taOS Self-Hosted Agent OS
Run, coordinate, and retain agents, models, memory, and files on hardware you control.
Per-dimension scores and reasoning
The evidence describes local-first operation, opt-in cloud use, per-category memory capture controls, explicit configuration before agents can read user memory, and worker pairing through a one-time PIN, human approval, and signing keys. Model manifests are said to use pinned checksums, and SECURITY.md provides a private reporting path. Deductions apply because the recommended installation pipes a script from the moving master branch into sudo, while Linux workers may create a privileged LXC, forwarding rules, and system services. The supplied material lacks a complete permission inventory, secret-lifecycle design, cloud-provider data-flow map, and system-wide revoke or uninstall procedure. Browser-data deletion has API tests, but universal confirmation for dangerous actions is not demonstrated. Several upstream sources are attributed, although publisher identity remains unknown.
The README, package configuration, and frontend tests collectively show defined platform behavior, dependency constraints, network-error fallbacks, and state handling. Installation is described as idempotent, and several dependency caps document concrete regression causes. Deductions apply because nothing was executed, many availability claims remain README assertions, and client helpers often collapse network failures into empty arrays, null, or false without preserving actionable detail. The stated total of 43 bundled apps is not immediately reconciled with the subsequent category counts, and the README acknowledges that many manifests have not been exercised on real hardware.
The material thoroughly identifies personal self-hosting, low-power devices, desktop and mobile use, heterogeneous clusters, multiple agent frameworks, and local or cloud models, with concrete operating-system and hardware paths. Deductions apply to capability boundaries and trigger precision: beta status, unverified manifests, explicit memory access, pairing, and hardware filtering are documented, but there is no complete agent-tool authorization model, per-skill triggering contract, or account of adapter failure boundaries across all claimed frameworks.
The README is well organized and supplies concrete quick-start, distribution dependency, development, worker installation, pairing, and verified-platform notes. Beta limitations, untested catalog entries, branch support policy, and a complete AGPL-3.0 license are explicit. Deductions apply because taOS and tinyagentos names and commands coexist, creating potential migration and operational ambiguity; the evidence lacks a dedicated FAQ, comprehensive uninstall, upgrade, or rollback instructions, and a formal changelog. Maintenance channels and response expectations are stated, but responsibility appears concentrated in one maintainer of unverified identity, with no clear governance or succession path.
The described web desktop, messaging, agents, files, models, monitoring, and clustering workflows appear directly usable, while shared memory across frameworks and aggregation of consumer hardware offer meaningful incremental value. Deductions apply because the outcomes are supported mainly by project descriptions and narrow frontend tests rather than static evidence for the main end-to-end workflows. Hardware price and offline benefits are mentioned, but the total cost and complexity of installation, privileged containers, model storage, maintenance, and optional cloud services are not systematically assessed.
The sources give concrete paths, interfaces, environment variables, dependency versions, issue references, tests, and design-document pointers. Some claims are corroborated across the README, pyproject, SECURITY policy, and tests, and the benchmark discussion explicitly distinguishes Judge accuracy from Recall@5. Deductions apply because major performance, scale, checksum-verification, and data-sovereignty claims cannot be traced through the supplied code excerpts, external benchmark and hardware reports are not independently corroborated, and several marketing conclusions extend beyond the available static evidence.
- Do not pipe the remote installer from the moving master branch into sudo before reviewing and pinning the script revision.
- The Linux worker setup creates a privileged LXC, btrfs pool, nftables forwarding rule, and root service; validate permissions, ports, and removal procedures on an isolated host first.
- The project is beta, and the README explicitly says many app, framework, and model manifests have not been tested on real hardware; validate each required component before deployment.
- Agent browsers can expose session cookies, while the platform stores API keys and long-term memory. Restrict agent skills, network reach, and secret access, and establish key-rotation and data-deletion procedures.
- This assessment is a static review of only the supplied files; no installation, tests, benchmarks, or security scans were executed.
What does this agent do, and when should you use it?
taOS is a self-hosted agent platform built around a FastAPI controller, React web desktop, taOSmd memory service, LiteLLM proxy, and distributed workers. It can run on one Raspberry Pi, Orange Pi, Mac, or PC, or combine heterogeneous devices into a hardware-aware inference cluster. The platform owns agent memory, files, secrets, messaging channels, model access, and container lifecycle, while frameworks such as SmolAgents, OpenClaw, the OpenAI Agents SDK, and Langroid act as replaceable execution engines. Users operate it through a browser desktop, mobile PWA, WebSocket chat, HTTP APIs, command-line utilities, and MCP plugins, producing conversations, files, searchable knowledge, generated media, games, and LoRA adapters. Local models can run fully offline, while OpenAI, Anthropic, DeepSeek, and compatible APIs are optional; the repository is explicitly labeled Beta, with some catalog manifests and management flows still incomplete or unverified.
The FastAPI controller manages agents, workers, applications, models, channels, secrets, scheduled tasks, and backups through the web desktop and APIs. Its deployment wizard follows a framework → model → configuration → deployment flow, placing each agent in an LXC or Docker environment with separate taOSmd memory, file storage, and network identity. taOSmd ingests messages, notes, file activity, searches, and documents, then applies fact extraction, a temporal knowledge graph, FTS5 keyword search, vector search, RRF merging, cross-encoder reranking, and token-budgeted context assembly; documented endpoints include POST /ingest/batch, GET /search?mode=bm25, and POST /vsearch. Channel Hub receives Telegram, Discord, Slack, Matrix, email, Web Chat, and webhook traffic, normalizes it, and dispatches it through framework adapters. LiteLLM exposes an internal OpenAI-compatible gateway with per-agent virtual keys, budgets, and rate limits, routing requests across available local and optional cloud backends. Workers report hardware, load, and loaded models so the scheduler can place inference, embedding, generation, and training jobs; the Store also installs frameworks, streamed desktop applications, models, and 47 MCP plugins.
- A privacy-conscious individual or small team that wants conversations, documents, and long-term memory to remain on a LAN while retaining optional access to cloud models.
- A home-lab owner with spare Orange Pi, Raspberry Pi, Mac mini, and gaming-PC hardware who wants to combine those devices into one heterogeneous inference cluster.
- A development team evaluating several agent frameworks that wants memory, channels, files, and secrets to survive a change of execution engine.
- An organization routing several agents through Telegram, Discord, Slack, email, and web chat while keeping bot credentials and delivery logic under central control.
- A creator who wants to generate images, video, speech, games, or LoRA adapters locally and already has an appropriate GPU, NPU, or CPU worker.
- A software team using Claude Code, aider, opencode, or another CLI coding agent that wants approval-based onboarding, scoped identities, Kanban tasks, and A2A coordination.
What are this agent's strengths and limitations?
- Memory, files, messaging connections, secrets, and model access sit outside the execution framework, allowing supported frameworks to be changed without discarding platform-managed state.
- One controller can coordinate Apple Silicon, NVIDIA, AMD, Rockchip NPU, Raspberry Pi, Android, and CPU-only workers and route work from their live capability reports.
- taOSmd combines a temporal knowledge graph, hybrid keyword/vector retrieval, a zero-loss archive, reranking, and optional source-support verification instead of storing only chat transcripts.
- The same deployment can remain fully offline or use multiple cloud providers through LiteLLM, with per-agent virtual keys, budgets, and rate limits.
- The browser desktop, mobile PWA, channel integrations, MCP catalog, and application store put deployment, conversation, monitoring, and generation tools in one interface.
- The project is in Beta: agent management, worker connectivity, and model-routing flows are still being refined, and many application, framework, and model manifests have not been tested on real hardware.
- A full controller brings Python, Node.js 22, PostgreSQL, SQLCipher, several services, and multiple ports, creating substantially more operational overhead than a single chat or inference server.
- The recommended Linux worker setup creates a privileged Incus LXC, a btrfs storage pool, and an nftables DNAT rule, which require host-level privileges and careful security and storage review.
- Running taOS inside an unprivileged Proxmox-style LXC blocks nested agent containers; the documented workaround requires a privileged container with nesting enabled, changing the isolation boundary.
- The reported 97.0% LongMemEval-S result is an end-to-end Judge measurement and cannot be compared directly with the Recall@5 figures published for MemPalace or agentmemory.
How do you install or deploy this agent?
The controller documents support for Debian, Ubuntu, Fedora, Arch, Alpine, and macOS. Run:
curl -fsSL https://raw.githubusercontent.com/jaylfc/taOS/master/scripts/install-server.sh | sudo shThe installer provisions Node.js 22, a Python environment, system dependencies, tinyagentos.service on port 6969, and qmd.service on port 7832. Running it without sudo installs a user-mode systemd unit. When installation finishes, open http://your-host:6969 or, where mDNS works, http://taos.local:6969. For a development checkout, use:
pip install -e .
python -m uvicorn tinyagentos.app:create_app --factory --host 0.0.0.0 --port 6969No cloud credential is required for local models. To use OpenAI, Anthropic, DeepSeek, or another compatible service, add and test its credentials in the Providers app. An optional owner password can protect the dashboard. Before production adoption, review the documented PostgreSQL installation, system services, ports, and persistent-data paths; the project identifies itself as Beta.
How do you use this agent?
After opening http://your-host:6969, install the desired framework and models from Store, then use the Agents app's five-step wizard to select a framework, choose a model, configure it, and deploy it. Chat through Messages or the standalone /chat-pwa, and use Memory, Files, Channels, Secrets, and Tasks to manage each agent's knowledge, workspace, connections, credentials, and schedules. Add a Linux or macOS worker with:
curl -fsSL https://raw.githubusercontent.com/jaylfc/taOS/master/scripts/install-worker.sh | sudo bash -s -- http://your-server:6969The worker prints a pairing code on first launch; approve it in Cluster. To avoid LAN announcements, set TAOS_PAIR_MANUAL=1 and enter the worker address and one-time PIN within its 15-minute validity window. An external coding agent begins by calling POST /api/agents/auth-requests with its identity claim and requested scopes; after user approval, it uses the issued Ed25519 registry JWT for scoped board, A2A, Canvas, decision, or project-file operations. Upgrade an installed controller by running bin/update.sh from its checkout.
How does this agent compare with similar options?
Unlike systems that bind memory and integrations to one execution framework, taOS keeps memory, channels, files, secrets, and model access at the platform layer and describes switching among engines such as SmolAgents, LangChain, and OpenClaw. Its user-memory experience is positioned as a self-hosted analogue to Pieces App. The reported 97.0% taOSmd LongMemEval-S score measures retrieval, answer generation, and LLM judging, whereas MemPalace's 96.6% and agentmemory's 95.2% are Recall@5 retrieval figures; the source explicitly states that these are not apples-to-apples metrics.