Automation agent through Claude MCP

Hello, Oleksiy!

I carefully studied your document for Pocket Rocket. The breakdown into phases (from Read-only monitoring to Write actions) is the only correct approach for working with the Meta API to avoid "draining" the budget with a faulty script.

Most developers here suggest simply writing the integration code (Node.js/TS + MCP SDK). But the main problem of AI agents in marketing is not in the code, but in the business logic. My background is in marketing analytics of sales funnels and identifying lead loss points. I understand how a media buyer thinks and will be able to correctly "encode" this logic in Claude.

Here’s how I propose to architecturally solve your task:

Niche specifics (White-hat or High-risk): This is something purely technical specialists often forget. It is important to understand which niches Pocket Rocket works with. If it’s a classic business (low risk) — we can safely use the official and built-in Meta Marketing API. However, if you are working with high-risk niches, standard API requests may provoke bans, and we will need to build a more complex architecture.
…
Not just a search engine, but a Senior analyst: Claude makes mistakes if given "raw" numbers. I will build MCP tools (for example, get_insights) so that they return data already with context: comparison of current CPL with baseline over 7 days, checking frequency of impressions (Frequency) and audience burnout markers.

Team player (Human-in-the-Loop): As modern practice of AI agents shows, a bot should not operate in isolation. We will implement a Telegraf bot, where the agent will drop anomalies with inline buttons: [Stop ad set], [Increase budget by 20%], [Ask for advice in chat].

Feedback Loop (Agent training): When your team presses a button in Telegram, the database (SQLite/Postgres) records this decision. Thus, the agent will learn from the actions of your real media buyers, adapting to the Pocket Rocket strategy.

Technical base: Node.js or TypeScript, official MCP SDK from Anthropic, handling Rate Limits from Meta (caching requests), isolation of access tokens.

The agent should act like your best employee: see the drop in the funnel, suggest solutions, and wait for your approval.

I have two questions to start:

In which niches do your advertising accounts primarily operate (is it worth laying the architecture for high-risk)?

Do you already have a compiled historical database (baseline metrics) for clients, with which the agent should compare current results in Phase 1?

Show original

Translate

Hello!
I read the technical specifications completely — the architecture and stack are clear. I am actively building StratBase.ai (my SaaS) with Claude Code + MCP in production; this is my daily tool. Proposal for §7 of your technical specifications: I am ready to do a test (get_insights + CPL in 7 days) before starting the main project. This also mitigates the main risk of Phase 1 — System User Token + correct permissions (ads_read, ads_management, business_management, pages_read_engagement). You will see the quality of the code on a real account in 2-3 hours. Regarding Phase 1: 7 get_* tools, node-cron 08:00, detection of 6 anomalies (CPL +30% vs 7d baseline, EMQ 5, drop without conversions), Telegram-digest via telegraf, SQLite. I will handle the rate limit of 200 requests/hour through caching repeated requests to Claude. I am new to Freelancehunt. Phase 1 — 8,000 UAH / 7 days. Phase 2 (write-actions + inline buttons) — separately after agreement.

Show original

Translate

Hello.

I have reviewed the technical specifications.
For the MVP, the logic is clear: first, a read-only contour that collects metrics, generates a daily report in Telegram, detects anomalies based on CPL / ROAS / CTR / Frequency, and provides an audit for the team. Write-actions, confirmation buttons, and autonomous actions will follow in the next phase.

From a technical perspective, it is important to correctly separate:
- MCP server as the tool layer
- Meta Marketing API as the source of data and actions
- Telegram bot as the interface for reports and confirmations
… - SQLite for configs, logs, and state in PoC (but better to use PostgreSQL right away)
- scheduler for daily and weekly tasks
- separate control of permissions, tokens, rate limits, and logging

This approach allows for a safe start with the MVP, where there is no risk to client accounts, and only after that move on to semi-automatic actions based on rules.

For the MVP / Phase 1, my estimate is 20,000 UAH.

This project is personally interesting to me, and I am open to dialogue in private messages.

Show original

Translate

Hello! I can implement it. Write to me privately to discuss all the details. I will be glad to cooperate!

Show original

Translate

Hello, Alexey! I am Nina, the manager of architect engineer Valentin.

We have thoroughly studied your technical specifications for Pocket Rocket. Your phased implementation strategy (from monitoring to full automation) is the only right way to work with the Meta Ads API. Valentin specializes in AI-augmented workflow and uses Claude MCP daily to develop complex systems, so we understand the protocol architecture from the inside.

How we will implement the project:

MCP Server (TypeScript): We will develop a custom tools server that will allow Claude to directly "communicate" with your advertising account.

Security and control: We will implement a Human-in-the-loop mechanism (via Telegraf). The agent will suggest actions (stop campaigns, rescale budget), but the final decision is yours through a button in Telegram.
…
Smart alerts: We will set up a system of triggers (ROAS, CPA, Spend) that the agent will analyze in context, not just by dry numbers.

Infrastructure: SQLite for deterministic logs and long-lived Meta tokens for stable operation 24/7.

Why Valentin: Due to his deep immersion in the Anthropic/Claude ecosystem, he will assemble Phase 1 (monitoring) in the shortest possible time, ensuring code cleanliness and scalability for Phases 2 and 3.

Timeline (Phase 1): 7–10 days.
Budget: 20,000 UAH (as a base for the first stage with full infrastructure setup).

Alexey, are you ready to provide access to the test application in Meta for Developers to start the integration?

Best regards, Nina.

Show original

Translate

Hello!

I read the technical specification — everything is clear. I have such a stack (TypeScript, MCP SDK, Meta API, telegraf).

The main thing that distinguishes my approach: I start with Meta API authorization and a test get_insights on a real account — this mitigates the main risk of the project in the first few days.

Phase 1 (read + Telegram reports) — 2 weeks.
Phase 2 (write + confirmation buttons) — another 2 weeks.

… Write to me — we will discuss the details and cost.

Show original

Translate

Good day, Oleksiy!

Regarding the described task — this is a job on the MCP agent over the Meta Marketing API: tools for requests to the advertising cabinet, wrappers around reports and actions (pausing campaigns/budget adjustments), orchestration through Claude.

I am a Senior Full Stack developer with 14 years of experience, specializing in Node.js / TypeScript / NestJS. I actively work with the MCP protocol and tool-use Claude — I build servers with validated inputs and safe actions (read-before-write, dry-run for critical operations). I understand the Meta Marketing API and its pitfalls (rate limits, ad-account permissions, batch endpoints).

Approach:
- MCP server in TypeScript (stdio or HTTP transport of your choice),
- a set of tools for key scenarios from the technical specifications (I will open a Google Doc after the call starts),
… - secure storage for Meta access tokens (refresh + scope minimization),
- audit log of all write actions,
- end-to-end tests for key tool flows.

Approximately 12-14 days, a budget of 20,000 UAH is adequate for a Proof of Concept. If it then expands in functionality — we will switch to hourly at $30/h.

I am ready for a 20-minute call to clarify the scope.

github.com/fun4life1990 · linkedin.com/in/marchenkovsky

Show original

Translate

Good day. I am ready to complete this project as I have extensive experience in application development.

Show original

Translate

Good day
My name is Angelina, I represent the company King Kong Lab
I have reviewed the technical specifications — the task is clear, we can implement an AI agent for working with Meta Ads based on Claude + MCP Server + Meta API
We have experience in
Node.js / TypeScript
working with APIs (including advertising ones)
creating Telegram bots
automation and analytics
Regarding the approach
… we will implement the MCP Server with the necessary tools (get_* and write)
we will set up a stable integration with the Meta Marketing API
we will create a Telegram bot with commands (/report, /audit, /status, etc.)
we will implement a scheduler for daily reports
we will add anomaly detection logic (CPL, ROAS, CTR, budget)
we will build a database (SQLite) for configs and logs
Next, we can move on to
semi-automated actions (with confirmation via Telegram)
and full automation according to the rules
We will create a clean, scalable architecture so that the agent can be expanded for new tasks
Regarding timelines
MVP (Phase 1) — 2–3 weeks
I am ready to discuss the details and start in the near future
I look forward to collaborating

Show original

Translate

I thought for 12 seconds

Hello! We are Devoxen, a company with extensive experience in developing automation, AI agents, and integrations for working with advertising accounts, APIs, and business processes.

We have reviewed your technical specifications regarding the agent for working with the Meta advertising account. We have practical experience in implementing similar solutions: task automation, data processing, logic for interacting with external services, and building stable agent operation scenarios.

I suggest we move to private messages for a more detailed discussion of the technical specifications, terms of cooperation, and cost.

Show original

Translate

Hello, Alexey!

I read the full project brief in Google Doc — this is exactly the stack I work with every day: Claude + MCP SDK + external APIs. PocketRocket Agency, 14 Meta accounts, reducing monitoring time from 2 hours to 15 minutes through the agent. I will outline how I see the implementation of Phase 1 (Read-only monitoring).

MCP server and tools:
— get_ad_accounts / get_campaigns / get_adsets / get_insights / get_pixel_events / get_audiences / get_account_summary — implementation through Meta Marketing API, caching repeated requests with Claude
— Node.js 20 + TypeScript + mcp-server-sdk from Anthropic — as in the project brief
— Node-cron scheduler for daily launch at 08:00, Telegraf for the agency's Telegram bot
— SQLite for thresholds, client configs, action log. Deploy Hetzner CX21.
…
Anomaly detector — rules directly from the project brief:
— CPL +30% to the 7-day average → flag
— Spending >120% of daily limit without conversions → flag
— CTR -40% → flag
— Frequency >5 on the ad set over 7 days → flag
— Active campaign does not spend >4 hours → flag
— EMQ

Show original

Translate

Good day. I have studied the technical assignment for PocketRocket. My main stack of Node.js and TypeScript fully covers your tasks. The logic of working with the Meta Marketing API, long-lived tokens, and request limits is completely clear to me. There are no questions regarding the architecture of the MCP server integration. Instead of the specified SQLite, I can suggest using Postgres in conjunction with Docker for greater reliability in storing logs during offline actions, but the final decision is yours. I am ready to complete a test task for the implementation of the get_insights function. We can discuss the budget and deadlines for the phases in private messages.

Show original

Translate

Hello!

Very high-quality project brief — it's clear that you have thought through the architecture and understand what you want to achieve. The project is right in my profile.

I work with Claude and MCP tools daily — this is my main work stack. I build AI agents with a complete pipeline: from integration with external APIs to Telegram bots and automation. There are market-released products developed in this format.

I am familiar with the stack: Node.js/TypeScript, REST API integrations, Telegram bots (grammy/telegraf), SQLite, cron jobs, deployment on VPS. I understand the logic of working with advertising APIs and the specifics of rate limits.

The phased approach with increasing responsibility (read → semi-auto → auto) is the right architectural solution, I fully support it.
…
I am ready to discuss the details and get started.

Show original

Translate

Hello, Alexey! I am ready to develop this agent for you according to your specifications, I will do it quickly and efficiently, feel free to contact me.

Show original

Translate

Hello. I have been working with Node.js for over 8+ years. I am ready for collaboration. Please contact me.

Show original

Translate

Hello! We are the SDEV team, specializing in the development of complex automation systems and integrations. We have experience in building solutions for working with advertising accounts, including task automation through APIs and external tools.

Your project is to create an automation agent for Meta Ads using Claude MCP. We understand that it is necessary to implement the interaction logic between the MCP agent and the advertising account, including request processing, executing actions (starting/stopping campaigns, changing budgets, analyzing metrics), and reliable data transmission. An important aspect is stability, logging, and scalability.

We have already implemented similar projects — in particular, an automated system for monitoring and managing advertising in other ecosystems, where accuracy and minimizing manual intervention were crucial. Our approach involves clear architecture, testing at every stage, and flexibility to changes in the technical specifications.

We are ready to take on the project after a detailed analysis of the document — we will clarify the agent's operation logic, use cases, and expected functionality.

Show original

Translate

Hello! I have reviewed the technical task regarding the development of the Meta MCP Agent for PocketRocket. The project looks technologically deep, and this is exactly the type of tasks our team specializes in.

Considering the specifics of working with autonomous agents (through the Model Context Protocol) and potentially high-load mechanics of the PocketRocket ecosystem, we do not use template solutions. Our approach is to create a fault-tolerant microservice infrastructure that will ensure the uninterrupted operation of the product.

What we offer from an engineering perspective:

1. Deterministic logic through State Machine: To ensure that the agent does not lose the context of interaction and flawlessly executes multi-step algorithms, we will implement a finite state machine architecture based on Node.js or Python. This guarantees that no action will be lost even during asynchronous requests or emergency restarts.
2. Hybrid database architecture: We propose to handle all unstructured context of AI sessions and logging through a high-performance document-oriented database (MongoDB), and for convenient monitoring of the agent's operational status by your team, we can set up transparent two-way synchronization with visual tables (for example, Airtable).
3. Infrastructure stability: We will set up deep control of I/O operations, embed mechanisms to protect against polling stalls, and isolate task schedulers. Your application will be insured against identifier shifts in memory or loss of connection with the server.
…
We design such systems in a way that AI is solely responsible for semantic routing and flexibility, while critical transactions and logic remain under strict deterministic control of our code. This minimizes the risks of hallucinations and system failures.

You can evaluate our approach to developing complex ecosystems, autonomous agents, and integrations in Telegram by reviewing our completed projects on the agency's website: https://vaysed.me/

Show original

Translate