Projects & Code

Find the right building block—or a complete assistant.

/
20 projects

Runtimes & SDKs

Keep agents running and activate them on events or schedules.

Personal assistants

Context-aware assistance built around a person’s ongoing work and life.

Research implementations

Implementations of specific research ideas, not necessarily ready-made products.

Products

Commercial references; capabilities follow the linked vendor sources.

Supporting components

Memory, perception and gating tools—not complete proactive assistants.

Source guide / 完整来源说明

Projects & Products

README · Applications · Personal Agents · Streaming Proactivity · Research Map · Benchmark Matrix

This guide covers implemented resources alongside research papers. Entries are grouped by activation mechanism and role, so a scheduler, a trained intervention policy, a product, and a memory library can be compared without conflating their evidence.

Check dates: baseline entries were reviewed on 2026-10-04; WakeGate, Meta Muse, Manus Cue, and Grok Bot were checked on 2026-10-31; Claude Duet and Mistlabs Ambi on 2026-11-01. Sources are official repositories and documentation. Product capabilities are vendor-described, not independently benchmarked; source inspection is not an installation test. An older entry's date is not refreshed by this targeted update.

Persistent Runtimes and Activation SDKs

Project Role and activation mechanism Available resources Boundary / what to inspect Official sources
OpenClaw Personal-agent runtime. Heartbeat monitors check for meaningful changes; automations and incoming events can activate work. Runtime source and operational docs. Monitor activation and notification suppression are runtime mechanisms; downstream helpfulness depends on the configured agent and standing instructions. Repository · Heartbeat · Automation
Hermes Agent Agent harness with memory/skills and scheduled tasks delivered through messaging channels. Source, CLI/gateway, and cron documentation. Include as scheduled activation: a user-defined recurring job does not by itself demonstrate discovery of an unstated need. Repository · Scheduled tasks
Letta Code Stateful agent harness supporting always-on operation, persistent memory, and self-managed schedules. Source, CLI, and scheduling/permissions docs. Memory persistence and scheduled execution are distinct from calibrated intervention timing. Repository · Schedules
Proactivity SDK Framework adapter layer with durable goals, cross-wake state, model-selected cadence, and event-driven waking. TypeScript SDK, adapters, examples, and governance primitives. Inspect cadence, goal continuity, idempotency, and action caps. These controls describe orchestration, not a trained streaming model or benchmark result. Repository

Ambient and Personal Assistants

Project Observed context → proactive behavior Available resources Boundary / what to inspect Official sources
MineContext Screenshots and desktop context → background summaries, tips, todos, and resurfaced information. Desktop application, context-processing source, configuration, and releases. Ambient information delivery is the relevant behavior. Broader multi-source inputs described as future support should remain separate from current capabilities. Repository
GAIA Connected services → automatic inbox triage, meeting briefings, and schedule/event-driven workflows. Application source and self-hosting instructions. A concrete personal-workflow reference; source is under PolyForm Noncommercial, so record it as source-available. Proactive utility is described by the project, not independently measured here. Repository · License
memUBot Persistent user/team context → recurring automations and message, keyword, or platform-event-triggered skills. Assistant source built around memU. Treat the assistant and its memory layer separately. Enterprise-readiness and cost-reduction claims in the README are not validated by this list. Repository

Research Implementations

Project Proactive mechanism What to reuse Evidence boundary Official sources
Proactive Agent Predicts tasks and assistance opportunities from activity/event streams. Event collection, data generation, agent/reward checkpoints, and evaluation pipeline. ICLR 2025 method + ProactiveBench; distinguish generated training scenarios from real-user deployment outcomes. Paper · Repository · Notes
ContextAgent Uses open-world sensory context to decide on assistance and tool use. Inference, SFT, tools, prompts, and sensory-context data layout. NeurIPS 2025 research system; inspect sensing assumptions and the context-to-assistance policy. Paper · Repository · Notes
PASK IntentFlow detects latent needs and chooses silence, fast intervention, or deeper assistance with hierarchical user/workspace/global memory. Official repository, project page, system descriptions, and client download links. Technical report and preliminary user evidence. Public client/documentation links do not establish that every training component or model weight is released; no installation or evaluation was repeated here. Paper · Repository · Project · Notes

Streaming research implementations have a dedicated model and framework matrix, including JoyAI-VL-Interaction, MOSS-VL-Realtime, OneStreamer, Proact-VL, Response-G1, Eyes Wide Open, MiniCPM-o 4.5, and Gander.

Product Reference

Product Documented initiative Evidence and access boundary Official source
OpenAI dot Always-on ongoing responsibility; background research, proactive connected-app review and memory formation, suggestions, and scheduled checks between conversations. Official product documentation. User-defined rules and pause controls govern activity. No open implementation or independent proactive benchmark is established by this documentation; plan/region rollout can change. Getting started with your dot
Meta Muse Personal-agent product: a standing goal drives background work, progress tracking and suggestions; it returns when conditions change or approval is needed. Distinct from the Muse Spark model. Checked 2026-10-31: official product/app information describes a persistent Secure VM, connected services and user-controlled access. Availability varies; no open agent implementation or independent proactivity result is established here. Product · Official introduction
Manus Cue Personal-agent product with persistent identities, computers and group-chat delegation. Included as a persistent personal-agent reference; task delegation alone does not establish autonomous need detection. Checked 2026-10-31: Manus documents Cue as a standalone phone/desktop app with agent email, phone and budgeted wallet. The same announcement describes Manus Automations, but does not establish identical trigger support inside Cue. No open implementation or independent proactive evaluation verified. Product · Manus 2.0 / Cue announcement
Grok Bot Persistent cloud-agent product with named Bots, cross-session context and scheduled/event-triggered routines. Background work and escalation to the user are the relevant initiative mechanisms; not the generic Grok chatbot. Checked 2026-10-31: official documentation describes app/webhook triggers, cloud execution, user approvals and shared account-level computer access. Configured workflows are not proof of unsolicited need discovery; no open full runtime or independent proactive benchmark verified. Product documentation · Skills and routines · Approval boundaries
Claude Duet Standalone proactive agent: checks connected work apps every few hours, sends a morning brief, messages when something needs attention, and keeps working on a cloud computer while the user is away. It reads, drafts and prepares on its own; sending, paying, signing and deleting wait for approval. Checked 2026-11-01: research preview by waitlist for US Pro, Max and Ultra subscribers; Team and Enterprise are not yet available. The announcement lists iOS, Android and claude.ai, while the help center describes mobile setup only. Securities trading and moving money between accounts are blocked even with approval. Anthropic says it may misread a thread or miscalibrate urgency and is still tuning when it messages. Security testing is vendor-reported; no open implementation or independent proactive benchmark is established. Announcement · Help center · Safety
Mistlabs Ambi Ambient assistant: builds context from the conversations, screens and tools a user permits, then offers help such as turning a promise into a follow-up. Routines run longer multi-agent tasks, and it researches while the user is away. Actions run only on instruction or confirmation. Checked 2026-11-01: early access. The download page lists Apple silicon macOS, Windows 11 x64 and iOS, with Android coming; the launch release names only Apple devices. Agent-to-agent communication is planned, not shipped. Mistlabs' wearable timing conflicts between its release (later this year) and homepage (early next year). When listening is on, bystander voices may be processed in the cloud, and recording consent is the user's responsibility. Official pages give no decision mechanism, measured evaluation or independent privacy audit. Website · Launch release · Terms · Download

Supporting Components

Component Role in a proactive system What it does not establish by itself Official sources
WakeGate (苏醒之门) Condition-based wake-up gate for an existing agent, not a complete runtime. An external timer runs a model-free local judge; a state snapshot and HTTP request ask the host to resume a session. The resumed agent, not this gate, decides whether to speak or act. Checked 2026-10-31: public Python integration skeleton; context collection and host wake endpoint must be supplied. Missing conversation time can request a wake. Cooldown/retries exist, but several documented controls are not implemented; see the source audit below. Repository · Judge · Wake request
memU Shared personal memory across sessions, agents, and devices; scheduled background memorization and later retrieval. Memory formation is not an intervention policy. The current README emphasizes a shared wiki and host adapters; use memUBot for the assistant layer. Repository
Screenpipe Desktop screen/audio history, workflow mapping, and context for agents/automations. Capture and workflow discovery require an integrated agent to choose whether and when to help. Source-available, with separate licensing terms. Repository · License
Live VLM WebUI WebRTC camera interface for connecting and comparing VLM inference backends. A live interface does not make a conventional model natively streaming or give it a learned speak/silence gate. Repository

WakeGate: implemented mechanism versus integration design

The public code separates external scheduling → local wake eligibility → host session activation → agent decision. This makes it a reusable activation component, not a standalone personal assistant or trained intervention policy. Its author's L0–L4 labels are not adopted as a field-wide standard.

judge.py checks an argument/crisis flag, sleep flag, quiet hours, silence duration and optional presence. gather() returns unknown placeholders until integrated. knock.py applies cooldown and HTTP retries; sending a request does not independently verify resumed inference or delivery of useful assistance.

The earlier description that exceptions override every limit was inaccurate: the argument flag bypasses judge checks, but knock.py still enforces cooldown. Promise/forced-wake options, event listeners, hourly/daily caps and ignored-message backoff are configuration/design entries, not wired behavior in the inspected scripts. action and dry_run are copied into state but do not prevent the HTTP request. Do not rely on them as implemented no-send controls. The reported two-month Windows use concerns the author's private precursor, not an independently tested public release.

How to Add a Project

Record the resource type, activation source, context input, observable behavior, release/access status, source, and checked date. Link a concrete heartbeat/scheduler, event hook, intervention gate, or documented product workflow. Keep a component in its supporting role unless its own implementation selects user-facing interventions.

Prefer canonical upstream repositories over mirrors and forks. A paper implementation belongs here when it offers reusable code; retain its bibliographic record and decision card in the paper collection.