vatic

I wanted a single binary that could run AI prompts on a schedule, respond to messages on Telegram or WhatsApp, and email me the results -- all configured with plain TOML files. No Python glue, no orchestration layer, no YAML indentation nightmares. Just a Rust daemon that reads a config directory and does the thing.

That's vatic. It's a TOML-configured AI agent framework. You define jobs that run prompts through LLM backends (Claude CLI, Ollama), on a cron schedule or triggered by channel messages, with templated prompts and multiple output targets.

What it does

• Sandboxed execution -- Agents run in Podman containers or Guix shell (--container --network) with isolated filesystems. Credentials are mounted read-only, nothing else leaks. I've found this matters a lot when you're letting an LLM run shell commands on your behalf.
• Channels -- Telegram, Matrix, WhatsApp, email (via Himalaya), or plain stdin. The daemon listens on all configured channels and routes incoming messages to matching jobs.
• Cron scheduling -- Jobs run on cron expressions (0 8 * * *), on channel triggers, or both. Mix scheduled and interactive jobs freely.
• Secrets proxy -- API keys and tokens live in secrets.toml, referenced via {% proxy:name %} tags. Secrets stay out of prompts and job configs, which is the whole point.
• Templated prompts -- Date math, dictionary lookups, memory from previous runs, loops over collections, and pipe transforms -- all in a simple {% tag %} syntax. It's not Jinja, but it covers what I actually need.
• Multiple outputs -- Send results to desktop notifications, email, shell commands, or back to the channel. Stack multiple outputs per job.
• Conversation memory -- Sessions track message history for context-aware assistants. History summarization compresses past runs into memories so prompts don't bloat over time.
• Flexible agents -- Claude CLI or Ollama. Swap backends per job without changing anything else.

Install

Method	Command
Cargo	cargo install vatic
Debian/Ubuntu	Download .deb -- sudo dpkg -i vatic_*_amd64.deb
Fedora/RHEL	Download .rpm -- sudo rpm -i vatic-*.x86_64.rpm
Guix	guix install -L <panther> vatic (Panther channel)

Pre-built binaries for Linux (x86_64), macOS (Apple Silicon, Intel) on GitHub Releases.

Quick start

Run /setup in Claude Code to get started -- it'll walk you through configuring your first channel and job. Use /new-channel or /new-job to add more later.

# Build
cargo build --release

# Run a job
vatic run weather

# List configured jobs
vatic list

# Start the daemon (channels + scheduled jobs)
vatic daemon

Configuration

Everything lives under ~/.config/vatic/:

~/.config/vatic/
  dictionary.toml          # variable substitution
  secrets.toml             # API keys for proxy tags
  jobs/*.toml              # job definitions
  channels/*.toml          # channel connections

Data goes to ~/.local/share/vatic/vatic.db (SQLite).

Job example

name = "Today's weather"
alias = "weather"

[agent]
name = claude

[job]
interval = "0 8 * * *"
prompt = "What's the expected weather for {% date %}, in Lisbon? Keep it to a 1-liner."

[environment]
name = local

[output]
name = notification
message = "Good morning {% custom:name %}; {% result %}"

Agent backends

Backend	Config	How it works
claude	name = "claude"	Spawns claude --print CLI
ollama	name = "ollama", host, model	HTTP POST to /api/generate

Channels

Channel	Config	How it works
stdin	type = "stdin"	Terminal I/O, good for getting started
telegram	type = "telegram", token	Long polling via getUpdates, strips @bot mentions
matrix	type = "matrix", homeserver, user, password	Sync loop via matrix-sdk, e2e encryption
whatsapp	type = "whatsapp"	QR pairing, feature-gated (--features whatsapp)
himalaya	type = "himalaya", poll_interval	Polls email via himalaya CLI

Telegram in groups: By default, Telegram bots have privacy mode enabled -- they only see messages that @mention the bot or start with /. Vatic automatically strips the @botname from incoming text so triggers match cleanly. If you want the bot to see all group messages (without requiring @mention), disable privacy mode via @BotFather: send /setprivacy, select your bot, choose Disable.

Environments

Environment	What it does
local	Runs commands directly
guix-shell	Wraps with guix shell -m manifest.scm -- (or named packages)
guix-shell-container	Like guix-shell but isolated (--container --network), shares ~/.claude
podman	Runs in a Podman container. Auto-builds vatic-agent image on first use

The guix-shell and guix-shell-container environments accept a packages list. Without it, they fall back to manifest.scm:

[environment]
name = "guix-shell"
packages = ["curl", "jq"]

[environment]
name = "guix-shell-container"
packages = ["node", "rust"]
pwd = "/home/franz/project"

Podman accepts an optional image (defaults to vatic-agent:latest, built automatically):

[environment]
name = "podman"
image = "node:22-slim"   # optional
pwd = "/home/franz/project"

Template tags

Tag	Description
{% date %}	Today's date (YYYY-MM-DD)
{% date minus=1d %}	Date offset (supports d, h, m)
{% datetime %}	Date and time
{% datetimeiso %}	ISO 8601 datetime
{% custom:name %}	Dictionary lookup
{% result %}	Job result (in output templates)
{% message %}	Incoming channel message
{% sender %}	Message sender identifier
{% memory %}	Last run result
{% memory minus=2 %}	Result from N runs ago
{% proxy:name %}	Secret proxy URL substitution

Loops:

{% for i in memories limit:3 %}
Date: {% i.date %}
Result: {% i.result %}
{% endfor %}

{% for i in (1..3) %}
{% date minus=i"d" %}
{% endfor %}

Pipes: {% i.result | summary %} -- transforms the value through the agent.

Outputs

Output	Description
notification	Desktop notification via notify-send
msmtp	Email via msmtp (requires to, optional subject)
command	Shell command execution
channel	Reply on the input channel

Multiple outputs with [output] and ["output:1"].

Channels + Sessions

Jobs can listen on channels and maintain conversation history:

[input]
channel = "telegram"
trigger = "vatic"
trigger_match = "anywhere"   # "anywhere" (default), "start", or "end"

[session]
context = 20

[output]
channel = "telegram"

History summarization

Summarize results before storing them as memories -- useful when the raw output is too verbose to carry forward:

[history]
prompt = "Summarize: list the email IDs and key topics."

Building on Guix

guix shell -m manifest.scm -- sh -c "CC=gcc cargo build --release"