OtoDock
The self-hosted agentic company OS: run a team of collaborative AI employees on your own Claude and ChatGPT subscriptions, all under one dashboard.
Evidence shows a detailed sandbox model (kernel namespaces, network isolation, role-based mounts), encrypted per-session brokered credentials, API-key redaction asserts, and a lint guard banning transcript logging at INFO+ (sensitive_data_handling earns full marks). Deductions: least_privilege and external_effects rest on README claims — the sandbox and tool-authorization source is not in evidence, so 2; user_confirmation is unaddressed anywhere (unattended runs, phone calls, full access on paired machines), 1; rollback is limited to a mention of backup/restore scripts with no agent-run recovery story, 1.
Evidence shows pinned CI dependencies (Postgres 16.14, Python 3.13, Node 24), ruff/pytest/vitest gates, and tests explicitly covering error surfacing (pop_fatal_error once-only semantics) and rate-limited logging — failure_messages and dependency_availability are fully earned; no concrete weakness found to deduct.
Evidence shows a clear audience (self-hosted teams, BYO subscriptions), four sharing modes, three roles, per-agent engine switching — audience full marks. Deductions: capability boundaries are asserted only via per-agent tool assignment with no boundary documentation, 2; triggers (schedule/webhook/once) are listed without precision detail, 2; environment fit is narrow — CI comments say amd64-only, quick start requires Linux+Docker, 2.
Evidence shows a complete FSL-1.1-Apache-2.0 LICENSE with Apache 2.0 future grant and named copyright holder — license full marks (the registry's NOASSERTION is a registry-side mismatch, not a repository defect); README structure, quick start, and architecture map are excellent, so information_architecture and install_notes are full. Deductions: CHANGELOG, VERSIONS.md, CONTRIBUTING are referenced but absent from evidence, so versioning_changelog and naming_stegrity get 2; known limitations appear mainly as SECURITY.md scope exclusions, 2; examples/FAQ live on the external docs site, 2; maintenance is a single author with a clear security process (72h acknowledgment) but bus-factor risk, 2.
Evidence shows genuine differentiation: BYO Claude/Codex subscriptions, self-hosting, per-user cost budgets — marginal_value full marks. Deductions: output_usability is inferable only from screenshots and prose (streaming replies, in-chat file editing) with no static proof of output quality, 2; cost_benefit direction is sound but unquantified in evidence, 2.
Evidence shows test comments carefully separating live-reproduction facts (dated) from design inference — fact_inference_separation full marks. Deductions: core claims (sandbox escape protection, SSO, multi-tenant isolation) point to external docs not present in evidence, claim_traceability 2; CI and tests corroborate engineering practice and the audio layer, but not the agent sandbox or security model itself, cross_source_corroboration 2.
- Sandbox, network-isolation, and multi-tenant isolation claims are README assertions only; this static review could not verify the implementation — validate sandbox-escape protection independently before sensitive deployments.
- Agents run unattended, place phone calls, and get full access on paired machines, yet no user-confirmation mechanism appears in evidence; restrict triggers and tool grants yourself before production use.
- FSL-1.1 is not an OSI-approved open-source license and restricts competing commercial use; the registry's NOASSERTION conflicts with the actual LICENSE — reconcile before procurement.
- Release images are amd64-only and the maintainer is a single author (bus-factor risk); consult the CHANGELOG before upgrading (its contents were not in evidence).
What does this agent do, and when should you use it?
OtoDock is a self-hosted AI agent platform branded as 'the agentic company OS,' built on Claude Code and Codex and consuming your own Anthropic and OpenAI subscriptions, with API keys and local models also supported. Every agent consists of six editable parts — Persona, Memory, Workspace, Knowledge, Skills and Tools — and a sharing mode that determines how workspaces are divided among members. The platform organizes agents into departments, supports delegation between agents, meetings, scheduled runs, webhook triggers, and phone calls over Twilio or Asterisk/FreePBX. Technically, agents run as Claude Code / Codex processes inside per-session kernel sandboxes, talk to their tools over MCP, stream every step to a React dashboard through a FastAPI core (proxy/) that handles sessions, security, scheduling and the WebSocket hub, with PostgreSQL holding platform state. It is multi-tenant by design, ships with SSO, two-factor auth and per-user cost budgets, and is licensed FSL-1.1-Apache-2.0 — free to self-host for up to 5 users.
The installer (scripts/install.sh) checks Docker on a Linux server, generates a .env with a database password, downloads the release-pinned docker-compose.yml plus the phone-service overlay, and starts the stack. On first visit to http://localhost:8400 a setup wizard creates the owner account and auto-installs a Personal Assistant; users then connect their own Claude or ChatGPT subscription under Setup → AI Engines. Each agent runs as an isolated Claude Code / Codex process in its own mount and process namespace, reaches its capabilities through the MCP tool servers in mcps/ (files, memory, tasks, meetings, notifications, and community mirrors), and streams every tool call back to the React dashboard in dashboard/ over WebSocket. Agents run autonomously on schedules, webhook events, or one-off times, produce files and reports in their workspace, and raise notifications in four severities. They place and answer phone calls through the phone/ daemon (Twilio or FreePBX/Asterisk), and satellite/ pairs remote laptops or workstations with a one-line install so the same agent runs locally with full machine access. Users edit Word, Excel and PowerPoint files live in chat, generate and edit images, video and audio, open a Claude Code / Codex terminal inside the dashboard, and install ready-made agents, tools and skills from the community catalog in one click.
- A small-team lead who wants a workforce of 'digital employees' on existing Claude/ChatGPT subscriptions instead of buying separate API credits per member
- A system admin who needs agents confined to kernel sandboxes with service-by-service grants to LAN services and MCP tools
- Marketing or ops staff who want agents to run on schedules, draft reports, process Excel/Word files, and escalate via severity-graded notifications
- A developer who wants an agent working directly on a paired laptop or workstation (macOS/Linux/Windows) with local file and terminal access
- A support or assistant scenario needing an agent with a real phone number that takes and places calls via Twilio or an existing Asterisk/FreePBX
- A home user who wants agents controlling smart-home devices through MCP tools (Home Assistant ships in the community catalog)
What are this agent's strengths and limitations?
- Bring-your-own-subscription economics: reuse the team's existing Anthropic and OpenAI subscriptions, one per user, instead of pooled API spending
- Security-first defaults: every agent session runs in its own kernel sandbox, private ranges and cloud metadata endpoints are unreachable by design, and credentials are encrypted at rest and injected per session so agents use but never see them
- Complete agent form factor: persistent memory, scheduled/webhook autonomous runs, agent-to-agent delegation and meetings, real phone calls, and in-chat editing of Office files
- Remote machine pairing is one command with no open ports (single outbound connection), and the server takes over when a paired machine goes offline
- SSO/OIDC, two-factor auth, and per-user/per-agent cost budgets are standard from the first install — genuinely team-ready
- FSL-1.1-Apache-2.0 (fair source): commercial use competing with OtoDock is prohibited, and self-hosting beyond 5 users requires seat licensing — not pure open source
- Production deployment requires a Linux server with Docker (4 GB RAM minimum); there is no managed, zero-ops option beyond self-hosting
- Core execution depends on the Claude Code and Codex runtimes — using those engines requires active Anthropic or ChatGPT subscriptions, a substantive vendor tie
- Telephony requires your own Twilio account or an existing Asterisk/FreePBX server, adding cost and operational burden
- The GitHub License field shows NOASSERTION, and the young ecosystem (community catalog of agents/MCPs/skills) is still largely unproven
How do you install or deploy this agent?
Requires a Linux server with Docker (4 GB RAM minimum). In your install folder run:
mkdir otodock && cd otodock
curl -fsSLO https://raw.githubusercontent.com/OtoDock/oto-dock/main/scripts/install.sh
bash install.shThe installer checks Docker, writes a .env with a generated database password, handles the Ubuntu 24.04+ host step automatically, downloads the release-pinned docker-compose.yml plus the phone-service overlay, and starts the stack. If users reach the server by name or IP, set DASHBOARD_PUBLIC_URL in the .env; behind a reverse proxy also set TRUSTED_PROXY to your proxy's IP. Fresh installs only — the script stops rather than touching an existing install.
How do you use this agent?
Open http://localhost:8400 and let the setup wizard create your owner account; OtoDock auto-installs a Personal Assistant with its required tools. Connect your Claude or ChatGPT subscription under Setup → AI Engines (each user connects their own subscription; API keys and local models also work). Open the Personal Assistant from the Agents page and send a message — replies, tool calls and created files stream live. From there, add agents from the community catalog or build your own (editing Persona, memory, workspace, knowledge, skills and tools), pick Claude Code or Codex per agent (switchable per chat), set a sharing mode and schedules, and invite teammates with Admin/Creator/Member platform roles and Manager/Editor/Viewer roles per agent.
How does this agent compare with similar options?
The README positions OtoDock as 'the agentic company OS' in contrast to using Claude Code or Codex directly in a terminal: it wraps those two engines into a multi-tenant, shareable, autonomously running team platform. Functionally it occupies the same self-hosted personal-agent space as OpenClaw-style projects, but emphasizes multi-user sharing, departmental delegation, and enterprise security (SSO, sandboxing, budgets).