BetaEditorial Listing

OpenHuman

TinyHumans · OpenHuman: Open-Source Desktop AI Agent with Persistent Memory, App Integrations, and Workflows

Open OpenHuman

OpenHuman is an open-source (GPL-3.0) personal AI agent for macOS, Windows, and Linux that connects your email, chat, and work apps, keeps a persistent memory of your documents, conversations, and preferences, and answers questions, runs workflows, and acts across those apps with your approval. It suits individuals who want a desktop agent that knows their context without wiring API keys or running a server.

PricingFreemium
Setupeasy
Runs onDesktop · Self-hosted
APIYes
Open sourceYes
DocsYes
CategoryProductivity
Open SourcePersonal AssistantDesktop AppPersistent MemoryWorkflow AutomationMulti-AgentVoiceMCPBYOKLocal ModelsDesktop AgentAgent Harness

Best for

Individuals and solo operators on macOS, Windows, or Linux who want a desktop agent that knows their email, chat, and work apps and can act across them, without wiring API keys or running a server

Not ideal for

Teams that need a supported, stable product today, people who want an assistant that lives only inside chat apps on a server, mobile-first users, and anyone who needs memory and voice to stay on their own machine without running their own CortexDB

Who it's for

Knowledge workers, founders, and technical individuals who want a personal desktop agent with persistent memory of their own apps and data

Capabilities

  • Memory engine that recalls a short, token-budgeted memory pack with citations before each reply, drawn from shared documents, the agent's earlier conversations, and saved learnings, and builds beliefs from them in the background
  • Two memory engines: TinyHumans' hosted CortexDB (requires sign-in) or your own CortexDB endpoint, managed or self-hosted; with neither, memory is off
  • Brain of shared documents fed by synced sources (folders, files, web links, GitHub repositories, RSS feeds) on a schedule you set per source, plus a one-time import of memory from earlier versions
  • 119 third-party integrations (Gmail, Outlook, Google Calendar, Drive, Notion, GitHub, Linear, Jira, Slack, Salesforce, HubSpot, Stripe, Shopify, and more) through one-click OAuth brokered by Composio, or a direct mode with your own Composio key
  • Connected services work as agent tools and as live triggers (a new email, a Notion edit, a GitHub issue) that can fire agent actions; services with a native provider, such as Gmail, can also sync into memory
  • Agent-proposed workflows on the open-source tinyflows engine: describe an automation in chat, review the graph on a canvas, then save it to run on schedules or app events with approval steps
  • Orchestrator that spawns sub-agents up to three levels deep, with checkpointed runs that pause for your input and resume, and a circuit breaker that stops repeated-call loops
  • Built-in tools: web search, web scraper, coder toolset (filesystem, git, lint, test, grep), a real Chrome or Chromium browser driven over the DevTools protocol (fresh or attached to the one you already run), desktop control on macOS and Windows, cron scheduling, and memory tools
  • Voice: dictation, spoken replies (with local Piper text-to-speech as an option), and live voice agents you can talk to and interrupt, on Gemini Live (managed or your own Google key), an ElevenLabs agent, or Sarvam AI for Indian languages
  • Approval gate that classifies tool calls (read, write, network, install, destructive) and holds side-effecting actions until you approve
  • Model choice: managed TinyHumans routing with automatic per-task model selection, 26 bring-your-own-key providers (OpenRouter, OpenAI, Anthropic, Google, Groq, Mistral, DeepSeek, and others), Claude Code or the Claude Agent SDK, or local runtimes you run yourself (Ollama, LM Studio, MLX, OMLX)
  • Privacy Mode with a local_only setting enforced in the Rust core that refuses cloud model calls, network tools, web search, integrations, cloud embeddings, and the hosted memory engine
  • Two-way messaging channels: Telegram, Discord, iMessage, Lark/Feishu, DingTalk, and Yuanbao have in-app setup, and Slack, WhatsApp, Signal, Mattermost, IRC, QQ, Linq (SMS), and email (IMAP/SMTP) are also supported, several of them configured by hand in config.toml
  • MCP server registry and a browsable catalog of about 90,000 SKILL.md skills
  • Headless core can run on a Linux server or VPS (Docker Compose, Fly.io, DigitalOcean) with the desktop app connecting to it, and an embeddable Rust library for other products

Limitations

  • Labelled early beta in its README, which warns users to expect rough edges. The newest published release is v0.64.15 (October 9, 2026); development has continued on the main branch since then
  • The default setup routes sign-in, model calls, memory, managed integrations, real-time integration triggers, and web search through TinyHumans' hosted backend
  • Memory is not stored on your computer: items live in CortexDB, hosted by TinyHumans or at an endpoint you run, and memory is off if you are signed out with no CortexDB key. Switching engines does not migrate existing memory
  • Version 0.64.14 removed several earlier features, including the local Memory Tree and Obsidian vault, the Supermemory, Mem0, and Cognee memory engines, the Google Meet agent, and in-app downloads of local models
  • Voice is not covered by local-only Privacy Mode: there is no local speech-to-text engine, so dictation and voice-agent audio still go to the hosted proxy or a third-party provider
  • Not every integration feeds memory: only services with native providers (such as Gmail) can be synced into memory, while others are callable as tools only
  • Local models are bring-your-own: you install and run Ollama, LM Studio, MLX, or another server and pull models with its own tools
  • Browser automation and desktop control are not available in the Linux build
  • The iOS companion app is experimental and not part of the shipped product, so there is no supported mobile client
  • A remotely hosted core does not serve the full app as a web UI; desktop features still need the Tauri app
  • Licensed GPL-3.0, which matters if you plan to embed or redistribute the core inside your own product
  • On Linux, the AppImage can fail on some newer distributions, where the .deb package is recommended

Use cases

  • Asking for a morning briefing of what changed across connected Gmail, Slack, and Calendar accounts
  • Querying everything a colleague said about a project across email and chat
  • Having the agent draft a workflow that summarizes new customer emails and posts them to Slack, then reviewing and saving it
  • Talking a task through with a live voice agent and interrupting it while it works
  • Running a private assistant on a local Ollama model and a self-hosted CortexDB with Privacy Mode set to local-only
  • Letting the agent tidy and rename a project folder inside a boundary you set, with each change approved

Our take

OpenHuman's pitch is that a personal agent is only as useful as what it already knows, so it wires up the context for you: connect Gmail and Slack once through OAuth and the agent can search them, act on them, react to their events, and remember what matters. That, plus a real desktop UI, agent-drafted workflows you approve on a canvas, live voice agents, and a default managed account that hides API keys, is what separates it from OpenClaw and Hermes Agent, which are built around chat apps or a terminal and are usually run on your own model keys. The convenience has a cost: the easy path runs inference, memory, and integrations through TinyHumans' backend, and since the October 2026 memory rewrite memory sits in CortexDB rather than in local files, so keeping data on your own machine means local models, a self-hosted CortexDB, and Privacy Mode, with fewer features. It is also early beta and changes quickly, with features removed as well as added between releases, so it suits people comfortable with change more than teams that need stability.

Who should use it

Individuals who live in Gmail, Slack, Notion, and GitHub and want an agent that already has that context, people who prefer a desktop UI over configuring an assistant in a terminal, and privacy-minded users willing to run local models and their own CortexDB with Privacy Mode.

Who should skip it

Teams that need vendor support and a stable release, users who want a chat-app-only assistant on an always-on server (OpenClaw or Hermes Agent fit better), people who need a mobile app, people who want their agent's memory as local files they can edit by hand, and companies that cannot accept GPL-3.0 for embedding.

Strengths

  • Memory recalls cited context from your documents, past conversations, and saved learnings before each reply
  • Desktop app with one-click OAuth integrations, so no API keys are needed on the default path
  • Agent-drafted workflows that you review and approve before they run
  • Choice of managed models, your own keys, or local runtimes, with an enforced local-only mode
  • Approval gate on side-effecting actions by default
  • Open source under GPL-3.0 with very active development

Weaknesses

  • Early beta with frequent releases and rough edges
  • Default path sends inference and integration calls through TinyHumans' servers
  • Memory lives in CortexDB, hosted by TinyHumans unless you run your own, and the local Obsidian vault was removed
  • Voice and some triggers still depend on hosted services
  • No supported mobile app yet

OpenHuman pricing

Open source / pay as you go

Free

  • OpenHuman software is free under GPL-3.0
  • Bring your own provider keys or run local models at no TinyHumans cost
  • Managed usage billed from a credit balance topped up from $5
  • Optional auto-recharge when the balance runs low

Basic

$20 ($200/year)

Billed monthly

  • $21 of credits every month
  • LLM, embeddings, voice, and OAuth APIs through one TinyHumans key
  • 1 GB of memory storage

Pro

$200 ($2,000/year)

Billed monthly

  • $210 of credits every month
  • Higher rate limits
  • 20 GB of memory storage

Enterprise

Custom

  • Custom usage
  • On-premise deployment
  • SSO / SAML, dedicated onboarding, and priority support

Free tier limits: No subscription is required. The app runs free with your own keys or local models; managed inference, hosted memory, and other services draw from a prepaid credit balance.

Note: Verified November 8, 2026. Plans are TinyHumans account plans shared across its products and are sold as credits, where one credit is one dollar of metered usage; the pricing page describes them in terms of TinyHumans' OpenCompany, Medulla, and API products without naming OpenHuman, and OpenHuman's managed services run on the same TinyHumans key. The page says pay-as-you-go credits cost the same per call as plan credits. OpenHuman's own billing documentation lists Basic at $19.99 a month ($199 a year) and Pro at $199.99 a month ($1,799.99 a year) with per-call discounts instead of bonus credits, so confirm the current terms on TinyHumans' pricing page. Calls made with your own provider keys are billed by that provider. TinyHumans states that subscription and credit purchases are non-refundable.

Technical specs

Modalities

Text, Voice, Vision, Browser automation, Computer use

Available models

TinyHumans managed routing (automatic per-task model selection)26 BYOK providers including OpenRouter, OpenAI, Anthropic, Google, Groq, Mistral, DeepSeek, Together, and FireworksClaude Code / Claude Agent SDKOllama, LM Studio, MLX, OMLX, and local OpenAI-compatible servers (run yourself)

Where OpenHuman excels

A daily briefing across inbox and chat

Because Gmail, Slack, and Calendar are connected as tools, and Gmail can also sync into memory, a question like "what do I need to know from the last 12 hours?" can be answered from your own data without copying anything in.

Turning a repeated chore into a workflow

You describe the automation in chat, the agent drafts a workflow graph, and nothing runs until you save it on the canvas. Runs can pause at approval steps and resume, with step-by-step history.

Hands-free work with a live voice agent

Live voice agents on Gemini Live, ElevenLabs, or Sarvam AI let you talk to the agent in real time and interrupt it while it uses tools. The audio goes to the provider you pick rather than staying on the device.

A private assistant on local models

With Ollama, LM Studio, or MLX, a self-hosted CortexDB for memory, and Privacy Mode set to local-only, the core refuses cloud model, search, integration, and hosted memory calls, which suits sensitive personal or client material (voice excepted).

OpenHuman vs. competitors

OpenHuman vs. OpenClaw

OpenClaw is an MIT-licensed assistant that routes messages from 20+ chat apps to agents you configure with your own model keys or local models, with skills and cron. OpenHuman is a GUI-first desktop app that connects your apps through OAuth, keeps a persistent memory, adds agent-proposed workflows and an approval gate, and defaults to a managed TinyHumans account. OpenClaw fits people who want an assistant inside their chat apps under full self-managed control, and OpenHuman fits people who want app context and automations without configuring providers.

OpenHuman vs. Hermes Agent

Hermes Agent is an MIT-licensed agent you run on your own machine or server with any model, building memory and skills from its own sessions and reachable from 20+ messaging platforms. OpenHuman is designed around a desktop GUI with OAuth-connected apps, live triggers, and visual, agent-proposed workflows, and uses managed TinyHumans inference and memory by default. Hermes suits developers who want a long-running server agent they control end to end, and OpenHuman suits individuals who want a desktop agent that already knows their inbox and tools.

OpenHuman vs. Claude Cowork

Claude Cowork is Anthropic's proprietary agent inside the Claude app that completes multi-step work across your files and connected tools on Claude models. OpenHuman is open source, lets you choose managed, bring-your-own, or local models and hosted or self-hosted memory, and adds saved workflows, integration triggers, and messaging channels. Cowork suits people already on a Claude plan who want a polished, supported product, and OpenHuman suits people who want an open, model-flexible desktop agent.

Frequently asked questions

What is OpenHuman?

OpenHuman is an open-source personal AI agent from TinyHumans that runs as a desktop app on macOS, Windows, and Linux. It connects apps such as Gmail, Slack, and Notion, keeps a persistent memory of your documents, conversations, and preferences, answers questions across them, runs workflows, and takes actions with your approval.

Is OpenHuman free and open source?

The software is free and open source under GPL-3.0. Managed model inference, memory, web search, voice, and integrations on the default path draw on a TinyHumans account, which has pay-as-you-go credits and paid plans. You can also bring your own provider keys or run local models.

How much does OpenHuman cost?

The app is free. TinyHumans' pricing page lists pay-as-you-go credit top-ups from $5, Basic at $20 a month ($200 a year) with $21 of monthly credits and 1 GB of memory storage, Pro at $200 a month ($2,000 a year) with $210 of monthly credits and 20 GB, and custom Enterprise plans. That page now describes the plans in terms of TinyHumans' OpenCompany, Medulla, and API products without naming OpenHuman, and OpenHuman's own billing documentation lists slightly different prices with per-call discounts, so check the current terms before buying.

Can OpenHuman run fully locally?

Partly. Privacy Mode's local_only setting refuses cloud model calls, web search, integrations, cloud embeddings, and the hosted memory engine, allowing only local runtimes such as Ollama, LM Studio, and MLX that you run yourself. Memory then works only with a CortexDB endpoint you host, because memory is no longer stored in a local database. Voice is the exception: there is no local speech-to-text engine, so dictation still leaves the device. Sign-in and billing calls also continue.

OpenHuman vs OpenClaw: what is the difference?

OpenClaw is an MIT-licensed assistant that runs locally and works mainly through chat apps, using model keys or local models you configure. OpenHuman is a GPL-3.0 desktop app that connects your apps through one-click OAuth, keeps a persistent memory, adds agent-proposed visual workflows, and by default uses a managed TinyHumans account for models, memory, and integrations.

Which models does OpenHuman support?

By default it uses TinyHumans' managed routing, which picks a model per task. You can also connect 26 bring-your-own-key providers including OpenRouter, OpenAI, Anthropic, Google, Groq, Mistral, and DeepSeek, use Claude Code or the Claude Agent SDK, or point it at Ollama, LM Studio, MLX, or another local OpenAI-compatible server that you run yourself.

Integrations & fit

GmailOutlookGoogle CalendarGoogle DriveGoogle DocsNotionGitHubLinearJiraSlackDiscordMicrosoft TeamsTelegramWhatsAppiMessageSignalSalesforceHubSpotStripeShopifyComposioCortexDBElevenLabsOpenRouterOpenAIAnthropicGoogle GeminiOllamaLM StudioMLXMCP servers
Good fit forSolo / individual, Startup / small team
Pricing modelFreemium· Free tier available
See pricing on OpenHuman →

Alternatives to consider

About OpenHuman

OpenHuman, built by TinyHumans, is a Tauri desktop app around a Rust core, and its developers now also pitch that core as a lightweight open-source agent harness for running many agents on a small server. After sign-in you connect services such as Gmail, Slack, GitHub, Notion, or Google Calendar through one-click OAuth (119 Composio toolkits in the app's catalog). The agent can call them as tools, react to their live events as triggers, and, for services with a native provider such as Gmail, sync them into memory on a schedule you set. Memory was rebuilt in v0.64.14 (October 8, 2026): before each reply OpenHuman recalls a short, cited memory pack from shared documents (folders, files, links, GitHub repositories, and RSS feeds you add), the agent's earlier conversations, and saved learnings. Those items are stored in CortexDB, either TinyHumans' hosted engine or your own CortexDB endpoint, not in a database on your machine, and the earlier local Memory Tree and Obsidian vault were removed. On top sit an orchestrator that spawns sub-agents with pausable, resumable runs, agent-proposed workflows that you review on a visual canvas before saving, a built-in toolbelt (web search, scraper, coder tools, a real Chrome browser, desktop control on macOS and Windows, cron), voice dictation, and live voice agents you can talk to in real time. Side-effecting actions such as sending an email or running a shell command go through an approval gate by default. The difference from OpenClaw and Hermes Agent is mostly about setup and who supplies the services: those are built around chat apps or a terminal and are usually set up with your own model keys, while OpenHuman is a GUI-first desktop app that, by default, routes models, memory, search, voice, and integrations through a TinyHumans account, with bring-your-own-key, local models you run yourself (Ollama, LM Studio, MLX), and an enforced local-only Privacy Mode as options. The same core also ships as a terminal client, a headless server, and an embeddable Rust library. The tradeoffs: the README still labels it early beta, the easy path depends on TinyHumans' backend, keeping memory off TinyHumans' servers means running your own CortexDB, voice still leaves the device even in local-only mode, and the project is GPL-3.0 rather than MIT.

Updates from OpenHuman

New Featurev0.64.14: Memory v2, live voice agents, and a Files folder

OpenHuman v0.64.14 replaced the local Memory Tree with an engine-based memory stored in CortexDB (TinyHumans-hosted or your own), added live voice agents on Gemini Live, ElevenLabs, and Sarvam AI, a visible Files folder for agent outputs, and cron output delivered back to the requesting conversation. The release also removed the Google Meet agent, the Obsidian vault, third-party memory engines, automatic syncing of Composio integrations into memory, and in-app local model downloads. A v0.64.15 hotfix followed on October 9.

New Featurev0.64.0 "The Intelligence Upgrade"

OpenHuman v0.64.0 merged 674 pull requests, adding always-on listening and global push-to-talk for hands-free desktop control, faster command routing, and upgrades to memory sync, ingest, and retrieval. Point releases continued through v0.64.10 on September 30.

Are you the founder? Claim this listing →