MAX

Recipes: your agent and your conversations — MAX

Follow practical MAX CLI workflows for your AI agent, scripts and everyday messaging tasks.

Give your agent a useful task without writing a command sequence yourself. These examples show requests you can adapt to your chats, the expected result and where to find the exact command details.

Use these recipes for recurring MAX tasks with your AI agent, such as unread summaries, chat reports, commitments, unanswered requests and group checks. Each includes a ready-made request, a command and whether it changes MAX. You will learn to schedule tasks and set restrictions enforced by max.

Terms used below:

  • Agent — your AI agent running commands on this computer.
  • Skill — instructions for using max; max skill install puts them where the agent looks.
  • MCP client — an agent application without terminal access, such as Claude Desktop; it calls max personal mcp serve tools instead of commands.
  • Noninteractive mode — the agent receives one request, executes it and exits: Claude Code -p or codex exec.
  • cron — the Linux/macOS scheduler running commands at specified times.

What you can do

RecipeWrites to MAXSuitable for schedule
Morning Briefnoyes
Weekly work chat reportnoyes
Who owes whomnoyes
To whom you did not answernoyes
Find what was saidnoyes
Draft answeronly after your “yes”no
Group you leadonly if allowed by group rulesyes

One-time setup

  1. Install max and log in: see installation and login and sessions.

  2. Install the skill for your agent:

    max skill install

The skill goes to Claude Code (~/.claude/skills/max-cli/) and Codex/Gemini CLI (~/.agents/skills/max-cli/). max skill show prints it for agents using other locations.

  1. Save recipe requests in files, such as ~/max-recipes/. The schedules below read those files.

The MCP client does not need the skill. Connect the MCP server (max personal mcp config prints an entry for its settings) and call the ready-made commands - /catch-up, /review, /reply, /find. See ready command MCP servers.

Agent permissions

A scheduled agent runs while you are away. Set restrictions in configuration the agent cannot edit: a request such as “send nothing” can be misinterpreted, whereas the tool enforces its permissions.

Prohibition at the agent level. This is a setting for the agent itself. For example, Claude Code in -p mode only runs the command from --allowedTools. If max personal messages send is not in the list, the agent cannot call it:

claude -p "$(cat ~/max-recipes/morning.md)" --allowedTools "Bash(max personal inbox:*)"

Codex does not have such a list. Its sandbox by default closes the network and file writing, and max needs both: connect to MAX and save what it reads to its copy. Therefore, Codex starts with --sandbox danger-full-access, and sending is limited by the max settings below.

Restrictions in max apply to every agent:

for resource in messages reactions polls topics chats contacts account bot conversations; do
  max personal config set "permissions.$resource" readonly
done
max config show                       # проверить эффективные права
max personal config set sends_per_hour 5          # или: не больше пяти сообщений в час
max personal recipients add "Иван Петров"       # и писать только в эти чаты

readonly restricts only its named resource: restricting messages does not block reactions, voting or chat changes. The loop sets every resource. A more specific key such as permissions.messages.send: allow takes precedence; remove such allowances for a read-only profile.

permissions also affects your own commands until removed with max personal config unset permissions.<ресурс>, restoring inherited settings and defaults. Use a separate profile to restrict only the agent. max personal sends list records every send attempt, including rejected ones. See send protections and permissions.

Reading does not reveal your activity. None of the commands below marks messages read, so other people do not see that the agent opened the chat.

Scheduling

  • Claude Desktop - scheduled tasks. The task runs on your computer, so max and your login are available to it. The skipped run (the computer was sleeping) is executed once after waking up. See Claude scheduled tasks.

  • max schedule. A built-in schedule: it makes the entry in the system scheduler itself (systemd on Linux, launchd on macOS, Task Scheduler on Windows) and needs no crontab editing. The profile is named as the first word, as with any command:

    max personal schedule command inbox-daily --at 08:30 -- inbox      # команда `max personal inbox` каждый день
    max personal schedule agent morning claude "прочитай max skill show follow-ups и выполни ежедневную проверку" --at 08:30
    max personal schedule list                                         # что назначено и как прошёл последний запуск
    max personal schedule run morning                                  # запустить сейчас, как это сделал бы таймер
    max personal schedule remove morning

    schedule agent runs claude or codex with the MAX_READ_ONLY=1 variable: the agent reads and writes drafts but cannot send anything. The agent's answer is saved, and you get a notification. --max-cost stops a claude run that costs more than the given amount in dollars. The schedule is set with --every 30m, --at 08:30 (every day) or --at-time 2h (a single run).

  • cron and claude -p. Without an application, on any computer that is turned on at this time:

    30 8 * * * claude -p "$(cat ~/max-recipes/morning.md)" --allowedTools "Bash(max personal inbox:*)" >> ~/max-recipes/morning.log 2>&1

Cron jobs run in an almost empty environment. On Linux this breaks max in two ways:

  • node: not found, code 127. The Node supplied via nvm, fnm or Volta is not in the system PATH, and cron knows only the system one.
    • no token found for profile "personal", although it has logged in on this machine, code 4. max cannot reach the system password storage and does not see the token. Don't login again: login is fine, it's a matter of the environment.

Both lines go to the beginning of crontab -e, with their own values. The folder is from dirname "$(which node)", the number is from id -u:

PATH=/home/ivan/.nvm/versions/node/v24.19.0/bin:/home/ivan/.local/bin:/usr/local/bin:/usr/bin:/bin
XDG_RUNTIME_DIR=/run/user/1000

The password store is available while you are logged in. Check the first run with your hands and look in the log. See Claude's noninteractive mode.

  • Cron with codex exec. The equivalent for Codex:

    30 8 * * * codex exec --sandbox danger-full-access "$(cat ~/max-recipes/morning.md)" >> ~/max-recipes/morning.log 2>&1

See Codex noninteractive mode.

  • Inside an open session Claude Code - /loop or “remind me at 15:00”. Works while the session is open. See Claude Code Scheduled Tasks.

Claude's cloud tasks (routines) are not suitable: they are executed on another computer where max and your login do not exist.

Daily follow-up check

The ready-made follow-ups skill keeps track of your commitments: once a day it reads max tasks agenda, looks only at the chats the agenda names, sets and links tasks, and returns a short list with draft replies. It sends nothing. To read it: max skill show follow-ups. To run it on a schedule:

max personal schedule agent follow-ups claude "выполни ежедневную проверку: max skill show follow-ups" --at 08:30

Morning summary

Writes to MAX: no. Permit: Bash(max personal inbox:*).

Run max personal inbox --new --json. Group messages by chat. Give each chat one line: who is writing and what they need. Put items needing a response today first. Combine advertisements and service notifications into one final line.

--new shows each message once. max remembers where it left off in each chat, and the next run starts from that place. The very first run looks at the last 24 hours.

Where “new” starts

  • max personal inbox — unread messages, as MAX counts them: everything you have not opened on any device.
  • max personal inbox --new — since the previous --new run. Only max knows this point; your contacts do not see it.
  • max personal inbox 'date>2d' — the past two days; the saved point does not move.

Neither command marks messages as read. To do that, run max personal chats mark-read <чат>. The other person will see the read receipt.

Direct chats, groups and channels separately

kind: leaves only chats of the required type: dialog - personal, group - groups, channel - channels. The channel summary and conversation summary can be run separately, at different times. Each chat has its own point, and one run does not hide what the other has not yet shown:

30 8 * * * claude -p "$(cat ~/max-recipes/morning.md)" --allowedTools "Bash(max personal inbox:*)"
0 19 * * * claude -p "$(cat ~/max-recipes/news.md)" --allowedTools "Bash(max personal inbox:*)"

morning.md contains max personal inbox --new 'kind:dialog OR kind:group' --json; news.md contains max personal inbox --new kind:channel --json and a request to pick out the main points. Without kind:, everything comes together.

Weekly work-chat report

Writes to MAX: no. Permit: Bash(max personal messages list:*).

Read the past 7 days in the Project Alpha chat: max personal messages list "Проект Альфа" 'date>7d' --limit 200 --json. If the response has a page.next step, continue reading with its --cursor. Report decisions, responsibilities and deadlines, and unanswered questions. Include the date and author for each item.

Commitments and follow-ups

Writes to MAX: no. Permit: Bash(max personal review:*), Bash(max personal messages context:*), Bash(max personal search messages:*).

Run max personal review --new --transcribe --json (initially the past 3 days, then since each chat's previous --new). Group what I owe, replies I await and matters to clarify. Include chat, date and source message ids; give a deadline only when stated. Before calling something overdue, check later messages and work groups for completion. If "complete": false, explain missing coverage. Finish with unresolved items.

--transcribe first translates voice messages into text; this may take minutes. The next review is the same request plus unclosed items from the past: the agent will check them first. A chat that has not been read in full will save its point and come again. In the MCP client this is command /review.

People you have not answered

The shortest option is max personal review 'date>7d' --unanswered --json: unanswered questions addressed to you in direct chats, or to you and admins in groups. The broader recipe below also catches requests without a question mark.

Writes to MAX: no. Permit: Bash(max personal chats list:*), Bash(max personal messages list:*).

Run max personal chats list kind:dialog --offline --limit 30 --json. For each chat active in the past 7 days, read max personal messages list <id чата> --limit 5 --json. Show chats whose latest message is not mine and contains a question or request: who it is with, the topic and days elapsed.

Find what was said

Writes to MAX: no. Allow: Bash(max personal search messages:*), Bash(max personal messages context:*).

Find invoice messages using max personal search messages счёт --json. For each match, read max personal messages context <locator> --json and explain who said what and when.

By default, the search is carried out both on the copy on this computer and on the MAX server. The local copy contains only what max has already saved. To search your entire chat history locally, first download it. This is a request from your account, so make it yourself: max personal messages sync <чат>. See download history.

Drafting a reply

Writes to MAX: only after your confirmation. This recipe is for an interactive conversation, not a schedule.

Read the last 20 messages with Ivan Petrov and suggest a reply to his latest question. Do not send it; show me the text.

When you agree, the agent sends himself: max personal messages send "Иван Петров" "…". The MCP client connects via max personal mcp serve. Don't let the client pre-approve max_write and he will ask you before every submission. The profile rights still limit the recording. See MCP Server.

A group you manage

Writes to MAX: only as permitted by group rules. Permit: Bash(max personal review:*), Bash(max personal chats events:*), Bash(max personal chats members list:*), Bash(max personal chats moderate:*).

Run max personal review 'chat:"Поход"' --unanswered 4h --json, max personal chats events "Поход" 'time>=7d' --json and max personal chats moderate "Поход" --dry-run --json. Summarise questions awaiting replies, authors and waiting times; joins/additions this week and who added them; and rule violations with proposed actions. Send no replies and remove nobody: list commands for my approval.

With --dry-run command chats moderate only makes a plan. Without it, it acts within the group rules, so remove Bash(max personal chats moderate:*) from the list if the agent should never act. All scenarios and rules are groups that you lead.

There are no ready-made recipes for MAX yet. For Telegram there is; their requests are also suitable for max, if you replace the tools with command:

  • Telegram MCP: complete guide - morning analysis of incoming messages, draft replies, channel summary, search in several chats.
  • pioh/tg - AI agent with a personal Telegram account: summary once every N minutes, “remind me if I haven’t answered my mom in 15 minutes,” monitoring people and chats.
  • Gorgias MCP cookbook - recipes for the support service, but well organized: everyone is told whether he is writing something, and what to correct for himself.