paraphe
Allows sending decision cards to the owner's phone via Telegram and receiving their replies, so agents can ask for approval without the owner at the machine.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Paraphe
The owner-decision inbox for agents. An agent raises one decision, you answer it on your phone or on the machine running it, and that same agent picks the answer up and carries on.
It is for people who run agents that should ask before they act. Paraphe is the small, self-hosted thing that holds the question until the human answers it — not a task tracker, not a chat client, and not a workflow engine.

Install and run
Python 3.11 or newer, and nothing else.
pip install parapheIn a checkout, the same install inside a virtual environment, with a configuration file to start from:
python3 -m venv .venv
.venv/bin/pip install .
cp config.example.toml paraphe.toml # then fill in two values you make up
.venv/bin/paraphe --config paraphe.tomlThose two values are yours:
Value | Who holds it |
| the agent. It creates cards and reads answers. |
| you. It answers them, and it is never given to an agent. |
With pip install paraphe and no checkout, create paraphe.toml with those
two lines; Paraphe reads it from the directory it runs in.
Paraphe ready: mcp http://127.0.0.1:8787/mcp answer-path onTo send decisions to your phone:
Open Telegram's BotFather, run
/newbot, and keep the token it gives you private.Open the new bot and send it
/start. Find your own numeric Telegram user id using a user-info bot you trust.Set
bot_tokenandowner_telegram_idinparaphe.toml, then check the pair before starting the server:.venv/bin/paraphe check telegram .venv/bin/paraphe --config paraphe.toml
The check names the bot and sends a plain setup message to your phone. It does not print the token or create a decision card. If delivery is refused, verify the numeric id and make sure you opened a private chat with the bot first. Without the two phone settings, cards are printed where you are looking, which is what the local run above does.
A repeat ask with the same external_id returns the existing card and does
not notify again — after a restart, read the answer with get_response by
external_id.
Or run the container, with the data location mounted:
mkdir -p paraphe-data && chmod 0777 paraphe-data
docker build -t paraphe .
docker run --rm -v "$PWD/paraphe-data:/data" \
-e PARAPHE_MCP_CREATE_BEARER=... -e PARAPHE_OWNER_ANSWER_TOKEN=... parapheIt binds loopback inside the container, because the answer path refuses a
non-loopback bind. To reach it from outside, publish 8787
(-p 127.0.0.1:8787:8787), configure the phone destination (a bot token and
your Telegram id) and set PARAPHE_MCP_HOST=0.0.0.0; the answer path is then
off, and taps arrive on the phone.
Related MCP server: Agent Escalation Harness
Connect your agent
Paraphe serves Streamable HTTP on the address the ready line prints
(http://127.0.0.1:8787/mcp by default), authenticated with the create bearer:
{
"mcpServers": {
"paraphe": {
"type": "http",
"url": "http://127.0.0.1:8787/mcp",
"headers": { "Authorization": "Bearer <mcp_create_bearer>" }
}
}
}The block is Claude Code's .mcp.json shape; any Streamable HTTP client takes
the same URL and header in its own MCP config. The create bearer creates and
reads. It cannot answer. Keep the file out of version control: that bearer
reads every card.
The loop
ask → you see the card → you answer → the asking agent resumesThe answer returns through the ask itself: the call may wait (wait_seconds),
paraphe wait can hold the card's lifetime, and any run that missed both
drains the answer from the store at its next boundary. Nothing has to be said
in chat.
On the phone the card is ordered rich text, not a wall of prose: it names who is asking and from where (agent, runtime, repository, worktree, ticket), bolds its title, numbers the choices with their one-line notes and marks what is recommended, and lists the limits, links and expiry. A long-press reply answers it in your own words — or asks a question — and that text returns to the exact session that asked.

The approval card as it arrives on the phone.
/pending keeps one pinned message in the owner's private chat: every open
card, oldest first, eight to a page, each row opening that card's own
rendering and answer buttons in the same message. It refreshes within five
seconds of a card changing, and it never sends or unpins while refreshing.

The pinned Pending list: eight open cards, oldest first, each with the agent, the repository and the expiry.
Both images in this section are rendered from the adapter's own output over a demo store, in the destination's own palette. They are not device captures, and the cards in them are fictional.
docs/demo/console-loop.md is a recorded run of
exactly that, on a clean checkout, with no third-party credential and no
external service: the agent asks, the card is printed, the agent's own
credential is refused when it tries to answer, the owner answers, and the
agent picks the answer up.
The line this product draws
The credential an agent holds cannot answer a decision. It creates and it reads. Answering needs the owner's credential, on a path the agent's cannot reach. An inbox an agent can approve on its behalf is not an owner-decision inbox, so this is not configurable.
Documentation
Document | What it covers |
the MCP tool surface and the card lifecycle | |
adding a destination, in two methods | |
optional Hermes approval transport, installation and bounded diagnostics | |
the async return protocol, per runtime | |
writing a card the owner can answer at a glance | |
what the words mean (card, tap, return path, owner) | |
the decisions behind the shape | |
the v1 specification | |
what is next, and what is not planned | |
the generated wiki index | |
what Paraphe is, and every route to a running inbox | |
how the inbox runs, page by page |
Dependencies
The runtime imports nothing outside the Python standard library, so using Paraphe pulls no dependency tree. Installing it fetches the build backend once, at install time, and nothing after that.
Licence
AGPL-3.0-or-later (LICENSE). Modify it and serve it to other
people over a network and you must offer them your modified source; nothing
here obliges a company to publish the application it builds on top.
Where your data lives
Cards are a SQLite file in your per-user data directory
($XDG_DATA_HOME/paraphe, else ~/.local/share/paraphe) unless you set
store_path. The directory is created 0700 and the file 0600, and Paraphe
never reads or writes outside that location.
Upgrading from a release that used /var/lib/paraphe: stop the running
Paraphe server, then copy its existing store into the current per-user location:
paraphe store relocateThe command refuses to overwrite a store at the target, verifies the SQLite
copy before installing it, and leaves the old store in place as a backup. To
remove the old store instead of keeping it as a backup, run the same command
with --move:
paraphe store relocate --moveParaphe refuses normal startup with the new default while a store exists only
at /var/lib/paraphe/inbox.sqlite; this prevents an upgrade from quietly
beginning a second, empty inbox. Setting store_path to the old location
remains available when relocation is not wanted.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Let AI agents ask a real person for a decision on their phone and wait for the answer.
Let your AI agents ask you for decisions and get your answers back, from your phone or the web.
Give any AI agent a way to ask a person — for approval, a decision, an answer or a review.
Human-input bridge for AI agents with voice-first answer links, MCP tools, and HTTP APIs.
Related MCP Servers
- AlicenseAqualityBmaintenanceAn MCP server that enables AI agents to pause and request human approval or information via Slack, Telegram, or macOS dialogs before proceeding with actions.29 PyPI15Apache 2.0
- FlicenseNot gradedqualityDmaintenanceAn MCP server that lets an AI coding agent pause on human-only tasks, request structured input via a form, and resume with the answer, all locally without cloud dependencies.-
- AlicenseNot gradedqualityAmaintenanceMCP server that provides an ask_human tool so AI agents can request decisions from users via chat apps, with deadlines, defaults, and an append-only audit log.93 npmApache 2.0
- AlicenseNot gradedqualityNot gradedmaintenanceEnables any MCP-compatible AI agent to ask humans questions in real time via a mobile-friendly PWA with push notifications, operating local-first without third-party services.MIT