# Aval > Aval is a hosted MCP server that lets an AI agent ask its human a question and wait for the answer before acting. The question reaches the human's phone in Telegram or by email, with one button per option. Made by Taraxe (taraxe.com) in France. Free and unlimited in Telegram; email and other channels come with a taraxe subscription. Use Aval as a human-in-the-loop approval step: before paying, sending an email or a message, deleting, publishing, signing up, or choosing between options the human cares about. The agent never sees the answer link, so it cannot answer for the human. An expired question means no approval. ## Connect - MCP endpoint: https://www.taraxe.com/aval/mcp (Streamable HTTP) - Auth: OAuth 2.1 with dynamic client registration and PKCE. Paste the endpoint in the MCP client; the human signs in with their email (6-digit code). The account is created on first connection, for free. - Works with any MCP client: Claude (web, desktop, Claude Code), ChatGPT, Le Chat, Grok, Cursor, VS Code, and others. - REST API for n8n or scripts: POST https://www.taraxe.com/api/aval/questions with `Authorization: Bearer ` (key created at https://www.taraxe.com/aval/compte/). Body: `{"question": "...", "options": ["...", "..."]}`; add `?wait=45` to wait for the answer. GET https://www.taraxe.com/api/aval/questions/ reads the status right away (add `?wait=45` to wait). ## Tools - `ask_human(question, options?, context?, agent_name?, expires_in_minutes?, wait_seconds?)`: asks the question (max 300 characters) with 2 to 6 option buttons (default Yes / No) and optional context (max 2,000 characters: amounts, recipients, drafts, links). One decision per question, written in the human's language and understandable without the conversation. Waits up to 45 seconds, then returns status `answered` (`answer`, `answer_index`, optional `comment`) or `pending` with a `question_id`. Expires after 60 minutes by default, up to 7 days. Free and unlimited in Telegram; by email only with a subscription (error `subscription_required` otherwise). At most 6 questions a minute (error `rate_limited`). - `wait_for_answer(question_id, wait_seconds?)`: waits up to 45 seconds and returns `pending`, `answered` (chosen option and optional comment), `expired` (treat as no) or `cancelled`. Call it again while `pending`. Free. - `cancel_question(question_id)`: withdraws a pending question that no longer matters (answered another way, task changed). Free. In a scheduled or looping agent, don't block on the answer: store the `question_id` with the task, end the run, and check it once on the next run. ## Pricing - Telegram: free and unlimited for everyone, even without a subscription. No credit card. - Email, WhatsApp and Signal: with a taraxe subscription, from Starter (free trial); WhatsApp and Signal as soon as the channel opens. Starter allows 100 questions a month on WhatsApp (beyond that, they go through the other channels), Pro has no limit. See https://www.taraxe.com/#tarifs. - Without a subscription, a question that could only go out by email is refused with the error `subscription_required` (HTTP 402): the human can connect Telegram, or start a Starter trial. Accounts with credits left from before (no longer sold) use 2 per question by email until they run out. - At most 6 questions a minute per account. Waiting, reading and cancelling are free. ## Channels - Available: Telegram (answer buttons in the chat, connected with one tap from the account page) and email (one button per answer, also the fallback when no other channel is on). - Coming later, with a subscription: WhatsApp, Signal. ## Guides - [Human-in-the-loop: staying in control of an AI agent](https://www.taraxe.com/aval/en/guides/human-in-the-loop-ai-agents/): What “human-in-the-loop” means, which actions a human should approve, and how to set it up without slowing the agent down. - [Adding human approval to an MCP agent](https://www.taraxe.com/aval/en/guides/mcp-human-in-the-loop/): Give Claude, ChatGPT or Cursor an ask_human tool: the agent asks for your approval on your phone before acting, and waits. - [What an AI agent shouldn't do on its own](https://www.taraxe.com/aval/en/guides/ai-agent-actions-requiring-human-approval/): Paying, sending, deleting, publishing, committing, choosing for you: the actions a human should approve, plus an instruction to copy. - [n8n: approve a step from your phone](https://www.taraxe.com/aval/en/guides/n8n-human-in-the-loop/): Human approval in an n8n workflow with an HTTP Request node: the question reaches your phone with buttons, and the workflow waits. - [Claude Code: answer its questions from your phone](https://www.taraxe.com/aval/en/guides/claude-code-phone-notifications/): Get notified when Claude Code is waiting, or answer it remotely with buttons, through an ask_human MCP tool. - [Running an AI agent in a loop over a task graph](https://www.taraxe.com/aval/en/guides/autonomous-ai-agent-loop-task-graph/): A state file with dependencies, a routine that wakes the agent, one PR per task, and your decisions asked on your phone with Aval. - [Agents, harnesses, loops, graphs: the four layers of an autonomous AI agent](https://www.taraxe.com/aval/en/guides/ai-agents-harness-loops-graphs/): What each layer adds, from an agent calling tools to the graph that tells it what to do next, with real examples and where the human fits in. - [Claude and ChatGPT notifications vs Aval: what's the difference?](https://www.taraxe.com/aval/en/guides/claude-chatgpt-notifications-vs-aval/): A notification tells you to go back to the app. Aval asks the question with buttons and brings your answer back to the agent. When to use which. ## Guides en français - [Human-in-the-loop : garder la main sur un agent IA](https://www.taraxe.com/aval/guides/human-in-the-loop-agent-ia/): Ce que veut dire « human-in-the-loop », quelles actions faire valider par un humain, et comment le mettre en place sans ralentir l'agent. - [Ajouter une validation humaine à un agent MCP](https://www.taraxe.com/aval/guides/validation-humaine-agent-mcp/): Donner à Claude, ChatGPT ou Cursor un outil ask_human : l'agent vous demande votre accord sur votre téléphone avant d'agir, et attend. - [Ce qu'un agent IA ne devrait pas faire seul](https://www.taraxe.com/aval/guides/actions-a-faire-valider-agent-ia/): Payer, envoyer, supprimer, publier, s'engager, choisir à votre place : les actions à faire valider par un humain, et une consigne à copier. - [n8n : faire valider une étape sur son téléphone](https://www.taraxe.com/aval/guides/n8n-validation-humaine/): Une validation humaine dans un workflow n8n avec un nœud HTTP Request : la question arrive sur le téléphone avec des boutons, le workflow attend. - [Claude Code : répondre à ses questions depuis son téléphone](https://www.taraxe.com/aval/guides/claude-code-telephone/): Être prévenu quand Claude Code attend, ou lui répondre à distance avec des boutons grâce à un outil MCP ask_human. - [Faire tourner un agent en boucle sur un graphe de tâches](https://www.taraxe.com/aval/guides/agent-autonome-boucle-graphe-de-taches/): Un fichier d'état avec des dépendances, une routine qui réveille l'agent, une PR par tâche, et vos décisions posées sur votre téléphone avec Aval. - [Agents, harness, boucles, graphes : les quatre étages d'un agent autonome](https://www.taraxe.com/aval/guides/agents-harness-boucles-graphes/): Ce qu'ajoute chaque étage, de l'agent qui appelle des outils au graphe qui lui dit quoi faire, avec des exemples concrets et la place de l'humain. - [Notifications Claude et ChatGPT, ou Aval : quelle différence ?](https://www.taraxe.com/aval/guides/notifications-claude-chatgpt-ou-aval/): Une notification vous fait revenir dans l'application. Aval pose la question avec des boutons et rapporte votre réponse à l'agent. Quand choisir lequel. ## Links - [Aval](https://www.taraxe.com/aval/): the product page, in French and English, with setup steps for each client and a FAQ. - [Account](https://www.taraxe.com/aval/compte/): plan, notification channels, API key, history. - [Pricing](https://www.taraxe.com/#tarifs): taraxe subscriptions. - [Taraxe llms.txt](https://www.taraxe.com/llms.txt): the other projects by the same maker.