Tars

Tars is a compact, local-first AI assistant for Discord and terminal workflows. It runs as a bare-metal Node.js process managed by PM2, supports cloud providers and OpenAI-compatible local endpoints, and adds durable memory, explicit scheduled tasks, local skills, and MCP extensions.

Tars is designed for operators who want a small self-hosted assistant with a transparent workspace at ~/.tars/. It is competitive with broader local-assistant projects such as OpenClaw in this narrower operating model; it does not claim parity with their multi-agent, mobile, voice, or broad multi-channel ecosystems.

What it provides

• Discord and interactive terminal interfaces backed by one active agent.
• Google, OpenAI, Anthropic, local, and custom OpenAI-compatible model configuration.
• Durable facts, searchable notes, and scheduled tasks through built-in MCP extensions.
• A PM2-managed supervisor with explicit cron execution, maintenance, and bounded initiative modes.
• Local skills and explicitly enabled MCP servers with restricted subprocess environments.
• Backup and restore commands with secret exclusion by default and transactional imports.
• Defensive defaults for Discord ownership, dashboard access, outbound fetching, attachments, and sensitive log redaction.

Tars is not a sandbox. Model tools and trusted extensions run with the permissions of the operating system user. Use a dedicated least-privilege account, review installed skills and extensions, and keep credentials out of prompts and repositories. See the security model.

Install

Prerequisites:

• Node.js 22.19 or newer
• npm 10.9 or newer
• A supported model provider or an OpenAI-compatible endpoint
• A Discord bot token and preconfigured owner user ID for the Discord channel

npm install -g @saccolabs/tars
tars setup
tars extensions audit
tars start
tars status

Use tars chat --no-discord for a foreground terminal session with Discord disabled. It does not start the background heartbeat, cron service, dashboard, or a second Discord connection. It still uses the configured TARS_HOME. Run tars stop first: Tars refuses foreground chat while a matching PM2 engine is active and holds a cross-process home lease for the full terminal session. Exit chat before running setup, start, stop, restart, import, export, refresh, update, uninstall, secret writes, or memory index operations; those commands refuse a live lease.

Common commands

Run tars --help for the complete command reference.

Keep secret values out of shell history, process arguments, and tool logs. Read a value without echoing it, pipe it to Tars, and then clear the temporary shell variable:

read -rs TARS_SECRET_VALUE
printf '%s' "$TARS_SECRET_VALUE" | tars secret set KEY
unset TARS_SECRET_VALUE

Custom extensions must explicitly list the host environment-variable names they receive. Audit them with tars extensions audit; use the guided tars extensions migrate flow for older entries. The migration scans extension source for likely names but requires an operator decision and never prints or copies secret values. tars restart launches the same review when required and continues the restart only after every blocking policy is resolved.

Common memory, search, and task tools stay directly available to the model. Specialized MCP schemas are discovered and invoked through a compact catalog, avoiding the recurring context cost of every browser, brokerage, commerce, and health tool definition on unrelated turns.

Runtime layout

Discord / terminal
        │
        ▼
Channel manager ──► Supervisor ──► Tars engine / model
                         │                    │
                         │                    └── MCP extensions
                         ├── active session
                         ├── cron service
                         └── maintenance heartbeat

~/.tars/
├── config.json          non-secret settings
├── .env                 local secrets (restricted permissions)
├── system.md            system instructions
├── skills/              trusted instruction packages
├── extensions/          trusted MCP servers
├── chats/               conversation history
├── data/                sessions, tasks, memory, and indexes
└── logs/                redacted operational logs

Configuration precedence is exported environment variables, then ~/.tars/.env, then ~/.tars/config.json, then validated defaults. Values such as intervals, context limits, compression thresholds, and rate limits are bounded during loading.

Documentation

• User documentation
• Architecture
• Development guide
• Operations guide
• Contributing

Build the documentation locally with npm run docs:build or run it with npm run docs:dev.

Project status and boundaries

• Discord is the supported daemon channel; terminal chat is available in the foreground.
• Foreground chat uses the same TARS_HOME; daemon/chat and chat/mutation exclusivity are enforced.
• Heartbeat maintenance continues while the owner is idle. The initiative service can observe, propose, or apply registered safe repairs from explicit objectives and runtime findings.
• Scheduled tasks are explicit and poll on a separate cron loop.
• Tars has one active agent and does not implement sub-agent orchestration.
• Multiple PM2 process names can be created, but the convenience lifecycle commands are not fully instance-aware. Treat multi-instance deployments as advanced and isolate each TARS_HOME.
• Redaction reduces accidental disclosure in logs and events; it is not data-loss prevention, prompt-injection protection, or a security boundary.

Development

npm ci
npm run ci:extensions
npm run check
npm run build
npm run test:extensions

See CONTRIBUTING.md before submitting a change. Releases and npm publishing are managed by Release Please; do not edit the package version manually.

License

MIT © Agustin Sacco