Todoist Daily Agent

A Cloudflare Worker + Next.js 16 reference implementation that plans a focused day with Workers AI, streams the intermediate events to the browser, and syncs the resulting tasks into Todoist through the official Model Context Protocol (MCP) Streamable HTTP endpoint. The project is distributed under the Apache 2.0 License (see LICENSE).

Architecture Overview

• Frontend: React 19 App Router page (src/app/page.tsx) that captures a single natural-language prompt, streams NDJSON events from /plan, and optionally records a short voice note that flows through /api/transcribe.
• Planning pipeline: /plan (documented in openapi/plan.yaml) validates the request, runs intent classification + scenario-specific planning with @cf/openai/gpt-oss-120b / @cf/openai/gpt-oss-20b, discovers Todoist projects/labels dynamically, and streams every stage as newline-delimited JSON. Details live in docs/PLAN_PIPELINE.md.
• Todoist MCP integration: Uses StreamableHTTPClientTransport to connect to https://ai.todoist.net/mcp, lists whatever metadata/tools are available (legacy or official), and calls add-task(s) with both projectId and project_id fields plus normalized priorities (p1…p4).
• Voice transcription: /api/transcribe proxies base64 WebM/Opus audio to @cf/openai/whisper-large-v3-turbo, enforcing an 8 MB limit so the front-end can overwrite the prompt and immediately submit /plan.

Features

• Intent-aware planning: single_reminder, multi_step_plan, recipe_plan, and general_plan templates enforce task counts, dependencies, and tone.
• Metadata-driven prompts: Todoist projects/labels fetched via MCP are injected into the AI prompt and exposed through debug.metadata events to minimize “Inbox” fallbacks.
• Priority normalization: Natural-language cues such as P0, P1, or “high priority” map to Todoist REST priority numbers (4 = P1, 1 = P4). Bulk tools automatically convert to p1…p4 strings.
• Streaming UX: /plan replies with application/x-ndjson events (status, ai.plan, todoist.task, final, error) so the UI can display progress and failures instantly.
• Voice-first flow: Browser MediaRecorder → /api/transcribe → prompt submission in one click, with graceful fallbacks when permissions or size limits fail.

Prerequisites

• Node.js 20+
• pnpm 9+
• Wrangler CLI 4.45+
• Cloudflare account with Workers AI enabled
• Todoist account + API token approved for the MCP beta (https://ai.todoist.net/mcp)

Configuration

Name	Description
FRONTEND_ORIGIN	Single allowed origin for CORS and Basic Auth prompts.
TODOIST_MCP_URL	MCP Streamable HTTP endpoint (default https://ai.todoist.net/mcp).
TODOIST_TOKEN	Bearer token recognized by Todoist MCP.
BASIC_AUTH_USER / BASIC_AUTH_PASS	Credentials for HTTP Basic auth across /plan and /transcribe.
AI binding	Configured in wrangler.jsonc to access Workers AI (e.g., binding: "AI").

Populate .dev.vars for local runs, then set the same names in Cloudflare via wrangler secret put.

Local Development

pnpm install
pnpm dev          # Next.js dev server (no Worker bindings)
pnpm lint         # ESLint flat config for Next.js 16 + TS strict
pnpm preview      # Build via OpenNext + Wrangler preview
wrangler dev      # Run the Worker locally at http://127.0.0.1:8787

Example streaming request:

curl -u "$BASIC_AUTH_USER:$BASIC_AUTH_PASS" \
     -N -H "Content-Type: application/json" \
     -H "Origin: $FRONTEND_ORIGIN" \
     -d '{"prompt":"Plan a mindful evening"}' \
     http://127.0.0.1:8787/plan

Deployment

pnpm run deploy   # Builds with opennextjs-cloudflare and runs `wrangler deploy`

The script produces .open-next/worker.js, uploads static assets, and publishes to wrangler.jsonc's Worker name.

Observability & Debugging

1. Tail logs with sanitization disabled to capture Todoist payloads:

   LOG_FILE=$(mktemp -t wrangler-tail).log
   WRANGLER_LOG_SANITIZE=false npx wrangler tail cf-todoist-daily-agent --format json > "$LOG_FILE" 2>&1 &

2. Trigger /plan while the tail is running. Look for:
   • [todoist.tools] – discovered MCP tools.
   • [todoist.debug.metadata] – first few projects/labels (confirms metadata scope).
   • [todoist.debug.call-args] – exact add-task(s) payload (check project_id, priority, due*).
3. Cross-reference the Todoist Activity Log to confirm tasks landed in the expected project and priority.

API Summary

• OPTIONS /plan – CORS preflight (204).
• POST /plan – Main planner endpoint returning application/x-ndjson (see openapi/plan.yaml).
• POST /transcribe – Voice helper, returns { text, language } or an error payload.

OpenAPI schemas define request/response bodies for automation and client generation.

Testing & Verification

• pnpm lint
• pnpm preview
• wrangler dev --local followed by a sample curl (as above) to validate NDJSON streaming and Todoist MCP connectivity.

License

Licensed under the Apache License, Version 2.0.