Omnigent
github.com/omnigent-ai/omnigent- Category
- AI Agents
- Rank
- No. 780Tools index
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Agent Building · Security & Identity
- Interfaces
- CLI · Web · Desktop · Mobile
- Builder
- omnigent-ai
- GitHub
- 10.2k stars
- Latest release
- v0.15.0
- Date
About
An open-source meta-harness that sits above existing agent harnesses like Claude Code, Codex, and Cursor, letting users compose or swap them without rewriting code. It adds a shared policy layer for spend caps, tool access, and OS-level sandboxing, plus real-time collaborative sessions accessible from terminal, browser, mobile, or desktop.
What it does
Omnigent runs agent sessions defined through YAML or selected interactively. A session can combine different coding runtimes, custom ACP agents, MCP tools, sub-agents, terminals, policies, and scoped operating-system access. The same conversation can continue through its terminal, web, phone, or macOS desktop interfaces.
Why it's ranked here
The project tackles orchestration, governance, remote access, and deployment as one coherent system. Its declarative agent format and layered policy model are unusually concrete. However, the alpha status, substantial dependency surface, and degraded Windows isolation make it better suited to technical adopters than teams seeking a quiet, mature default.
What's good
Policies return allow, deny, or ask decisions and can apply per session, agent, or server. Budget gates cover model requests as well as tool calls. YAML specifications describe models, authentication, tools, sub-agents, policies, terminals, and sandbox access. Cloud sandbox choices and custom ACP support keep execution options broad.
Tradeoffs
The baseline requires Python 3.12, while native terminal wrappers also need tmux and coding harnesses may need Node.js. Linux wrappers require Bubblewrap. Native Windows lacks the terminal wrappers, filesystem isolation, network isolation, and egress proxy available elsewhere. Building from source may also require the web workspace toolchain. The repository labels the project alpha.
How to use it well
Use Omnigent when several agent runtimes must share governance, session history, collaboration, or reproducible YAML configuration. Start with one agent, narrow filesystem and network permissions, then add policy gates and specialist sub-agents. It does not replace an underlying model, coding harness, provider account, or production identity and storage stack.
Technical notes+
pyproject.toml defines a setuptools Python 3.12+ package, two client package dependencies, server libraries, baseline Claude and OpenAI agent SDKs, and numerous optional provider, sandbox, tracing, harness, storage, and memory extras. setup.py builds the pnpm web workspace, bundles example agents and maintenance scripts, and generates wheel build metadata before normal packaging. package.json pins pnpm for the private workspace. docs/AGENT_YAML_SPEC.md documents declarative executors, MCP and Python tools, ACP agents, policies, terminals, containers, and platform-selected sandbox backends. docs/POLICIES.md specifies ordered ALLOW, DENY, and ASK evaluation across session, agent, and server scopes. docs/openclaw.md records ACP gateway boundaries and manual validation limits.
Observed
- License
- Apache License 2.0
- Primary language
- Python, requiring Python 3.12 or newer
- Packaging
- Published Python package installable with uv or pip, with a Homebrew option
- Interfaces
- CLI aliases, local web UI, macOS desktop app, REST API, MCP tools, and ACP agents
- Platform support
- Linux and macOS support native wrappers and OS sandboxing; Windows runs in degraded mode
- Project status
- Classified and badged as alpha
- Configuration
- Agents and server policies use YAML specifications
Read from README.md, setup.py, package.json, pyproject.toml, docs/POLICIES.md, docs/openclaw.md, docs/databricks.md, docs/QWEN_FOLLOWUPS.md, docs/AGENT_YAML_SPEC.md, docs/UNINSTALL_DESIGN.md.
What it can do
Wrap and swap multiple coding agent harnesses (Claude Code, Codex, Cursor) without rewriting code
Agent harness configuration → Unified agent interface
Sandbox agent sessions at the OS level
Agent session → Sandboxed execution environment
Sync collaborative sessions in real time across terminal, browser, mobile, and desktop
User session → Synchronized session state
Intel on Omnigent
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.