Hosted and self-hosted platforms and apps, from agent runtimes and workflow builders to self-hosted services an agent can work with.
65 entries. Latest source check .
What a source check covers
Documentation checked means the source was read. It does not mean the software was installed, tested or security audited.
Reading an entry does not authorise installation, credential access or spending. Tech Horizon Labs projects are labelled, and there are no affiliate links or paid placement.
MCP clients, aggregators and people looking up a server by name.
Access and use
The registry is in preview. Breaking changes or data resets may occur before general availability. Listing a server is not a security audit. Installing a listed server still needs its own credentials and terms.
People who already have provider accounts and want one local endpoint with fallback.
Access and use
Install with npm install -g omniroute. It runs locally (default http://localhost:20128). Bring your own provider credentials. OmniRoute does not grant tokens. MIT.
Teams who want to train and score a specialised agent from their own traces.
Access and use
SDK and CLI: pip install overmind, then overmind init and overmind sync. Hosted console is console.overmindlab.ai with a project API key. MCP is https://api.overmindlab.ai/api/mcp/ with X-Api-Key. Self-host is docker compose; the API needs OpenRouter, Modal, AWS and Hugging Face keys before it boots. SDK and CLI are MIT. The platform is AGPL-3.0. SDK and CLI send anonymous usage to PostHog; opt out with OVERMIND_ANALYTICS_ENABLED=false or DO_NOT_TRACK=1.
Teams that want a clone-and-own agent workspace on their own infrastructure, with their own model keys.
Access and use
Clone https://github.com/CopilotKit/OpenBot. Needs Docker, Bun 1.3 or later, a CopilotKit Intelligence project, and a model key. Default OPENBOT_SINGLE_USER admits every request as one administrator until you turn sign-in on. Credentials are encrypted at rest. Published image ghcr.io/copilotkit/openbot. MIT. Marked beta.
People who want several CLI coding agents working in parallel on one machine, with a phone app to follow them.
Access and use
Download from https://onorca.dev/download or GitHub Releases. macOS Homebrew is brew install --cask stablyai/orca/orca; a bare brew install --cask orca is a different Plotly app. Mobile companion is on the App Store and as an Android APK. The README points to privacy and telemetry docs for usage data and how to opt out. MIT.
People who want answers from their own files that stay current as those files change, on a machine they control.
Access and use
pip 23.1 or later. pip install serviette. Needs a free Pathway licence in PATHWAY_LICENSE_KEY from pathway.com/framework/get-license (GitHub or LinkedIn sign-in). Default server binds localhost:8989 with no built-in auth. Local embeddings extra is serviette[local]. MIT.
People who want a local super-agent stack they can configure with their own model keys.
Access and use
Clone https://github.com/bytedance/deer-flow then make setup. Docker is make docker-init and make docker-start. Python 3.12 or later and Node 22 or later for a local run. The Docker stack and local start bind to 127.0.0.1. Treat Gateway admin as code execution on the host. Set BIND_HOST only after you have put auth in place. MIT.
People who want one local assistant that can talk on WhatsApp, Slack, Telegram and other channels, using their own model keys.
Access and use
Installer is curl -fsSL https://openclaw.ai/install.sh | bash. Read the script before running it. Also npm install -g openclaw@latest --allow-scripts=openclaw. Needs Node 24.22.1 or later, or Node 26.12.1 or later. After npm, run openclaw onboard --install-daemon. Daily version check is on by default; anonymous feature statistics are opt-in; update.checkOnStart: false turns both off. Tools run on the host unless you configure sandboxing. Treat inbound messages as untrusted. MIT.
People who want to run DeepSeek's plugin-based agent stack on their own machine.
Access and use
npx @deepseek-ai/dsh web. Default web UI is http://127.0.0.1:3080. Marked developer preview with expected breaking changes. The SAFETY notice says it can execute model-generated code, load plugins, and has not been security audited. MIT.
People who want IFTTT-style automations on a server they control, with their data staying there.
Access and use
Quick try: docker run -it -p 3000:3000 ghcr.io/huginn/huginn then open http://localhost:3000 and log in as admin with the password printed once in the container log. Production needs MySQL or PostgreSQL and a web process plus a worker. MIT.
People who want a vision-capable agent that can drive a browser or desktop, using their own model keys.
Access and use
Needs Node 22 or later. npx @agent-tars/cli@latest, or npm install @agent-tars/cli@latest -g, then agent-tars --provider anthropic --model MODEL --apiKey KEY. The same repository ships UI-TARS Desktop. Treat computer and browser operators as full desktop control. Apache-2.0.
Teams adding an agent to an app they already ship, or connecting one agent to Slack, Teams and similar channels.
Access and use
From the app root: npx --yes copilotkit@latest onboard start. New project: npx copilotkit@latest create. Optional skills: npx copilotkit@latest skills install. CopilotKit Intelligence is a separate hosted or self-hosted product. MIT.
Teams that want a self-hosted or hosted agent workspace with MCP access from Claude Code and Codex.
Access and use
Cloud at sim.ai. Self-host: npx sim-setup. Read the script before running it. Then open http://localhost:3000. Needs Node 20 or later and Docker. Chat on a self-hosted install uses a Chat API key from sim.ai. Cloud MCP is https://mcp.sim.ai/mcp. Apache-2.0.
People who want a local copy of bookmarks, research pages or public web content they choose to keep.
Access and use
Installer is curl -fsSL https://get.archivebox.io | bash. Read the script before running it. Also docker compose from docker-compose.archivebox.io, or uv tool install archivebox. Docker -p 5797:5797 and archivebox server 0.0.0.0:5797 bind every interface. Pages are submitted to archive.org by default. MIT.
People who want coding agents and chat apps to call models that run on a Mac they already have.
Access and use
Needs macOS 15 or later and Apple Silicon. Download the .dmg from GitHub Releases, or brew tap jundot/omlx then brew install jundot/omlx/omlx. Default server is http://localhost:8000/v1. A non-loopback bind (--host 0.0.0.0) needs a main API key. Optional MCP via pip install mcp then omlx restart. Apache-2.0.
Teams who want agents in the same rooms as people, with their own keys and an audit trail.
Access and use
Download the desktop app from GitHub Releases, or git clone then just setup && just dev. Needs Docker. Default relay is ws://localhost:3000. Agents use buzz-cli with BUZZ_PRIVATE_KEY; they can open repos, send patches and run workflows on that relay. The Windows installer is unsigned. Apache-2.0.
Teams who want one API in front of several model providers, with a budget per identity.
Access and use
pip install experiential then exp. Local gateway is http://127.0.0.1:8000/v1. Hosted API is api.experientiallabs.ai. Anonymous PostHog product telemetry is on by default; exp config telemetry disable turns it off. exp capture (experimental, macOS) installs an extension and certificate to read your Codex and Claude Code desktop traffic, and sends those traces to your Experiential account. Apache-2.0.
Startups who want each person to have their own agent workspace, plus shared rooms on Slack and the web.
Access and use
npm exec --yes --package=@yc-software/qm@latest -- qm init . --org <slug> --target <fly-or-aws>. Needs Node 24.15 or later. The agent acts with that person's credentials. Default security posture is Auto, which blocks private-network access. Strict pauses every harness tool call. Dangerous drops those gates. Sessions need DATABASE_URL and SESSION_STORE=postgres to survive restart. A third-party hosted version runs at agent37.com/qm. MIT.
Companies who want an internal agent workspace with sandboxed apps and a security layer they can copy.
Access and use
pnpm run-local then visit http://localhost:8787. Needs pnpm. Early-access v2. Deploy to a Cloudflare account at os.cloudflare.app/deploy. Gadgets are sandboxed. Gatekeepers log actions and can queue side-effect approvals. Agents get only the resources you introduce. Apache-2.0.
People who want OpenClaw hosted on Cloudflare Workers, with pairing and Access in front.
Access and use
Experimental proof of concept. Needs a Workers Paid plan for Sandbox containers. npm install, then npx wrangler secret put ANTHROPIC_API_KEY (or the AI Gateway secrets) and MOLTBOT_GATEWAY_TOKEN, then npm run deploy. Open the Control UI with ?token=YOUR_GATEWAY_TOKEN. Cloudflare Access plus device pairing at /_admin/ are required before the Control UI works. Optional R2 keeps pairing and history across container restarts. SANDBOX_SLEEP_AFTER defaults to never; a timeout such as 10m sleeps the container to cut cost. DEV_MODE=true is local only and skips Access and pairing. Apache-2.0.
People who want agent code isolated from the host process, with explicit mounts and network rules.
Access and use
Library: npm install @rivet-dev/agentos-core, then AgentOs.create() in your backend. The sidecar is a separate Rust process. Guest code reaches host files, environment or the network only through mounts, host functions or network allow rules you add. Hosted: npm install @rivet-dev/agentos, set RIVET_NAMESPACE and RIVET_TOKEN from the Rivet dashboard, and keep the token on the backend. Browser clients need a JWT scoped to that VM. Preview URLs expire. Apache-2.0.
People who run Claude Code, Codex, OpenCode, Hermes or OpenClaw and want them in one workspace.
Access and use
git clone then make dev. Needs Go 1.25+, Node.js 20+, Docker and at least one supported agent CLI on PATH. make dev starts PostgreSQL in Docker on port 5432, a Go API on 8080, a daemon on 8081 and the UI on 3000. Register at http://localhost:3000. Signup stays on until you set ALLOW_SIGNUP=false. LLM_PROVIDER defaults to local, so agents are your installed CLIs. Workspace data lives in PostgreSQL. MIT.
People who want a local assistant on Node 22 with their own model keys, or the HybridAI Cloud workspace.
Access and use
npm install -g @hybridaione/hybridclaw, or curl -fsSL https://raw.githubusercontent.com/HybridAIOne/hybridclaw/main/scripts/install.sh | bash; read the script before running it. Needs Node.js 22.23.4 or later, below 23. hybridclaw onboarding then hybridclaw gateway. Default bind is 127.0.0.1:9090; web chat is http://127.0.0.1:9090/chat. Config is ~/.hybridclaw/config.json. Provider keys live in the encrypted store under ~/.hybridclaw/credentials.json. Docker is the default sandbox. Hosted path is hybridclaw.io. Set ops.webApiToken before exposing the gateway beyond loopback. Treat inbound channel messages as untrusted. MIT.
People who want one local console and API in front of several coding-agent harnesses, using their own model keys.
Access and use
docker run -d --name harnessrouter -p 127.0.0.1:3000:3000 -v harnessrouter:/data harnessrouter/harnessrouter. Needs Docker and about 4 GB of disk. Console is http://127.0.0.1:3000. Default Console login is harnessrouter / harnessrouter; change it before exposing the instance. First launch installs harness CLIs into the volume. Provider keys go under Bring your own key. Product API keys are created in the Console and belong in a backend, not a browser. Gateway :8080 and Runner :8081 stay on loopback inside the container. Hosted path is harnessrouter.ai. Apache-2.0.
People who want to inspect and steer a town of LLM agents from one browser on their machine.
Access and use
git clone https://github.com/XiaoLuoLYG/GOD.git then ./scripts/god.sh start (Windows: .\scripts\god.cmd start). The first run installs dependencies and opens a browser wizard. Paste an OpenAI-compatible API key, base URL and model name. The control room prints a URL such as http://127.0.0.1:5174/pixel-replay/god_town/1. .env stores local model and port settings. The model endpoint is the external service. Curated browser replays on the public site need no install or API key. Apache-2.0. Bundled AgentSociety and JiuwenClaw checkouts keep their own licences.
An open-source notes and knowledge workspace with three-pane editing, multi-device sync, and a native remote MCP so agents can create and edit notes in the same account.
People who want a self-hosted or Cloudflare-hosted notebook that coding agents can read and write through MCP.
Access and use
Cloudflare: fork the repo, create D1 edgeever and R2 edgeever-resources, then a Worker from main. Set the runtime Secret EDGE_EVER_AUTH_PASSWORD; upstream recommends 32 characters or longer. Username defaults to admin via EDGE_EVER_AUTH_USERNAME. Notes and attachments then live in that Cloudflare account. Docker: curl -fsSL https://edgeever.org/install.sh | bash (read the script before running it). The installer writes ~/edgeever, pulls ghcr.io/tianma-if/edgeever, generates an administrator password, and serves http://localhost:8787. Daily auto-update is crontab 04:17 unless EDGE_EVER_AUTO_UPDATE=false. Profile -> API / MCP issues a token with full CRUD on notes, notebooks, tags, attachments, history, templates and AI instructions. BYOK sends note text to the model endpoint you configure. The public demo at https://demo.edgeever.org resets daily; keep private notes off it. AGPL-3.0.
A multi-agent simulation engine that builds a digital world from seed documents and runs thousands of persona agents so you can rehearse what-if outcomes.
People who want to simulate public opinion, policy or story trajectories with an LLM swarm on their own machine.
Access and use
Needs Node.js 18+, Python 3.11 or 3.12, and uv. Copy .env.example to .env. Set LLM_API_KEY, LLM_BASE_URL and LLM_MODEL_NAME for any OpenAI-compatible endpoint, plus ZEP_API_KEY from https://app.getzep.com/. Your seed documents and agent memory are sent to Zep Cloud, and prompts go to the LLM endpoint you set. npm run setup:all then npm run dev. Frontend is http://localhost:3000; backend is http://localhost:5001. Docker: docker compose up -d with the same .env. Upstream says LLM use is high; try fewer than 40 rounds first. AGPL-3.0.
Teams who want issues, sprints and docs on a stack they can host.
Access and use
Cloud sign-up is https://app.plane.so. Community self-host: mkdir plane-selfhost && cd plane-selfhost, curl -fsSL -o setup.sh https://github.com/makeplane/plane/releases/latest/download/setup.sh, chmod +x setup.sh, then ./setup.sh action 1. Edit plane.env WEB_URL and CORS_ALLOWED_ORIGINS (defaults http://localhost on ports 80 and 443), then ./setup.sh action 2 to start. Commercial installer: curl -fsSL https://prime.plane.so/install/ | sh - (read the script before running it); that edition starts on a free plan with 12 seats. Needs 2 CPU cores and 4 GB RAM. Instance admins configure the host in God mode. Self-hosted Plane sends telemetry to Plane by default: the admin's name and email, the instance ID, anonymised usage counts and performance data including stack traces. Turn it off in God mode under General Settings, Telemetry. AGPL-3.0.
People who want a private search front-end they can host, or who will use a public instance they trust.
Access and use
Self-host with Docker or Podman: mkdir -p ./searxng/core-config && cd ./searxng, curl -fsSL -O https://raw.githubusercontent.com/searxng/searxng/master/container/docker-compose.yml -O https://raw.githubusercontent.com/searxng/searxng/master/container/.env.example, cp .env.example .env, then docker compose up -d. Compose publishes 0.0.0.0:8080. Edit core-config/settings.yml. Manual run: docker run --name searxng -d -p 8888:8080 -v ./config/:/etc/searxng/ -v ./data/:/var/cache/searxng/ docker.io/searxng/searxng:latest then http://localhost:8888. Each query is sent to the engines you enable. A public instance operator sees the queries you type. AGPL-3.0.
People who want a lighter YouTube player they can host, or who will pick a public instance they trust.
Access and use
Public instances are listed at https://instances.invidious.io; that operator holds your subscriptions, history and watch requests, and the instance fetches video from YouTube. Self-host: git clone https://github.com/iv-org/invidious.git && cd invidious. Generate two 16-character keys with pwgen 16 1, put them in docker-compose.yml as hmac_key and invidious_companion_key (companion SERVER_SECRET_KEY must match). Default compose binds 127.0.0.1:3000. Needs about 20 GB disk and 2 GB RAM; restart at least daily. Playback needs Invidious companion. AGPL-3.0.
People who want encrypted chat on iOS, Android or a terminal, with relays they can pick or host.
Access and use
Mobile: App Store, Google Play, F-Droid (https://app.simplex.chat) or the APK from GitHub Releases. Terminal: curl -o- https://raw.githubusercontent.com/simplex-chat/simplex-chat/stable/install.sh | bash (read the script before running it). Connect by scanning a QR code or opening an smp invitation link. Encrypted messages travel through the SMP relays you choose; the default public relays are operated by SimpleX Chat Ltd. AGPL-3.0.
People converting documents, images, e-books or video on a machine they control.
Access and use
Docker: map port 3000 and volume ./data:/app/data, set PUID to your host user id, then open http://localhost:3000 and create the first account. Anyone who can reach that URL can register that first account. Set JWT_SECRET. HTTP_ALLOWED=true only on a non-HTTPS local connection. Files land in ./data on the host; AUTO_DELETE_EVERY_N_HOURS defaults to 24. ALLOW_UNAUTHENTICATED and ACCOUNT_REGISTRATION stay off on a public host. AGPL-3.0.
Teams who want fill-and-sign workflows on a host they run, or on DocuSeal's cloud.
Access and use
Cloud sign-up is https://docuseal.com/sign_up; completed documents then live on DocuSeal LLC servers. Self-host: docker run --name docuseal -p 3000:3000 -v.:/data docuseal/docuseal then http://localhost:3000. Compose: curl https://raw.githubusercontent.com/docusealco/docuseal/master/docker-compose.yml > docker-compose.yml then sudo HOST=your-domain-name.com docker compose up. SQLite is the default; set DATABASE_URL for PostgreSQL or MySQL. Disk or S3/GCS/Azure for files. AGPL-3.0 with Section 7(b) additional terms: keep DocuSeal attribution in interactive UIs.
Freelancers and contractors who want invoices, time and bank feeds in one product they can host or run on Midday Cloud.
Access and use
Cloud is https://app.midday.ai, run by Midday Labs AB; your invoices, transactions and files are then processed in Sweden. Bank connections go through GoCardless (EU), Plaid (Canada and US) or Teller (US). Hosted analytics use OpenPanel. Local: git clone https://github.com/midday-ai/midday.git then bun i; you need a Supabase account, which holds the database, files and sign-in, and an Upstash account for its key-value store (https://docs.midday.ai/local-development). AGPL-3.0 for non-commercial use. For commercial use, Midday Labs AB asks for a commercial licence at engineer@midday.ai.
Teams adding sign-in, organisations and MCP-ready auth to a SaaS or agent product.
Access and use
Cloud is https://cloud.logto.io; identity data then lives on Logto Cloud. Local Docker Compose: curl -fsSL https://raw.githubusercontent.com/logto-io/logto/HEAD/docker-compose.yml | docker compose -p logto -f - up. That compose bundles Postgres and drops data if you run it again. npm init @logto@latest needs Node.js 26+ and PostgreSQL 14+. Core is http://localhost:3001; Admin Console is http://localhost:3002. The welcome page creates one admin account with username and password. Recommended host: 2 vCPU, 8 GiB RAM, 256 GiB disk. MPL-2.0.
Small teams who want kanban, gantt, goals and docs on a host they run.
Access and use
Cloud is https://leantime.io; project data then lives on Leantime's hosted service. Self-host needs PHP 8.2+ and MySQL 8.0+ or MariaDB 10.6+. Docker: docker run -d --restart unless-stopped -p 8080:8080 with LEAN_DB_HOST, LEAN_DB_USER, LEAN_DB_PASSWORD, LEAN_DB_DATABASE and LEAN_EMAIL_RETURN, then open http://localhost:8080/install. Files can go to local disk or S3. Plugins need the plugins volume mounted and writable by www-data. Self-hosted Leantime sends a daily anonymous usage summary (counts and versions) to telemetry.leantime.io by default; set LEAN_ALLOW_TELEMETRY=false to stop it. AGPL-3.0.
People who want chat and calls on Fluxer's hosted service or on an instance they run.
Access and use
Hosted: https://web.fluxer.app or the desktop and mobile apps; messages then live on Fluxer's servers. Self-host: curl -fsSL https://fluxer.dev/install.sh | sh (read the script before running it). Windows: irm https://fluxer.dev/install.ps1 | iex (read the script before running it). Needs 2 CPU cores, 4 GB RAM, 20 GB disk, a domain A record, TCP 80, 443 and 7881, and UDP 443 and 7882. rexctl creates one admin account in the terminal or with a setup token. AGPL-3.0-or-later for the source; artwork is CC BY-SA 4.0.
People who want a self-hosted manga library that talks to the same sources as Mihon.
Access and use
OS bundles (Windows, macOS, Linux) ship JRE, Electron and a launcher from GitHub Releases. Docker: ghcr.io/suwayomi/tachidesk, default http://localhost:4567. Jar needs JRE 21+: java -jar Suwayomi-Server-vX.Y.Z-rxxxx.jar. Library, backups and downloads live on the host. Installed extensions fetch chapters from the manga sites you enable; optional FlareSolverr and trackers (MyAnimeList, AniList, MangaUpdates) send data to those services. MPL-2.0.
An Apache-licensed FastAPI and Next.js executive persona backed by nine specialist agents, with local RAG, episodic memory and optional chat and mail connectors.
Teams who want a self-hosted executive assistant that drafts in one voice and leaves messages on approval cards unless Handle it for me is on.
Access and use
git clone https://github.com/SenteLabsAI/OpenExecutive.git && make install, copy .env.example to .env, set ANTHROPIC_API_KEY, then make dev. UI is http://localhost:3000; API is http://localhost:8000. Needs Python 3.11+ and Node 22.6+. First boot downloads an embedding model of about 90 MB. Local make dev with blank AUTH_GOOGLE_ID accepts connections from this computer only. Company documents and memory are stored on the host (ChromaDB and SQLite); prompts, retrieved document chunks and memory go to the Anthropic API. Slack, mail, Telegram, Google Chat and similar connectors send messages through those services when you set their tokens. Apache-2.0.
People who want mailboxes on their own domain, on a Cloudflare account or a Docker host they run.
Access and use
Cloudflare: Deploy to Cloudflare with Worker name mailflare, then open /setup to create the first admin account. CF_TOKEN is a scoped API token that can edit DNS and Email Routing on the zones you connect, and Email Sending if you send through Cloudflare. Docker: git clone https://github.com/hieunc229/mailflare.git && cd mailflare && cp .env.docker.example .env.docker && docker compose up -d --build. Setup is http://localhost:3000/setup. On Cloudflare, mailbox records live in D1 and files live in R2 (or B2 or S3) in that account. Docker stores SQLite, mail and backups in the mailflare-data volume. Send and receive through Cloudflare Email, Resend, Amazon SES or SMTP. The assistant sends mailbox content to the AI provider you set under Admin, Agent. MCP is Streamable HTTP at /mcp on your origin with a Bearer key; request_send still needs confirmation in Mailflare. Optional Pro and Team licence keys are checked with Paymug. AGPL-3.0.
Researchers who want a local workbench that keeps project files on the machine and sends model calls to a provider they choose.
Access and use
Desktop app for macOS, Windows and Linux from https://openscience.sh/download. CLI: npm install -g @synsci/openscience then openscience. Or npx synsci. Or curl -fsSL https://openscience.sh/install | bash (read the script before running it). First launch of the GUI asks you to sign in to a Synthetic Sciences account. Project files stay in the connected folder. Prompts go to the model provider you connect, a local endpoint, or Ace Wallet pay as you go. Ace auto reload is a separate opt-in that charges a saved card. Search, database connectors and remote compute send queries and selected files to those services. While signed in, session traces are shared by default, including sessions that use your own keys or a local model; turn that off under Customize, Permissions, Data and privacy. Apache-2.0.
Friends who want a private watch-together room they can run themselves.
Access and use
Hosted demo is https://demo.piik.tv, P2P only; video and audio travel between participants. App: download Piik App from https://piik.tv, then choose Public invite or Local room. Self-host: extract the Server archive from GitHub Releases and run ./piik-server (Linux) or ./piik-server.exe (Windows). Local UI is http://localhost:8787. Room records stay in SQLite. Docker: download compose.yaml and .env.example from the repo, set your domain, then docker compose up -d. Default is P2P with STUN on UDP 3478; optional SFU on UDP 7882 forwards video and audio through the server. Remote control, once you approve it, lets viewers use the host keyboard, mouse or a supported gamepad. MIT.
Developers who want an agent to diagram a change beside the local repository.
Access and use
macOS, Windows and Linux installers from https://dev.fast/install. On macOS: brew install --cask devdotfast/tap/whiteboard. Connect Claude Code, Codex or another coding agent from the welcome screen. Whiteboards stay in ~/.dev/review-api.db. Anonymous telemetry is on by default and goes to PostHog as feature counts, durations and versions under a random installation ID; it omits source code, diffs, paths, prompts and model output. Turn it off in Settings under Share anonymous usage data, or set DO_NOT_TRACK=1. After a crash, the next launch uploads the dump to bug.dev.fast, and it can contain process memory including open source text; with telemetry off, dumps are deleted without upload. whiteboard share uploads a copy of that whiteboard, traces and referenced source to the /dev/fast share service. Trace capture is off until you allow a repository. MIT.
Teams who want Linear-style tracking on their own Postgres, with agents as assignees.
Access and use
git clone https://github.com/croffasia/itsaplan.git && bun install && bun run setup, then answer Try it. UI is http://localhost:3001; API is http://localhost:3000. Production: copy .env.example to .env, set the secrets and public URLs, then docker compose up -d behind TLS. The first account is instance admin. Issues live in Postgres; files live in RustFS. Agent prompts go to the model or CLI you configure. A daily telemetry snapshot goes to https://telemetry.itsaplan.dev by default (counts as ranges, plus which features and integrations are on); set TELEMETRY_DISABLED=1 or DO_NOT_TRACK=1 to stop it. MCP is on the API origin. AGPL-3.0; the runner package is Apache-2.0.
People who want a desktop harness that can read the workspace and run shell commands after you approve the tool call.
Access and use
npx openwork, or npm install -g openwork then openwork. Needs Node.js 18 or later. From source: git clone https://github.com/langchain-ai/openwork.git && npm install && npm run dev. Set Anthropic, OpenAI or Google keys in the settings panel. Agents can read the filesystem and run shell commands; review each tool call and only open workspaces you trust. Prompts go to the model provider you choose. MIT.
People who want a small private mesh across home, work and a cloud box, without a SaaS overlay.
Access and use
Download the CLI or desktop app from GitHub Releases. Hub: ztm start hub --listen 0.0.0.0:8888 --names YOUR_PUBLIC_IP:8888 --permit root.json. Agent: ztm start agent, then ztm join MESH --as NAME --permit root.json. GUI is http://localhost:7777 after the agent starts. Hub state is ~/.ztm/ztm-hub.db; agent state is ~/.ztm/ztm.db. Traffic is TLS. The permit file is the join key. Apache-2.0.
Developers who want an AI pass over a GitHub repo for secrets, injection and quality.
Access and use
Self-host: git clone https://github.com/danish296/codevibes.git, npm install, then cd codevibes-backend && npm install, copy .env.example to .env, set DEEPSEEK_API_KEY, JWT_SECRET and a 32-character ENCRYPTION_KEY. Backend: npm run dev at http://localhost:3001. Frontend: npm run dev at http://localhost:8080. History is SQLite at codevibes-backend/data/codevibes.db. Repo files go to DeepSeek. Private repos need a GitHub token or GitHub OAuth (GITHUB_CLIENT_ID and GITHUB_CLIENT_SECRET). MIT.
Teams shipping Android, iOS, HarmonyOS or Electron builds who want agents and CI to publish through one API.
Access and use
Hosted dashboard https://app.hands.build and API https://hands.build, run by Oranix on Cloudflare Workers, D1 and R2. Login with Raft. CLI: npm install -g @botiverse/hands-cli, then hands login (browser JWT saved at $XDG_CONFIG_HOME/quiver/auth.json) or CI HANDS_API=https://hands.build with HANDS_AUTH_TOKEN or HANDS_BEARER_TOKEN. Builds, crash reports, feedback and installable artifacts go to that Cloudflare account. Self-host: git clone https://github.com/botiverse/hands.git, pnpm install, pnpm --filter @botiverse/hands-worker dev at http://127.0.0.1:8787, pnpm --filter @botiverse/hands-admin dev at http://localhost:5173 (proxies /api to 8787). Docker is needed for the APK parser container. Set RAFT_CLIENT_ID in worker config and wrangler secret put RAFT_CLIENT_SECRET. The production Worker config turns Cloudflare observability traces on (5 percent head sample, persist). MIT.
An Apache-2.0 Cloudflare-native multi-tenant runtime for autonomous agents, with Durable Object SQLite per tenant, V8 isolates and on-demand Run9 containers.
Teams who want agents to call real APIs and sandboxes while keeping credentials out of the model context.
Access and use
Hosted UI https://antiproton.ai and admin https://admin.antiproton.ai, run by botiverse on Cloudflare Workers, Durable Objects, D1 and R2. Custom domain antiproton.botiverse.dev sits behind Cloudflare Access. GitHub sign-in; UI_ALLOW_ANONYMOUS is 0; GITHUB_OPEN_SIGNUP is 1 so a GitHub account registers on first sign-in. Self-host: git clone https://github.com/botiverse/antiproton.git, npm install, Node.js 24+. cd cf && npx wrangler dev. npx wrangler secret put ANTHROPIC_API_KEY, or OPENROUTER_API_KEY or DEEPSEEK_API_KEY. Production deploy: cf/scripts/deploy-pinned.sh from origin/master. Default HARNESS_MODEL is deepseek-flash. USER_MODELS are deepseek-flash and gpt-5.6-luna through Cloudflare AI Gateway once AI_GATEWAY_TOKEN is set. Credentials stay behind the tool gateway and never enter model context prompts, tool outputs or execution logs. Three tiers: Durable Object plus SQLite per tenant and agent; V8 Dynamic Worker with globalOutbound null; Run9 containers with a 30 minute idle lease and a 5 minute warning. TRACE_INGEST_URL and TRACE_INGEST_TOKEN are unset so trace forwarding is off. Cloudflare observability is enabled. Raft provision APIs can create hosted agents. Daily spend limits apply on the operator model accounts. Docs: docs/agents-api.md, docs/agent-surface.md, docs/plugins.md, docs/runbook.md. Default branch is master. Apache-2.0.
Teams evaluating a Linux mail and newsletter stack they operate themselves.
Access and use
Linux amd64 or arm64 with Docker and Docker Compose. The stack includes PostgreSQL, Redis, Postfix, Dovecot, Rspamd and Roundcube; the management container has a Docker socket mount and network capabilities. Requires domain and DNS configuration. Self-hosting, maintenance and any external SMTP relay or AI provider have their own costs. SMTP relay is configurable per sending domain. AI features can call configured model providers, and an image-upload path uses the BillionMail CDN. AGPL-3.0. CVE-2026-111367 affects releases through 4.9; the reviewed dev source still has deterministic fallback generation for subscription signing secrets. A fixed deployment image has not been verified. Separate contact groups or domains do not by themselves keep one client's data apart from another's. We list it for its documentation and source only.
Teams who want to connect existing services in workflows they can host or run in n8n Cloud.
Access and use
Self-host with Docker using docker.n8n.io/n8nio/n8n; the editor runs on port 5678 and the n8n_data volume persists data under /home/node/.n8n. Workflows use the credentials and data access of the connected services; model and service charges still apply. n8n Cloud has its own account and plan. Source-available under the Sustainable Use License for internal business, personal or non-commercial use, with separate enterprise terms for designated code. Check the licence before embedding it in a customer product or offering hosted access.
Teams building AI applications with a visual workflow and document knowledge base.
Access and use
Self-host with Docker Compose 2.24 or later. The README specifies at least 2 CPU cores and 4 GiB RAM. From the docker directory, configure .env and start Compose; initialise through /install. Dify Cloud needs an account and has its own plan. Documents and conversation data are processed by the deployment and whichever models, tools or observability services you configure. Model-provider charges still apply. The modified Apache-2.0 licence restricts operating a multi-tenant environment without written authorisation or a commercial licence, and removing frontend logos or copyright notices. A workspace is a tenant under these terms.
Developers who want to design and serve AI workflows visually.
Access and use
The README supports a desktop app for Windows and macOS, a Python package, or Docker. Python 3.10 to 3.14; install with uv pip install langflow -U, then uv run langflow run. Local UI is http://127.0.0.1:7860. Set LANGFLOW_SUPERUSER_PASSWORD before startup. Official release packages and Docker images require authentication by default; development branches retain automatic login. Flow components can use connected files, models, databases and service credentials, with provider charges where applicable. MIT.
Python developers building agent workflows in their own application.
Access and use
Requires Python 3.10 to 3.13 and uv. Install the CLI with uv tool install crewai. Models and tool integrations need their own configuration, credentials and any provider fees. The framework can run tools with the access the host gives them. Its README describes usage telemetry, disabled with OTEL_SDK_DISABLED=true. Enabling share_crew additionally shares detailed agent and task execution data. MIT.
Teams evaluating searchable document collections and retrieval workflows they operate.
Access and use
Self-host through Docker and Docker Compose; the reviewed README quickstart selects release candidate v1.0.0-rc2. Recommended starting host: 4 CPU cores, 16 GB RAM and 50 GB disk, with more resources for local models or higher load. Docker 24 or later and Compose 2.26.1 or later. Configure an LLM, embedding model and reranker, including endpoint and API key as needed. Uploaded documents are processed and stored by the deployment, and configured model services may receive document or query content. Provider and hosting costs apply. Apache-2.0.
Teams who want a web interface for model access and document chat on infrastructure they operate.
Access and use
Use the official Docker image ghcr.io/open-webui/open-webui, or install the Python package with Python 3.12. The Docker quickstart serves the UI on localhost:3000 and persists the database and uploads in the open-webui volume at /app/backend/data. Local or remote model endpoints determine where prompts and documents go and whether provider charges apply. Offline use needs local models and assets prepared first. The current Open WebUI License includes branding-preservation requirements, with exceptions including deployments of at most 50 end users in a rolling 30-day period or written permission. Earlier components retain their original terms. Check the licence before rebranding.
A visual builder for internal tools, admin panels and dashboards that run on existing databases, APIs and SaaS systems, with an MCP server that lets coding agents build and edit ToolJet apps.
Teams building internal apps over their own data who want people and coding agents working in the same app builder.
Access and use
Use ToolJet Cloud with an account, or self-host on Docker, Kubernetes, AWS, GCP, Azure and the other documented targets. The README's quick-try image is tooljet/try:ee-lts-latest on port 80. Each data source needs its own credentials, which apps use to run queries. The MCP server is marked beta and runs agent operations on your own model subscription. AI app generation, workflows and the agent builder are listed under the Enterprise edition, whose code sits in non-public submodules (frontend/ee, server/ee). The repository is AGPL-3.0, which requires offering source for modified versions run as a network service.
A self-hostable platform for tracing, evaluating, simulating and guarding LLM and agent applications, with an OpenAI-compatible gateway and prompt optimisation.
Teams running agents in production who want traces, evals and guardrails in one system they can host themselves.
Access and use
Future AGI Cloud needs an account. Self-host clones the repository and runs ./bin/install, which writes secrets to .env and pulls about 800 MB of images. The default Standalone setup needs Docker Compose v2.24 or newer with 2 vCPUs and 4 GB of memory; a Helm chart is also published. Apps send traces with an API key and secret key; add LLM provider keys when you need them. Self-hosted installs send telemetry by default, including the email addresses of owner, admin and staff accounts. Opting out with --no-telemetry or FUTURE_AGI_TELEMETRY_DISABLED=true still sends one registration without email addresses; block api.futureagi.com to send nothing. The core is Apache-2.0, but code in ee/ directories is under the Future AGI Enterprise License and needs a paid subscription for production use.
An alpha personal agent app for iOS, Android and web, built on CopilotKit and AG-UI, with a persistent Chromium browser, an optional Linux workspace, task plans with approvals, and Gmail and Calendar adapters.
Developers who want a self-hosted personal agent template to fork and extend.
Access and use
Needs Node 24, pnpm 11.19.0 and a CopilotKit Intelligence project key in every mode; Intelligence is a separate service outside the MIT licence. A model needs an OpenAI, Anthropic or Google key. Gmail and Calendar need your own Google OAuth client plus live mode with an access key and token encryption key, and each send or calendar change waits for review. Web search is on by default and sends model-generated queries and context to Parallel's Search MCP; set WEB_SEARCH_ENABLED=false to turn it off. The CopilotKit SDK sends runtime telemetry to PostHog. The optional Linux computer needs Docker. The Render blueprint runs the API and browser on Standard instances. One owner behind a shared access key; not multi-tenant. MIT.
Microsoft's reinforcement learning framework for training the model behind an agent while the agent keeps running in its real harness, through a gateway that captures interactions as training data.
Machine learning teams who want to train an open-weight model on their own agent's tasks.
Access and use
Needs a CUDA GPU machine. The README example runs uv sync and a verl setup script for CUDA 13.0, and the trainer uses verl and vLLM. Agents run locally or as Kubernetes Jobs. The Calc-X example needs one GPU. Releases before v1.0 live on the v0.x branch. MIT.
Developers who test AWS SDK, CLI, Terraform or CDK code on their machine or in CI without an AWS account.
Access and use
docker compose up with the floci/floci image on port 4566, or floci start with the Floci CLI; a standalone binary is on GitHub Releases. Needs no AWS account or auth token, and any non-empty credentials work unless you turn on service auth checks. Lambda, RDS, ECS, EC2, EKS and other services run real containers and need the Docker socket mounted, which gives Floci control of Docker on the host. The web console at localhost:4566/_floci/ui starts a sidecar on port 4500, published on every interface unless you set FLOCI_SERVICES_UI_BIND_ADDRESS. Data stays in memory by default. MIT.
An open-source AI search visibility platform that tracks how a brand appears in ChatGPT, Gemini, Claude, Google AI Overviews and other engines, and uses agents to draft content fixes, with an MCP server and Claude skills.
Agencies, brands, SaaS companies and developers who want to track and improve how AI answer engines present them, including from Claude, Cursor or another MCP client.
Access and use
Self-host: clone the repo, copy the web and server .env example files, run the migration SQL in a Supabase project (the free tier works), then install the web app with Yarn and the server with npm, or run docker compose up --build. Needs Node.js 20 or later, a key for at least one AI provider (OpenAI, Google Gemini or Anthropic) and a Cloro key for collecting answers from AI platforms; Site Audit also needs Scrape.do, and DataForSEO, Tavily, OpenRouter and Composio are optional. Tracked prompts go to the AI providers you configure and to Cloro on a nightly cycle, and each run uses credits at those services. Connect Claude Code, Cursor or another MCP client to /api/mcp with an ans_ API key sent as a bearer header; app.ansvisor.com/api/mcp is the hosted endpoint on Ansvisor's servers. Docker Compose publishes the web app on port 3000 and the API on port 80, and the first user to sign up becomes the admin, so register first and put a firewall in front. Send to CMS sends WordPress pages as drafts for you to publish. Agents that publish on their own are off by default and need server flags, an Admin switch and per-agent permission. Telemetry is off on self-host unless you set a PostHog key. A managed Ansvisor Cloud is also offered. MIT.