Omnara is an open-source managed agent platform released under Apache 2.0 that positions itself as a self-hostable alternative to Claude Managed Agents and OpenAI's Agents API. The pitch is structural: instead of locking your agent orchestration into a single vendor's cloud and model ecosystem, Omnara lets you run agents on any compatible model — OpenAI, Anthropic, OpenRouter, Bedrock, LiteLLM, Ollama, or self-hosted endpoints — on any machine you control. The architecture is straightforward. A REST API handles agent lifecycle (launch, message, approve). Postgres stores every step of every agent run as durable state, so any stateless worker can resume an agent after a crash. Workers build model context, call models, and execute tools. Machines run an outbound-only daemon (omnarad) requiring no inbound ports or SSH keys, and agents can target sandboxes from eleven providers (Arker, Blaxel, E2B, Modal, and others) or your own laptops, VMs, and containers. Redis, blob storage, and a memory filesystem handle the plumbing. The feature set maps closely to what the closed platforms offer: built-in tools for shell commands, files, web search, and human-in-the-loop questions; MCP server support; custom tools; skills (packaged instructions reusable across agents); persistent memory stores shared between agents; subagent spawning; cron schedules with time zones; and integrations for Slack, Discord, and GitHub. Permissions are enforced at the platform level — each tool call can be auto-approved, require human approval, or be blocked entirely. Multi-tenancy is built in. A single deployment serves multiple organizations, each with isolated members, projects, secrets, model providers, and machines. Self-hosted operators can provision default models and sandboxes for new organizations, mirroring what Omnara Cloud does. Role-based access control covers users and API keys, and credentials are stored once and shared explicitly per project. The competitive comparison table is the most revealing artifact. Claude Managed Agents locks you to Claude models on Anthropic's cloud. OpenAI's Agents API locks you to OpenAI models (with Codex as the open-source harness but the service itself closed). Omnara claims model-agnostic orchestration, self-hostable infrastructure, and agent history stored in your own Postgres — queryable for analytics, evals, and training datasets. The data ownership angle is pointed: Anthropic's offering is noted as "not currently eligible for Zero Data Retention or HIPAA BAA coverage." The catch is operational burden. Self-hosting means you run Docker Compose, manage Postgres, configure model providers, handle security hardening (the local defaults are "intentionally insecure"), and maintain uptime. The project ships a TypeScript SDK, CLI, and OpenAPI spec, but the ecosystem is nascent. Whether Omnara can sustain the velocity needed to keep pace with Anthropic and OpenAI's rapid iteration on their closed platforms is the real question. This is infrastructure-layer competition, not model-layer competition. Omnara doesn't make models — it routes to them. The bet is that the orchestration layer should be open and portable, not locked to the model vendor. If the agentic computing thesis plays out, whoever controls the orchestration layer controls the switching costs. Omnara is trying to ensure that layer stays contestable.
