aegis-ag 1.0.0.dev20260419134211

Aegis CLI-first persistent agent runtime.

Aegis

Your own persistent AI.
An agent that remembers, reasons, and endures across time.

Aegis is building toward a personal AI agent that does not lose the thread when a single chat ends. The goal is not just tool calling or session orchestration, but continuity: an agent that becomes more useful as it accumulates the right history, goals, and context.

Positioning

Aegis is for long-running work that should survive interruption.

• It should remember what matters.
• It should reason across longer horizons than the current turn.
• It should recover the right context instead of collapsing when the window fills.

In product terms, Aegis is aiming at a persistent personal AI, not a disposable chat session.

Why Aegis Exists

Most current agents fail in three predictable ways:

• Amnesia: decisions, preferences, and project state disappear when the session ends.
• Myopia: current-turn optimization wins over next-week usefulness.
• Overflow: context windows fill up, and naive truncation replaces structured memory and retrieval.

The product thesis behind Aegis is that a useful agent should behave more like a persistent collaborator: it should remember, plan, and recover the right context when work resumes.

Thesis

• Long-term memory: durable episodic, semantic, and procedural memory instead of session-only state.
• Long-horizon decision making: persistent goals, task decomposition, and deadline-aware planning.
• Long-context understanding: layered context management, compression, and retrieval instead of naive truncation.

Quickstart

Aegis is used from the terminal. The path is simple: install it, run aegis, let it prepare your first named clone, and continue directly into wake. Use aegis clone <name> only when you want another Aegis individual, and give that new thread its first durable goal right away.

Install from remote

curl -fsSL https://aegis.agentic-in.ai/install.sh | bash

This installs the public aegis launcher into ~/.local/bin and prepares the default durable runtime under ~/.aegis. The public installer now tracks the latest published development package by default. If you want the latest stable package instead:

curl -fsSL https://aegis.agentic-in.ai/install.sh | bash -s -- --channel stable

Install from source

git clone https://github.com/agentic-in/aegis.git
cd aegis
bash scripts/install.sh

This creates a local launcher and the default durable layout:

• AEGIS_HOME
• AEGIS_STATE_DIR
• AEGIS_GATEWAY_STATE_DIR
• AEGIS_PROFILE_DIR

If ~/.local/bin is not already on PATH, use the launcher path printed by the install script directly.

CLI overview

aegis
aegis init
aegis wake
aegis clone nova --initial-goal "Own the release readiness thread"
aegis clones
aegis status
aegis skills
aegis gateway setup --default-clone-id demo
aegis gateway doctor
aegis gateway describe
aegis gateway feishu start --transport long-connection --detach
aegis gateway feishu status
aegis gateway feishu logs ops-feishu --follow
aegis gateway feishu restart
aegis gateway discord setup --account-id ops-discord --bot-token ...
aegis gateway discord start --transport gateway
aegis gateway discord doctor
aegis dashboard --dry-run
aegis wake --message "Who are you?"

The CLI surface stays intentionally small:

• aegis is the default entry: on the first run it routes into init, and later it re-enters wake directly or asks which clone to open.
• aegis init persists identity and provider posture, asks for the first durable goal, prepares the first clone, and now surfaces the IM chooser before wake resumes.
• aegis wake is the normal way back in and enters the current live clone directly when one exists.
• aegis wake --message "..." runs a single turn without staying in the TUI.
• aegis skills ... mirrors the skill-hub and install controls outside chat when you want to inspect, search, install, or toggle skills before opening wake.
• aegis gateway ... is the IM setup/runtime surface: use aegis gateway setup for the interactive chooser, aegis gateway doctor and aegis gateway describe for whole-surface inspection, and provider-specific commands such as aegis gateway feishu ... or aegis gateway discord ... for account setup and runtime control.
• aegis dashboard launches against the live CLI state database: repo checkouts build the dashboard once before Vite starts, while PyPI installs serve the prebuilt dashboard assets shipped in the wheel; pass --reuse-api only when you intentionally want to attach to an existing healthy API.
• aegis status checks whether the current Aegis runtime is ready to wake.
• aegis clone <name> creates another named Aegis individual only when you want a second continuity line, and requires a first durable goal either from --initial-goal or the interactive prompt.
• aegis clones lists known clones.
• aegis clones bye <name> retires one clone.
• aegis clones bye --all clears every clone.
• aegis wake --clone-id <name> opens one clone directly.

Inside wake, the conversation stays primary. Slash commands such as /status, /profile, /activity, /memory, /procedure, /audit, /skills, and /help are there when you need to inspect or steer continuity without leaving the session.

Shipped Capability Overview

Aegis already ships these core runtime surfaces on main:

• Durable memory, planning, and context recovery backing the normal wake path.
• Persistent clones with isolated continuity, personality, and canonical identity state.
• A dedicated aegis provider surface for configuring provider, model, reasoning effort, and context window together.
• Provider onboarding and a pluggable catalog covering openai-compatible, openai, openai-codex, openrouter, copilot, anthropic, claude-code, google, google-gemini-cli, groq, deepseek, xai, mistral, together, fireworks, moonshot, qwen-oauth, minimax, minimax-cn, zai, alibaba, xiaomi, huggingface, opencode-zen, opencode-go, kilocode, ollama, and vllm.
• Provider discovery now reuses configured profiles, encrypted local secrets, environment variables, local endpoint probes, and imported external auth sources such as Codex, GitHub Copilot, Qwen, and Hermes auth state.
• Provider setup and model selection now surface reasoning-effort choices for supported models instead of treating every runtime as flat chat completion.
• Built-in tools now ship as a curated runtime catalog:

• terminal: tool.terminal.exec
• process: tool.process.manage
• file: tool.file.read, tool.file.write, tool.file.patch, tool.file.search
• web: tool.web.search, tool.web.read, tool.web.extract
• browser: tool.browser.navigate, tool.browser.snapshot, tool.browser.click, tool.browser.type, tool.browser.scroll, tool.browser.back, tool.browser.press, tool.browser.images, tool.browser.vision, tool.browser.console
• clarify: tool.clarify
• cron: tool.cron.manage
• code_execution: tool.code.execute
• memory: tool.memory.recall, tool.memory.upload
• messaging: tool.message.send
• todo: tool.todo.manage
• skills: tool.skill.list, tool.skill.view, tool.skill.manage
• sub_agents: tool.sub_agents
• continuity-native: tool.profile.manage, tool.activity.manage, tool.procedure.inspect, tool.procedure.manage

• Built-in skills for shell, workspace search, web search/read, continuity, cron scheduling, and voice reply guidance.
• Local skill-hub search/install plus Aegis-authored skill creation inside the shared shelf.
• Durable agent-run checkpoints for resumable multi-step tool loops and long-horizon follow-up.

The public docs now cover the supported operator path plus the currently shipped provider, capability, automation, and extension surfaces. Internal design docs under docs/system-design/ still describe the broader target architecture.

Contributors