Omnara (YC S25) is an open-source, model-agnostic control plane for managed agents: you define agents in YAML, it handles execution and state, and you drive them from a web dashboard, mobile apps, CLI, REST API, or Slack. Facts below verified as of 2026-09-13.
Omnara has quietly pivoted from “run Claude Code from your phone” into the open-source alternative to Claude Managed Agents, and that repositioning makes it infrastructure for serving agents, not just supervising them.
What it is #
A Go platform (Apache-2.0) where an agent is a small YAML config: instruction, model provider config, tools, permissions, MCP servers, skills, and the machine pool it runs on. It exposes a dashboard, CLI, and REST API, plus first-party iOS/Android apps and a Slack connector, with approvals, questions, streamed events, and artifacts as first-class API concepts. Self-host it or use Omnara Cloud (app.omnara.com); models are BYOK through provider configs such as OpenRouter presets. Founded by Ishaan Sehgal (ex-Microsoft, led Kaito) and Kartik Sarangmath, per its YC profile. It sits in the orchestration layer of this section’s taxonomy (control plane and supervision, like Paseo), not the harness layer: it drives and manages agents rather than being one.
Status #
Active, well-capitalized by YC standards, and shipping daily.
2,837 stars and 220 forks as of 2026-09-13 on a repo created July 9, 2025, with pushes landing September 12, 2026 (GitHub API).
Two HN threads anchor its traction: a 310-point Show HN on August 12, 2025 (HN) and a 147-point Launch HN on February 12, 2026 (
HN).
The product’s own framing moved between those dates, from “run Claude Code from anywhere” to “the API for production-grade agents” with managed-agent execution.
Strengths #
- The mobile-and-API surface is the differentiator: approvals, questions, and streamed events work the same from an iPhone, Slack, or a REST call, which is exactly what unattended agent fleets need.
- Agent-as-YAML makes agents reviewable artifacts: configs live in your repo, execution and state live in Omnara, and machine pools separate where code runs from who can invoke it.
- Model-agnostic and BYOK by design, with self-hosting under Apache-2.0 if you want the control plane on your own hardware.
- Its positioning against Claude Managed Agents gives it a crisp comparison point for teams avoiding vendor lock-in.
Cautions #
- The pitch has moved twice already (mobile Claude Code remote, then managed-agents control plane), so the product you adopt today may be framed differently a year from now.
- You are trusting a young startup (YC S25) with the execution and state layer for your agents; self-hosting mitigates but does not eliminate that.
- The docs are thinner than the platform’s ambition; the feature surface (machine pools, integrations) is documented but young.
- As a control plane it still needs actual harnesses underneath; it does not replace your choice of agent.
Pricing #
The platform is open source under Apache-2.0 and self-hostable at no cost. Omnara Cloud now publishes usage-based pricing, a change from the no-pricing-page state I recorded on 2026-08-30: $0 platform fee with no per-seat charges, Omnara-managed models at provider token rates, and Omnara-managed machines at $0.0414 per GiB of memory per hour online (idle machines auto-sleep) plus $0.20016 per GiB of memory per 30 days of retention, as of 2026-09-13. Enterprise adds deployment, integrations, and support at custom terms. You pay your own model costs when you bring your own keys.
Compared to #
- Paseo: the closest sibling, also FOSS self-hosted supervision from your phone; Paseo wraps CLI agents you already run, while Omnara wants to own execution and state via its API.
- Conductor: the Mac-native multi-session manager for local parallel work; Conductor is a desktop app around Claude Code/Codex sessions, Omnara is a server-grade control plane.
- Claude Code: Anthropic’s own managed-agent offering is the closed counterpart Omnara positions against; pick Claude Managed Agents for turnkey, Omnara for openness and BYOK.
Bottom line #
Recommended for teams running agents as products or internal services who want approvals, events, and mobile control behind one API on their own infrastructure; not for solo developers who just want to drive one CLI agent from their phone (Paseo or the harness’s own remote features do that with less machinery). I believe every serious agent deployment eventually needs exactly what Omnara is building, and I think the open-source control plane will beat vendor-owned managed agents the same way Postgres beat proprietary databases.
Changes #
- 2026-08-30 - Created when the candidate-resolution run accepted it into Orchestration as a control plane rather than a harness.
- 2026-09-08 - Recorded Omnara Cloud’s newly published usage-based pricing and added the pricing page as a source.
- 2026-09-09 - Added the machine-retention line item, $0.20016 per GiB per 30 days.
See also #
- Orchestration Feature Matrix - the supervision layer Omnara belongs to
- Paseo - the closest open-source sibling for phone-based agent supervision
- Agentic Coding Tools Landscape - where control planes sit above harnesses
- OpenChamber - self-hosted scheduling, the unattended complement to interactive supervision
References #
https://github.com/omnara-ai/omnara - repository, Apache-2.0, 2,837 stars as of 2026-09-13
https://www.omnara.com/pricing - usage-based pricing: $0 platform fee, token-rate models, per-GiB machine runtime and retention (2026-09-13)
https://docs.omnara.com/introduction - agent YAML model, dashboard/CLI/API surfaces, use cases
https://raw.githubusercontent.com/omnara-ai/omnara/main/README.md - current positioning and getting-started flow
https://www.omnara.com/ - “open-source alternative to Claude Managed Agents” positioning
https://www.ycombinator.com/companies/omnara - YC S25 batch, founders, self-host-or-cloud model
https://news.ycombinator.com/item?id=44878650 - the August 12, 2025 Show HN, 310 points (verified via Algolia API)
https://news.ycombinator.com/item?id=46991591 - the February 12, 2026 Launch HN, 147 points (verified via Algolia API)