Omnigent
A shared control layer for coding agents, with multi-agent coordination, policy enforcement, and sandbox-backed execution.
Per-dimension scores and reasoning
Evidence shows policy system (e.g., ask_on_os_tools) and budget caps, supporting user confirmation; but data flow transparency only mentions telemetry, not detailed internal data flows; sensitive data handling has keyring and file permission notes; dependencies have version locks and comments, but no vulnerability scanning mentioned; external effects have sandbox and policy controls; rollback has uninstall and upgrade mechanisms; source attribution unverified publisher, but code has author info. Deductions: insufficient data flow transparency, unverified source attribution.
Self-consistency: README and pyproject.toml versions match, but dev0 version exists; dependency availability: version locks and comments, but no mirrors or fallback sources; failure messages: error handling examples, but not systematic. Deductions: version inconsistency (dev0), insufficient failure message documentation.
Audience clear (developers), scenarios diverse (multi-device, multi-agent); capability boundaries documented (Windows limitations); trigger precision via policies and commands; environment fit cross-platform. Deductions: no major flaws.
Information architecture clear (README, docs); install notes detailed; naming stable (omnigent/omni); examples abundant (Polly, Debby); known limitations documented (Windows); license clear (Apache-2.0); versioning has pyproject.toml but no CHANGELOG; maintenance responsibility has CONTRIBUTING and MAINTAINER. Deductions: no CHANGELOG, version is dev0.
Output usability: CLI and Web UI; marginal value: multi-agent orchestration and policies; cost-benefit: open-source free, but many dependencies. Deductions: no major flaws.
Claim traceability: README has links but not verified; cross-source corroboration: tests exist but not independently verified; fact-inference separation: documentation distinguishes facts and inferences. Deductions: claims unverified, tests not independently run.
- Publisher identity unverified; assess supply chain risk carefully.
- Version is dev0, may be unstable.
- Many dependencies; watch for security vulnerabilities.
- Telemetry enabled by default; be aware of privacy.
What does this agent do, and when should you use it?
Omnigent is an open-source meta-harness for Claude Code, Codex, Cursor, OpenCode, Hermes, Pi, and YAML-defined custom agents. It supplies a common session and orchestration layer, so teams can switch or combine harnesses, delegate work to sub-agents, and route work for cross-vendor review. The same session can be accessed from the terminal, browser, phone, or macOS desktop app, with messages, terminals, files, and sub-agents synchronized. Its Agent YAML format can expose local Python functions, MCP servers, and delegated agents, while policies govern shell access, file writes, tool usage, and spend. It can run locally or as a deployed server with registered hosts and cloud sandboxes; Windows supports the server, web UI, and SDK harnesses in a reduced mode but lacks the native PTY wrappers and filesystem/network isolation.
Running omnigent or omni starts a local session and a web UI at http://localhost:6767; dedicated commands include omnigent claude, omnigent codex, omnigent cursor, omnigent opencode, omnigent hermes, and omnigent pi. omnigent setup adds, selects, or removes per-agent API-key, Claude/ChatGPT subscription, compatible-gateway, and Databricks credentials, while /model changes a model during a session. omnigent run path/to/my_agent.yaml reads an agent definition containing prompt, executor, and tools, then invokes declared local Python callables, MCP endpoints, or sub-agents. A server can be started with omnigent server --background and a machine registered with omnigent host; omnigent attach <session_id> co-attaches to a session, while omnigent run --fork <session_id> forks one. Policies stack at server, agent, and session level and evaluate each action as allowed, blocked, or approval-required; documented built-ins include ask_on_os_tools, max_tool_calls_per_session, and cost_budget.
- An engineering lead who wants Claude Code, Codex, or Pi sub-agents to work in parallel git worktrees and send each resulting diff to a different vendor's agent for review can use the orchestration pattern demonstrated by Polly.
- A developer with both Claude and OpenAI credentials who wants parallel answers from two model heads, followed by critique rounds through
/debate, can run the Debby example. - A distributed engineering team that needs colleagues to watch a live agent, inspect its terminals and files, send messages, or co-drive from the host machine can use shared sessions and
omnigent attach. - A team that needs to continue a development session from a browser or phone can run a local server and host, or deploy Omnigent to a stable URL.
- A platform or security team that needs approval before shell commands or file writes, plus tool-call and spend limits, can configure built-in policies at server or agent level.
- A developer building a specialized assistant from Python functions, MCP services, and delegated specialists can define and run an Agent YAML file.
What are this agent's strengths and limitations?
- Provides one orchestration layer across Claude Code, Codex, Cursor, OpenCode, Hermes, Pi, and YAML-defined agents, including the ability to switch or combine harnesses.
- Synchronizes live chat, sub-agents, terminals, and files across terminal, browser, phone, and the macOS desktop app, with sharing, co-driving, and session forks.
- Supports stacked server, agent, and session policies with documented controls for OS-tool approval, tool-call ceilings, and cost budgets.
- Agent YAML can combine local Python functions, MCP services, and delegated agents while supporting API keys, subscriptions, compatible gateways, and Databricks credentials.
- The repository labels the project alpha, so adopters should plan for validation and potential interface or behavior changes.
- The full native terminal workflow depends on Python 3.12+, uv, git, Node.js, pnpm, and tmux; Linux native wrappers additionally require bubblewrap.
- Windows is explicitly reduced: native tmux/PTY wrappers are unavailable, and its Job Object backend does not isolate the filesystem or network or provide an L7 egress proxy.
- Cloud sandboxes, Databricks, Bedrock, Vertex, S3, Hindsight, and some SDK harnesses require optional extras or external-service configuration.
How do you install or deploy this agent?
Python 3.12+ is required. Install with uv tool install omnigent or pip install "omnigent"; both omnigent and omni become available. uv and git are required; coding-harness CLIs need Node.js 22 LTS+ and npm, while the web UI needs pnpm. Native omnigent claude, omnigent codex, omnigent cursor, omnigent hermes, omnigent kiro, and omnigent pi terminal wrappers require tmux; Linux additionally requires bubblewrap, while macOS uses seatbelt. For a first session, run omnigent; it can detect ANTHROPIC_API_KEY, OPENAI_API_KEY, or an already signed-in claude or codex CLI.
How do you use this agent?
Run omnigent setup to add an API key, Claude/ChatGPT subscription, compatible gateway, or Databricks credential and choose a default for the relevant agent. Start a default session with omnigent, or choose a runtime such as omnigent codex or omnigent claude; the web UI is available at http://localhost:6767. To run a custom definition, use omnigent run path/to/my_agent.yaml; executor.harness may be claude-sdk, claude-native, codex, codex-native, cursor, cursor-native, hermes, hermes-native, opencode, pi, pi-native, or openai-agents. For browser-created sessions, run omnigent server --background, then run omnigent host in another terminal.
How does this agent compare with similar options?
Rather than binding a workflow to one coding harness, Omnigent explicitly supports Claude Code, Codex, Cursor, OpenCode, Hermes, and Pi, and documents combining those harnesses in one session or replacing them with YAML-defined custom agents.
FAQ
Can I use an existing Claude or ChatGPT subscription instead of buying API keys?
omnigent setup supports Claude Pro/Max and ChatGPT plans through the official claude and codex CLIs, as well as first-party API keys.