Claude Coderefactor the implementation, and Grok review the PR — all without leaving a single terminal — I stopped paying for three separate subscriptions. That was two weeks ago. I've been running Proliferate on a home server since, and it's the first tool that actually delivers on the "one interface for every agent" promise without feeling like a compromised wrapper.
Proliferate (GitHub: proliferate-ai/proliferate) is an open-source, self-hostable AI IDE from YC S25 that lets you drive Claude Code, OpenAI Codex, OpenCode, Cursor, and Grok from the same UI. Each agent keeps its native inference path — Bedrock, Azure, local vLLM, whatever you configure — so you're not locked into a lowest-common-denominator abstraction layer. The AGPL-3.0 license means you can audit, fork, or run it behind your own VPN without asking permission.
Inter-agent delegation that works #
The feature that changed my workflow: a parent agent can spawn subagents and pass context between them. My current loop:
-
Fable (my planning agent) breaks down a feature spec into discrete tasks
-
Codex implements the core logic with its computer-use tooling
-
OpenCode with a local Qwen2.5-Coder model reviews the diff for style and security issues
-
** Claude Code** writes the integration tests and updates documentation
All four sessions share a persistent workspace. The handoff is explicit — you see each agent's output, approve or reject, then pass the baton. No hidden state, no "trust me bro" black boxes.
steps:
- agent: codex
task: "implement user-auth module per SPEC.md"
model: gpt-4o
- agent: opencode
task: "review diff for OWASP top 10 violations"
model: qwen2.5-coder:32b
approval_gate: true
- agent: claude-code
task: "write integration tests for auth flow"
model: claude-3-5-sonnet
Workflows = reusable automation, not chat history #
Workflows are where Proliferate separates from "just another chat UI." You define a chain of agent sessions with human approval gates, document passing, and per-step model selection. I built one for our PR process: spec ingestion → implementation → security review → test generation → changelog draft → final human sign-off. Runs in about 12 minutes end-to-end. The YAML is version-controlled alongside the repo, so the automation evolves with the codebase.
Rough edges worth knowing #
No Windows support yet— Docker on WSL2 works but expect path-mapping friction** Agent config UI is minimal**— you'll edit JSON for Bedrock/Azure credentials** Streaming output occasionally desyncsbetween the frontend and long-running Codex sessions Resource usage**— running three agents concurrently needs ~24 GB RAM on my box
The team ships weekly. Rough spots get patched fast.
If you're evaluating AI workflow tooling for a team that refuses vendor lock-in, spin up the Docker Compose stack and throw a real task at it. The self-hosted story is genuine — no telemetry, no external calls unless you configure them.
Built a shared brain for AI agents after markdown file chaos 1h ago
Built a schedule-aware PM copilot that actually respects 1d ago
Built a remote control for the AI agents living on my garage 2d ago
Does anyone actually care about llms. 3d ago
AI is making every side project feel like meaningless slop 3d ago
HarnessRouter makes it way easier to switch between Claude Code 4d ago
Next Why the current alignment toolkit still leaks — and what →