|

Databricks Open-Sources Omnigent: A Meta-Harness That Composes, Governs, and Shares AI Agents Across Claude Code, Codex, and Pi

ByRicardo

Databricks launched Omnigent, an open supply ‘meta-harness’ for AI brokers. The undertaking ships below the Apache 2.0 license. The Databricks AI group constructed it with Neon.

A harness is the wrapper round a mannequin that turns it into an agent. Claude Code, Codex, and Pi are harnesses. Omnigent sits one stage above them. It treats every harness as an interchangeable half of a bigger system.

Many engineers now juggle 4 or 5 brokers directly. They copy textual content between coding brokers, search instruments, Docs, and Slack. Each harness solely understands its personal classes. Omnigent provides a shared layer the place composition, management, and collaboration reside.

What is Omnigent

Omnigent is a standard interface above command-line brokers and agent SDKs. It wraps terminal coding brokers similar to Claude Code, Codex, and Pi. It additionally wraps SDKs like OpenAI Agents and the Claude Agents SDK.

The design rests on one commentary. However a harness calls its mannequin internally, the user-facing interface is identical. Messages and recordsdata go in. Text streams and instrument calls come out. Omnigent standardizes that interface so harnesses turn into swappable.

You provide the fashions and the infrastructure. Omnigent runs the brokers on high. It can coordinate a number of of them as interchangeable employees below one orchestrator.

How Omnigent Works

The structure has two components. A runner wraps any agent in a sandboxed session with a uniform API. A server gives insurance policies and sharing. The server exposes each session over the terminal, the app, and internet APIs.

One command begins a session in your terminal. It additionally launches an area internet UI at localhost:6767. The similar session seems within the browser or on a cellphone. Messages, sub-agents, terminals, and recordsdata keep in sync.

The CLI installs below two names, omnigent and omni. They are interchangeable. On first run, it detects mannequin credentials already in your atmosphere.

https://omnigent.ai/

Composition, Control, and Collaboration

Databricks group frames Omnigent round three capabilities:

  • Composition means combining fashions, harnesses, and methods with out rewriting code. You swap between Claude Code, Codex, Pi, and customized brokers with one-line adjustments.
  • Control means stateful, contextual insurance policies. They monitor agent actions and implement guardrails on the meta-harness layer, not via prompts. One instance pauses an agent after each $100 it spends. Another requires human approval to git push as soon as the agent installs a brand new npm package deal.
  • Collaboration means sharing reside agent classes by URL. Teammates watch the agent work and chat with it in actual time. They can touch upon recordsdata, co-drive the session, or fork the dialog.

An OS sandbox, referred to as Omnibox, underpins this. It can lock down OS entry and remodel community requests. For occasion, it might hold your GitHub token hidden from the agent. The token is injected solely within the egress proxy on authorized requests.

Use Cases and Examples

Two instance brokers ship with the repository:

  • Polly is a multi-agent coding orchestrator. It writes no code itself. It plans, then delegates work to coding sub-agents in parallel git worktrees. Each diff routes to a reviewer from a special vendor than the author. You merge the outcome.
  • Debby is a brainstorming companion with two heads. One head is Claude, the opposite GPT. Every query goes to each, with solutions proven facet by facet. Type /debate and the heads critique one another earlier than converging.

Other sensible patterns comply with the identical form. A frontier advisor mannequin can information a less expensive open-source employee. A lead agent can orchestrate parallel subagents. Different LLMs can deal with planning, search, and code era in a single circulate.

Interactive Concept Demo

Marktechpost group has created a interactive demo (under) that allows you to expertise Omnigent’s meta-harness workflow firsthand. You decide a job for the Polly orchestrator, which plans it and delegates to a few sub-agents: Claude Code, Codex, and Pi which are working in parallel and streaming their steps reside. A session value meter ticks up as they work, and the 2 coverage toggles present Omnigent’s management layer in motion: the price funds pauses the run at $3.00 to your approval, and a contextual coverage halts a git push that follows an npm set up till you enable it. Once the sub-agents end, every diff is cross-reviewed by a special vendor than the one which wrote it, then marked able to merge. Terminal, Web, and Mobile tabs present the identical session staying in sync throughout interfaces. It’s an illustrative simulation, no reside fashions are referred to as.

Omnigent Meta-Harness

One orchestrator. Many harnesses. One ruled session.

Interactive idea demo





Session LLM value
$0.00

Orchestrator · Polly (writes no code; plans & delegates)
Idle. Pick a job and press “Run session”.

Claude Codeready

Codexready

Piready

Illustrative simulation of the Omnigent workflow — no reside fashions are referred to as.
Learn extra at omnigent.ai ·
GitHub · Apache 2.0 · Alpha.




Marktechpost
·
AI Dev & Research Media

Omnigent Meta-Harness

One orchestrator. Many harnesses. One ruled session.

Interactive idea demo





Session LLM value
$0.00

Orchestrator · Polly (writes no code; plans & delegates)
Idle. Pick a job and press “Run session”.

Claude Codeready

Codexready

Piready

Illustrative simulation of the Omnigent workflow — no reside fashions are referred to as.
Learn extra at omnigent.ai ·
GitHub · Apache 2.0 · Alpha.




Marktechpost
·
AI Dev & Research Media

Omnigent vs a Single Harness

Capability Single harness (e.g., Claude Code) Omnigent meta-harness
Agents and fashions One harness; swap fashions inside it Claude Code, Codex, Pi, SDKs, customized — interchangeable
Switching value Re-integrate per instrument One-line change
Interfaces Terminal or that instrument's personal UI Terminal, internet, desktop, cellular, APIs — similar session
Governance Allow / deny lists, usually prompt-based Stateful contextual insurance policies on the harness layer
Cost management Manual monitoring Budget coverage pauses at set thresholds
Collaboration Copy-paste between instruments Live shared classes, co-drive, and fork
Sandbox Tool-dependent OS sandbox plus egress-proxy secret injection
Cloud execution Local machine Disposable Modal or Daytona sandboxes
License Varies Apache 2.0, open supply

Getting Started

Omnigent wants Python 3.12+, Node.js 22 LTS, and tmux. One command installs all the things:

Copy Code
curl -fsSL https://omnigent.ai/set up.sh | sh

Then arrange mannequin credentials:

Copy Code
omni setup

Omnigent accepts 4 credential varieties. They are a first-party API key and a Claude or ChatGPT subscription. The others are an OpenAI- or Anthropic-compatible gateway and a Databricks workspace. The /mannequin command switches fashions mid-session.

A customized agent is a brief YAML file. It declares a immediate, a harness, instruments, and optionally available sub-agents.

Copy Code
identify: my_agent
immediate: You are a useful knowledge analyst.

executor:
  harness: claude-sdk          # or: codex, codex-native, claude-native, openai-agents, pi

instruments:
  researcher:
    kind: agent
    immediate: Search for related data and summarize it.

Run it with one command:

Copy Code
omnigent run path/to/my_agent.yaml

Policies use the identical YAML strategy. This builtin caps spend with a gentle warning first:

Copy Code

Policies use the identical YAML strategy. This builtin caps spend with a gentle warning first:

insurance policies:
  funds:
    kind: operate
    handler: omnigent.insurance policies.builtins.value.cost_budget
    factory_params:
      max_cost_usd: 5.00          # arduous spend cap
      ask_thresholds_usd: [3.00]  # gentle warning on the best way

Policies stack throughout three ranges. They are server-wide, per-agent, and per-session. The stricter session guidelines are checked first.

Strengths and Limitations

Strengths

  • One interface to Claude Code, Codex, Pi, and customized brokers
  • Sessions reachable from terminal, internet, desktop, and cellphone
  • Policies that monitor state, not simply enable or deny
  • Live session sharing replaces copy-pasting between instruments
  • Cloud sandboxes on Modal and Daytona want no native laptop computer
  • Apache 2.0 license with targets like Fly.io, Railway, and Render

Limitations

  • The undertaking is alpha and early in its lifecycle
  • It requires Python, Node.js, and tmux setup
  • You convey your individual fashions, infrastructure, and spend
  • Roadmap gadgets just like the Omnigent Server MCP usually are not shipped but
  • Off-network teammates want an always-on deployed server to hitch

Check out the Repo, Quick start, and Technical details . Also, be happy to comply with us on Twitter and don’t overlook to hitch our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

Need to companion with us for selling your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar and many others.? Connect with us

The submit Databricks Open-Sources Omnigent: A Meta-Harness That Composes, Governs, and Shares AI Agents Across Claude Code, Codex, and Pi appeared first on MarkTechPost.

Similar Posts