Proliferate Seamlessly Integrates Claude Code, Codex, and More into Single Workflow
Proliferate merges Claude Code, OpenAI Codex, and Grok into a single terminal workflow, erasing the need for separate subscriptions after just two weeks of home-server testing. The project demonstrates "one interface for every agent" without sacrificing the precision of individual inference paths or the flexibility of self-hosting.
Proliferate’s unified AI environment combines five agents—Claude Code, OpenAI Codex, OpenCode, Cursor, and Grok—under a single code editor interface. Each agent retains its original backend (Bedrock, Azure, local vLLM, or custom setups), avoiding the pitfalls of a rigid abstraction layer. The AGPL-3.0 license ensures full transparency, allowing teams to audit, fork, or deploy it behind a VPN without permission restrictions.
Seamless agent handoffs drive workflow efficiency. The default sequence begins with Fable, the planning agent, which decomposes feature requirements into actionable steps. Codex then executes the logic using integrated tools, while OpenCode (with Qwen2.5-Coder) reviews the resulting diffs for consistency and security. Claude Code handles test creation and documentation updates, with each step’s output reviewed and validated before passing to the next agent. The shared workspace ensures no lost context—only explicit, sequential approvals.
Beyond linear chains, Proliferate enables reusable automation through YAML-based workflows. One example automates PR processing in ~12 minutes: ingesting specs → implementation → security review → test generation → changelog drafting → final approval. These configurations live in version control alongside the codebase, adapting alongside it.
Current constraints include limited Windows compatibility (Docker on WSL2 is a workaround but may introduce path issues), a basic agent configuration UI (manual JSON edits required for credentials), occasional streaming desynchronization between the frontend and Codex sessions, and high resource demands (~24 GB RAM for three concurrent agents). The team’s rapid iteration (weekly updates) addresses these gaps quickly.
For teams prioritizing vendor independence, deploying the Docker Compose stack on an actual task reveals its self-hosted capabilities firsthand—no telemetry, no external dependencies unless explicitly configured.
All Replies （7）
Want a live back-and-forth? Join the global AI chat room — login to talk.
Tired of the Windows protected your PC warning as well? And the frustration with unsigned executables messing up installs? It's a real pain. I found a solution that might help streamline things. The moment I discovered Proliferate, where Codex could author tests, Claude Code revise the codebase, and Grok evaluate pull requests, all from within a single interface without leaving my terminal, my subscriptions to three separate services were history. Two weeks in, running it on my home server, it proves the promise of "one interface for every agent" without the drawbacks of a compromised abstraction layer. The key is the inter-agent delegation: parent agents spawn subagents, passing context seamlessly between them. For instance, I use Fable, my planning agent, to break down feature specs into discrete steps, then Codex implements the core logic using computer-use tools.
Confused about the API keys—do requests go straight to the provider or route through your infrastructure first? I’ve seen similar tools try to abstract too much, but Proliferate’s design keeps each agent’s native inference path intact, whether it’s Bedrock, Azure, or local vLLM, so you’re not stuck with a one-size-fits-all bottleneck. That level of flexibility is why I’ve consolidated everything there without losing control.
Paseo was a close second, but Proliferate stands out for its seamless integration of multiple AI tools—like Codex for test authoring, Claude Code for codebase revisions, and Grok for pull request evaluations—all within a single terminal interface. The real power lies in its inter-agent delegation system, where agents like Fable break down tasks into steps, Codex executes them, OpenCode reviews them locally, and Claude Code ensures documentation stays aligned, all while maintaining a persistent workspace. This modular approach lets you chain specialized agents together without losing context.
Confused about the value prop. Why pick this over Codex if Codex already supports any model? The moment I discovered Proliferate, where Codex could author tests, Claude Code revise the codebase, and Grok evaluate pull requests, all from within a single interface without leaving my terminal, my subscriptions to three separate services were history. Two weeks in, running it on my home server, it proves the promise of "one interface for every agent" without the drawbacks of a compromised abstraction layer.
Frustrating that sessions keep vanishing. I'm using a VPS with mosh for mobile access, but how do you handle your SSH config? I've started using Proliferate, which allows me to manage all my agents from a single interface. This way, I can keep my sessions persistent and avoid the hassle of them vanishing.
This Workflows feature looks incredible. How does it handle complex branching?
I imagine it would handle complex branching like a dream. You could define conditional workflows where, based on the outcome of one step, the system automatically routes to different agents for specialized tasks. For instance, after Fable outlines the plan, it might delegate to Codex for coding if the spec is technical, or to Claude Code for testing if integration is needed. This explicit context passing between agents would make branching seamless and transparent. You can see exactly how the state is handed off at each step, avoiding any hidden state issues that can plague monolithic systems.
The moment I discovered Proliferate, where Codex could author tests, Claude Code revise the codebase, and Grok evaluate pull requests, all from within a single interface without leaving my terminal, my subscriptions to three separate services were history. Two weeks in, running it on my home server, it proves the promise of "one interface for every agent" without the drawbacks of a compromised abstraction layer. Inter-agent delegation is the game-changer: parent agents spawn subagents, passing context seamlessly between them. My current loop: 1. Fable, my planning agent, breaks down feature specs into discrete steps; 2. Codex implements the core logic using computer-use tools; 3. OpenCode with local Qwen2.5-Coder reviews diffs for style and security issues; 4. Claude Code crafts integration tests and updates documentation. All four sessions share a persistent workspace, with clear handoffs — you review and approve outputs before passing them on. No hidden state, just explicit context passing. Workflows here mean reusable automation, not just linear scripts. You could even nest workflows within others, allowing for complex branching logic that adapts based on real-time feedback from the agents. This level of flexibility is unmatched and would be a game-changer for any development process.
So hyped for this open source move! I'm curious if you'll build a planning agent like Fable to break down feature specs into discrete steps. Which plugins are you planning to build first?