Features Overview

Anyy is a resident personal assistant: one runtime, persistent context, and multiple ways to work with it.

Automatic Behavior at a Glance

Some Anyy behavior happens without a separate command. These mechanisms are part of the corresponding feature rather than hidden maintenance jobs:

Behavior	When it happens	Durable effect	Details
Proactive memory edit	During an ordinary model turn, when Anyy judges a confirmed fact or preference useful later	May add, replace, or remove an entry in USER.md or MEMORY.md	Memory
Automatic session title	After the first successful user/assistant exchange in a new, untitled session	Replaces New session with a short title; never overwrites a manual title	Conversations & Sessions
Context compaction	Before a new run when estimated context reaches the model's pressure threshold	Stores a summary for future model input; does not delete the transcript	Conversations & Sessions
Skill self-improvement review	After 15 eligible foreground model-loop iterations when managed Skill editing is available	May update the Profile's managed Skill library in a separate background session	Skills
Model metadata refresh	In the background at Gateway start when capability caches need refreshing	Updates Profile-local model capability caches; does not send conversation content or run inference	Models & Providers
Background result handoff	When delegated or background work reaches a terminal result	Adds a result to the parent session inbox and may notify the origin	Background Jobs
Heartbeat reporting	Once a minute while the resident runtime is healthy	Updates liveness state and reports a newly changed failed-Cron signature through the configured route	Heartbeat
Storage housekeeping	At Gateway startup, every 24 hours, and on object-store cleanup signals	Expires temporary data and reclaims unreferenced storage without selecting canonical user records	Maintenance

Model-generated titles, context summaries, and Skill reviews use the selected provider and can add token usage and cost. Memory is different: Anyy may use the memory tool inside the foreground run, but there is no scheduled “every N turns” memory model pass.

Conversations

• Conversations & Sessions — continue durable foreground and background work across surfaces and restarts.
• Queue and steer — keep later work separate or redirect the active execution.
• Clarifications and attachments — answer structured questions and add scoped files to a request.
• Session execution modes — choose approval or yolo for each conversation.

Identity and Personal Context

• Roles — maintain different assistant identities and Soul instructions.
• Memory — durable personal context separate from one conversation transcript.
• Profiles — isolated Anyy homes for different people, purposes, or environments.
• Workspace context — store task files and artifacts used by the assistant.

Models and Media

• Models & Providers — provider authentication, model selection, reasoning controls, and ordered availability routing.
• Media understanding — route image, video, and audio input through declared capable models.
• Image generation — create images with a configured output backend.
• Speech — optional text-to-speech and speech-to-text routes.

Tools and Integrations

• Web search — current public-information retrieval through a configured backend.
• Browser — an explicitly connected isolated browser or compatible browser backend.
• MCP — add, authenticate, probe, enable, and call external MCP servers/tools.
• Skills — install reusable instruction packages, invoke their commands, and understand managed self-improvement.
• Files and terminal — inspect a Workspace, propose edits, and run approved commands.

Messaging

• Messaging Platforms — Telegram, Discord, Slack, WhatsApp, Facebook, Email, SMS, DingTalk, Feishu/Lark, Weixin/WeChat, QQ Bot, IRC, and LINE.
• Access policies — pairing, allowlists, open/disabled delivery, and group mention rules where supported.
• Multiple accounts — keep platform/account connections under the resident Gateway.

Automation

• Scheduled automations — prompts, Skills, scripts, or notifications with an explicit approval or yolo mode.
• Heartbeat — resident runtime liveness reporting.
• Background sessions — run work independently of the foreground chat.
• Tasks — durable structured obligations with status, priority, due time, and labels.
• Notifications — route results to the current user or trusted Channel origin.

Trust and Recovery

• Approvals and ChangePlans — review state-changing work before it runs in approval mode.
• Audit history — retain authorization and execution records for inspection.
• Backups — create and restore Profile archives; credentials are opt-in.
• Maintenance — checkpoint or compact state and rebuild token-estimate caches.
• Doctor/status — inspect configuration and runtime health before changing anything.

Add capabilities in layers

Start with one tested main model and a persistent chat. Add one Channel, tool backend, Skill, MCP server, or scheduled job at a time, and verify it before adding the next.

See Quickstart for the complete first-run path.