OmniAgent

Inspiration

Web AI assistants are powerful, but each platform keeps its own conversation, context, and capabilities. Switching between them means losing continuity, while personal knowledge and tool permissions remain fragmented.

What it does

OmniAgent is a browser-extension-based personal agent system that lets users bring their own web AI as the reasoning engine while keeping the agent's memory, skills, tools, projects, tasks, and execution history under the user's control. It connects web AI providers through site adapters and gives a task the same durable context when the provider changes.

How we built it

We built a TypeScript monorepo with a Vue browser extension and modular packages for storage, memory, skills, tools, browser automation, site adapters, MCP, and the agent runtime. The extension integrates with web AI pages through adapters, while IndexedDB-backed storage preserves tasks, memory, project context, and execution steps locally.

How Codex and GPT-5.6 helped

We used Codex as a collaborative coding agent to inspect the monorepo, plan changes across packages, implement TypeScript and Vue features, and verify the build and type checks. GPT-5.6 supported iterative reasoning about the cross-provider agent loop, protocol boundaries, persistence, and recovery flows.

Challenges we ran into

The hardest part was maintaining a clean boundary between the web model that reasons and the local agent system that owns state. We designed the runtime so task state, tool results, permissions, memory, and retry/pause/resume controls do not depend on any single AI website.

Accomplishments that we're proud of

• A reusable agent core with persisted tasks, steps, pause, resume, and retry.
• Cross-provider web AI adapters for DeepSeek and Kimi.
• Local, scoped memory with evidence, revisions, candidate review, and safe context injection.
• A modular tool runtime, browser agent, skills, project context, and MCP foundation.

What's next for OmniAgent

Next we are completing the model-driven tool loop, a unified tool-call protocol, standard MCP transports, stronger browser automation, and semantic memory retrieval. The long-term goal is one durable personal agent that can use the best available web AI without making the user's tasks and knowledge belong to a single platform.

Built With

• typescript