written in Rust · MIT License · v0.11.33

AIdaemon — Durable, Self-Hosted AI Agent Runtime

Give work to an agent system that can finish it — and prove it.
Plan durable tasks, coordinate specialists, execute in isolated workspaces, and resume after interruptions with every side effect backed by a typed receipt.
Reach it from Telegram, Slack, or Discord.

View on GitHub $ cargo install

~/aidaemon

$ cargo install aidaemon
Installing aidaemon v0.11.33
Installed package aidaemon v0.11.33

$ aidaemon
✓ Semantic task contracts ready
✓ Durable work coordinator active
✓ Typed receipts + replay protection active
✓ Local workspace connected
✓ Memory + exact history ready
✓ Telegram / Slack / Discord connected
◆ daemon active — agents may resume safely
$

ChatGPT subscription, native providers, local models, and any OpenAI-compatible API

ChatGPT subscription

Anthropic

Google Gemini

xAI / Grok

OpenRouter

Ollama + local

Agent control from Telegram

Run coding agents from your phone. Keep every workspace under control.

Launch Claude, Codex, Gemini, or OpenCode in a real terminal inside Telegram. Coordinate parallel specialists, isolate collision-prone work with Git worktrees, and watch verified progress from anywhere.

end-to-end encryptedself-hostableclaudecodexgeminiopencode

Self-hosting guide

✕aidaemon - coding⌄⋮

AIdaemon.aiactive

Reviewing PR #47: “Add user auth middleware”

• Ran git diff main...feature/auth

└ 4 files changed, 186 insertions(+)

• Ran cargo clippy --all-features

└ warning: unused import (auth.rs:3)

• Applied fix → src/auth.rs

└ Removed unused import, added error handling

• Ran cargo test -- auth

└ 9 passed ✓

✓ Committed: “fix: address PR review”

Ask or run a command...Send

Features

A Durable Runtime for Agents That Do Real Work

The model interprets what you want. A deterministic kernel controls what can happen and records what actually did.

Durable work coordination

Goals become persistent projects, runs, tasks, journals, and handoffs. Interrupted idempotent work is re-queued; ambiguous writes pause for verification.

projects + runs + tasks

Parallel specialist teams

Assign research, code, review, or artifact specialists. Run up to four independent tasks in parallel while dependent work stays ordered.

specialists ×4

Workspace isolation

Choose shared, isolated, or Git worktree scope per task, then execute locally, over SSH, or inside an optional Docker workspace.

local + ssh + docker

Evidence-backed execution

Every tool action emits a typed receipt with its effects, outcome, and stable operation identity. Completion comes from evidence, not confident prose.

typed receipts

Resume without repeating

Approvals, partial outcomes, and resources live in the event log. After an interruption, AIdaemon reconciles the last receipt before acting again.

replay + resume

Source checkpoints

Opt-in checkpoints capture eligible source before the first mutation, with conflict-aware rollback that preserves independent edits.

checkpoints + rollback

Memory with provenance

Facts, episodes, entities, relationships, and exact retained history stay searchable with sources, timestamps, and scoped visibility.

graph + fts5 + vectors

Vision, voice, and image generation

Analyze images, understand or transcribe audio, inspect screenshots, and generate images through a connected ChatGPT subscription.

vision + audio + imagegen

Any model — even your ChatGPT plan

Use Anthropic, Gemini, xAI, local models, OpenAI-compatible APIs, or sign in with ChatGPT Plus, Pro, or Business instead of an API key.

chatgpt oauth + any llm

Browser and desktop control

Browse, fill forms, inspect console and network errors, and control native macOS apps with per-action risk checks and verification.

browser + computer use

Scoped team collaboration

Slack owners can grant time-limited read or edit access to one project. Every queued message keeps the authority of its actual sender.

scoped slack grants

Skills and integrations

Connect MCP servers, APIs, and OAuth profiles. Install skills from registries or promote proven procedures into reusable workflows.

mcp + oauth + skills

Scheduled automation

Natural-language and cron goals run durably, recover from stalls, honor budgets, and alert you when repeated failures need attention.

schedules + budgets

Always on and observable

Run as a background service with rotating logs, an authenticated dashboard, a live work board, health checks, and diagnostics.

dashboard + logs

Talk from anywhere

Use Telegram, Slack, or Discord with multiple bots, images, voice notes, file delivery, progress updates, and approval cards.

telegram + slack + discord

Agent runtime

Inside the AIdaemon Runtime

Natural-language reasoning stays flexible; execution stays deterministic. Task contracts, scoped operations, typed receipts, and durable state make long-running work resumable and auditable.

Telegram ──┐ Slack ──┼──> Conversation + Resources Discord ──┘ │ ▼ Semantic Task Contract │ ▼ Durable Work Coordinator ├── projects / runs / tasks ├── specialist agents ×4 └── shared / isolated / worktree Model proposes operation │ ▼ Deterministic Kernel schema • auth • scope • approval idempotency • policy │ ▼ local / SSH / Docker tools + MCP / APIs │ ▼ Typed Receipt + Resource Registry │ ▼ Durable Event Log / Run State └── replay • resume • verify Work board ──> /work + /dashboard Health ──> /health

Understand the request

Conversation and registered resources become a versioned semantic task contract with concrete observations, effects, constraints, and task shape.

Plan durable work

The coordinator creates projects, runs, tasks, acceptance criteria, journals, and handoffs that survive restarts.

Coordinate specialists

Named agents take independent work in parallel and use shared, isolated, or Git worktree workspaces according to collision risk.

Validate every operation

A deterministic kernel checks schemas, authorization, scope, approvals, policy, and idempotency before any tool performs I/O.

Execute where the work lives

Files, terminal, Git, CLI agents, MCP, and APIs run through one coherent local, SSH, or optional Docker execution backend.

Record typed evidence

Each result becomes a typed receipt. Files and generated artifacts receive opaque resource IDs, provenance, MIME metadata, and integrity hashes.

Resume and prove completion

The durable event log supports replay and interruption-aware resume. The agent can only claim completion when matching effect evidence exists.

Quickstart

Install AIdaemon in Minutes

Install one binary, run the setup wizard, choose your model and channels, then leave the daemon running.

Install
curl -sSfL https://get.aidaemon.ai | bash
Or: brew install davo20019/tap/aidaemon · Or: cargo install aidaemon

Run the setup wizard
On first run, aidaemon launches an interactive wizard to configure your LLM provider and channel tokens.

Use your ChatGPT subscription (optional)
aidaemon auth login openai lets a Plus, Pro, or Business plan power chat and image generation without an API key.

Install as service
aidaemon install-service
Runs as systemd (Linux) or launchd (macOS).

config.toml

$ aidaemon auth login openai
✓ ChatGPT subscription connected

[provider]
kind = "openai_chatgpt"
reasoning_effort = "medium"

[execution]
backend = "local"
workspace_root = "/path/to/projects"

[checkpoints]
enabled = true