Omnigent

Omnigent is the meta-harness for a team that wants to stop arguing about which agent to standardise on. One session can run Claude Code, Codex, Cursor, OpenCode, Hermes, Pi or a custom YAML agent, and you swap harnesses without rewriting the task. The trait that decides it is the label on the box: alpha, with a first release in June 2026.

Try it if you have a platform team whose job is exactly this. Do not put a product on it yet. Pick Paperclip for governance you can run today, and Ruflo if you want swarms over one host agent.

The risk is a half-built sandbox. The README states Windows support is degraded: no native terminal wrappers and no filesystem sandboxing, use Linux, macOS or WSL. A policy layer whose isolation depends on the operating system is a policy layer with a documented hole, and in alpha there is no schedule for closing it.

Run it on Linux or not at all. What it does right: the policy engine has a per-session tool-call limit, fifty calls in the example, so a looping agent stops at a number instead of at your patience.

The abstraction is the interesting part. 1. An agent is defined in YAML with a name, a prompt, an executor harness and a tool list, where a tool is a local Python function, an MCP server or a sub-agent. 2. Because the harness is a field, the same definition runs under a different agent by editing one line. 3. Collaboration is share, co-drive, where a colleague attaches and issues commands, and fork, which clones the conversation.

No benchmark is published, and none should be for an alpha. The observation: making the harness a parameter is the first design here that treats coding agents as interchangeable, which they are not yet.

Omnigent AI has published no price and no plan, which at 9,660 stars is a seed-stage posture: build the layer, sign the partners, price later. The partners are visible in the sandbox list, Modal, Daytona, Blaxel, E2B, CoreWeave, Kubernetes, Databricks, and a control plane that routes work into all of them is a channel every one of those vendors would like to own.

Moat: none yet; switching cost is a YAML file. Likely acquirer: a sandbox or compute vendor on that list, Databricks or CoreWeave first. Likely pivot: a hosted control plane with per-seat pricing. Position: watch; do not build on an alpha.

The demo is a spend cap stopping an agent mid-task. Procurement noted the controls: spending caps with soft warnings at thresholds, shell command approval, and tool restrictions defined centrally rather than per laptop, which is what sixty engineers need. What they arrive in is an alpha that requires Python 3.12 or newer and Node.js 22 on every host, with no SSO, no audit export and no support organisation.

It does not run in CI. Onboarding is a curl script and a YAML file per agent. Not yet; the policy layer is the right idea and we will re-read it after a stable release.

Apache-2.0 Python I can pip install or uv tool install, and the model layer takes Ollama, vLLM and LiteLLM alongside the hosted keys, so a session can run entirely on my own hardware. MCP is a tool type in the agent YAML, type: mcp with a url or a local command, which means my existing servers plug in without a separate config file.

The sandbox list leans on cloud vendors, but Kubernetes and OpenShell are on it, and the OS-level sandbox works on Linux. A fork would be a Python package with a new name. Alpha, readable, mine if it survives.