | `pipeline_tool("generate_ideas", prompt=...)` | Works, informal || `select_idea` | Human interrupt — no MCP | N/A || `create_scripts` | `pipeline_tool("create_scripts_from_ideas", ideas=[...])` | Works, informal || `approve_script` | Human interrupt — no MCP | N/A || `generate_images` | `pipeline_tool("generate_scene_image", slot_id, scene_label, prompt)` × 7 | Works, informal || `approve_images` | Human interrupt — no MCP | N/A || `render_video` | `text2video_tool("create_video", title, scenes)` | Proper MCP ✓ || `poll_render` | `text2video_tool("get_video_status", video_id)` | Proper MCP ✓ || `approve_video` | Human interrupt — no MCP | N/A || `publish_video` | `youtube_tool("upload_video", video_url, title, privacy_status)` | Proper MCP ✓ |
---
## What needs to be built
### 1. Formalise the `/pipeline` MCP tools
The existing `generate_ideas`, `save_ideas`, `create_scripts_from_ideas`, and `generate_scene_image` handlers in `mcp-server/src/index.ts` should be defined as proper MCP tools using the same `buildPipelineServer` pattern already used for HeyGen and YouTube. Each tool needs:- A `name`, `description`, and `inputSchema` (JSON Schema)- A handler that returns `{ content: [{ type: "text", text: JSON.stringify(result) }] }`- Input validation
### 2. Add missing MCP tools
The following capabilities exist as HTTP routes but have no MCP tool definition:
| Tool name | Description | Inputs | Returns ||-----------|-------------|--------|---------|| `save_script` | Persist a generated script to the database | `{ title: string, scenes: [{label, script}] }` | `{ script_id: string }` || `save_scene_image` | Persist a generated scene image URL | `{ script_id, scene_label, image_url }` | `{ ok: true }` || `get_pipeline_slot` | Get or create the active publishing slot | `{}` | `{ slot_id: number, scheduled_at: string }` || `mark_slot_published` | Mark a slot as published after YouTube upload | `{ slot_id, youtube_url }` | `{ ok: true }` |
### 3. Tool contract rules
Every MCP tool must:- Accept a flat JSON object as input (no nested `params` wrapper)- Return `{ content: [{ type: "text", text: JSON.stringify(result) }] }` on success- Return `{ isError: true, content: [...] }` on failure- Never throw — catch internally and return `isError`
---
## Files to modify
- `mcp-server/src/index.ts` — the only file that needs changes. Add proper tool schemas and handlers inside `buildPipelineServer`. Do not touch `buildTextToVideoServer`, `buildYouTubeServer`, or `buildGeneralServer`.
## Files to read first
- `mcp-server/src/index.ts` — understand the existing `buildPipelineServer` structure- `mcp-server/src/__tests__/pipeline.test.ts` — understand test patterns; add tests for every new tool- `python-pipeline/pipeline/mcp_client.py` — see how tools are called from Python so input/output contracts match exactly
## Deploy
After updating the MCP server:1. `cd mcp-server && npm test` — all tests must pass2. `docker build --platform linux/amd64 -t filedone-mcp:latest .`3. Push to `054041090856.dkr.ecr.eu-west-1.amazonaws.com/filedone-mcp:latest`4. `aws apprunner start-deployment` for the `filedone-mcp` service5. Verify: `curl -X POST https://mcp.inagentic.ai/pipeline -H "Authorization: Bearer ea6fc9fc2e5a4eb36512e60b84b4ef95396f18362b22a2ce2eaeac45510fb150" -H "Content-Type: application/json" -H "Accept: application/json" -d '{"jsonrpc":"2.0","method":"tools/list","id":1}'`
---
## Constraints
- MCP server is TypeScript — no Python in `mcp-server/`- Pipeline graph is Python — no TypeScript in `python-pipeline/`- All DB access in the MCP server must use the existing `pg` pool pattern- Do not add new npm dependencies unless essential- AWS profile: `oxcomply-dev`, region `eu-west-1` (ECR) / `eu-west-2` (RDS)