agentAI Agent Runtime on the BEAM

Multi-gateway, SQLite-backed, tool-augmented chat with any OpenAI-compatible API — built in Gleam on Erlang/OTP.

CLI + Daemon + Telegram

Interactive REPL for local use. Headless daemon with Telegram bot and TCP admin interface. Three run modes, one codebase.

Rich Tool System

Bash in a sandbox. Web fetch with SSRF hardening. Browser automation via Playwright. SQLite-backed memory. Code execution. Cron scheduling. All parallel-safe.

SQLite Persistence

Sessions, messages, and memories persist in WAL-mode SQLite with FTS5 full-text search, per-session token and cost tracking, crash recovery, and session lineage trees.

Guardrails & Safety

Circuit breaker detects and halts stuck tool loops. Hardline command blocks and dangerous-pattern detection. Tool-specific recovery hints. Idempotent no-progress tracking.

Config-Driven

Every knob exposed — model, timeouts, tool limits, gateway settings, DND windows, autonomy features. TOML files for settings, .env for secrets.

Extensible Plugins

Tools, Gateways, Hooks, and Memory plugins follow defined shapes. Drop custom plugins into ~/.agent/ and the registry auto-discovers them on startup.

~600

Tests

Built-in Tools

Gatways

Services

Configurable knobs

What can agent do?

Agent is designed for autonomous operation — it can run as an interactive CLI, a Telegram bot, or a headless daemon processing scheduled tasks.

In a conversation, agent can:

Execute shell commands in a sandboxed environment with approval gating
Browse the web via Playwright — navigate, click, type, scroll, screenshot
Fetch and analyze web pages with SSRF protection and content scanning
Store and retrieve memories in SQLite, with deduplication and validation
Search past conversations with FTS5 full-text search
Manage cron jobs — create, list, update, and delete scheduled tasks at runtime
Send Telegram notifications with DND-aware delivery coordination
Run code in Python, Node, or Bash

On its own, agent can:

Execute scheduled tasks every N minutes via Pulse (time-driven PULSE.md)
Reflect on conversations after each turn and consolidate memories automatically
Recover from crashes — active sessions resume on daemon restart
Auto-title sessions via LLM after the first exchange
Compress context when the conversation outgrows the model’s context window

Quick Example

# Set up
git clone https://github.com/p424p424/agent.git && cd agent
echo 'DEEPSEEK_API_KEY=sk-...' > .env

# Interactive chat
$ gleam run
You: What's in this directory?
agent: [runs ls] Here's what I found:
  src/  test/  agent.toml  gleam.toml  README.md

You: Search for the latest Gleam release notes and summarize them
agent: [fetches web page, analyzes content]
  Gleam 1.16 brings improved LSP support, faster compilation, and...

# Daemon mode (Telegram bot + admin TCP)
$ gleam run -m agent_app
# Bot is now live on Telegram

# Admin CLI
$ gleam run -m agent_admin sessions list
$ gleam run -m agent_admin db stats

Project Status

Area	Status
CLI REPL + Daemon + Telegram	Production-ready
Tool system (bash, web, browser, memory, code, cron, session search)	Complete
SQLite persistence (WAL, FTS5, sessions, memories)	Complete
Guardrails (circuit breaker, SSRF, safety scanning)	Complete
Context compression + tool loop detection	Complete
OTP supervision tree	Complete
Structured autonomy (Pulse, Reflection, Cron, Harness)	Complete
Notifications + DND coordination	Complete
Parallel tool execution	Complete
Multi-provider routing	Deferred
Streaming responses	Deferred
Sub-agent delegation	Deferred

Ready to try?

Jump to the Getting Started guide to install and run agent in under 2 minutes.