Omnigent on GitHub Trending
Omnigent is a new open-source project from omnigent-ai that reached GitHub Trending this week. The repository
Core architecture and agent swapping
Omnigent defines agents through YAML files that declare model endpoints and tool permissions. A session can load multiple agents at once, so one call might route a code review task to a Codex-backed agent while another spawns a custom Python script agent for data extraction. The harness translates requests into each backend's native format and merges their outputs into a shared context store.
Because the abstraction lives above the individual harnesses, existing agent code does not need modification. Sessions persist state for files, terminal processes, and sub-agent calls, allowing resumption from the CLI, a browser, or the macOS desktop app. The server component can launch disposable sandboxes on Modal or Daytona when a session requests cloud execution.
Policy enforcement and sandboxing
Policies are expressed as rules that trigger on specific tool calls or spend thresholds. A policy can require explicit approval before a file-write operation outside a designated directory or cap total token usage for a session. These rules are evaluated server-side before any agent action reaches the underlying model.
Sandbox isolation is handled at the provisioner level. When a managed host is used, each session receives its own container with restricted network and filesystem access. Local runs still respect the same policy checks but rely on the host operating system's user permissions rather than container boundaries.
Collaboration and device handoff
A session identifier can be shared with other users. Recipients join the live context and can inject messages or fork the conversation into their own session. State synchronization covers terminal output, file edits, and agent reasoning traces, so multiple participants see the same view without manual copying.
The implementation uses a WebSocket layer for real-time updates and an OpenAPI-defined REST surface for session management. The Python SDK exposes the same primitives used by the CLI, making it straightforward to embed Omnigent session control inside existing Node.js or Rails automation scripts.
Trade-offs observed so far
The YAML agent definitions add an indirection layer that requires initial setup for each new backend. Policy evaluation adds measurable latency on every tool invocation, which becomes noticeable in tight loops. Sandbox provisioning on external hosts also introduces cold-start delays compared with direct local execution.
On the positive side, the unified context store removes the need to maintain separate conversation histories when switching models mid-task. Teams can enforce spend limits centrally rather than relying on individual API keys.
FAQs
Does Omnigent require a paid subscription? No. It accepts first-party API keys, existing Claude or ChatGPT subscriptions, and any OpenAI-compatible gateway.
Can I run Omnigent entirely offline? Local agents work without network access once models are available, but cloud sandbox features and multi-device sync require connectivity.
How does Omnigent handle agent-specific tool schemas? The harness normalizes common tool categories and passes unknown schemas through to the original backend, preserving each agent's native capabilities.
---
📖 Related articles
- Agentic Coding: Una Trappola per lo Sviluppo Software?
- Claude Code: da prompt a agente con skills e subagent
- File agents.md: utili per gli agenti di coding?
Need a consultation?
I help companies and startups build software, automate workflows, and integrate AI. Let's talk.
Get in touch