I've just released vincent 0.14.0, and it's one of the most significant updates to the project so far.
For those unfamiliar with it, vincent is an open-source, local-first AI SDLC orchestration tool built around a simple idea: AI-assisted development needs more than a chat interface.
It needs visibility, control, reproducible workflows, and the ability to work with different AI coding agents without being locked into a single provider.
This release takes a major step in that direction.
First, see it in action
One of my favorite additions is the ability to run two different AI agents against the same prompt, compare their results side by side, and choose which one to keep.
Here's a demonstration using Claude Code and Codex:
Instead of guessing which agent would perform better, you can actually compare their approaches and results.
And this is only one of the new capabilities in 0.14.0.
A completely redesigned terminal UI
The biggest visible change is a complete overhaul of the TUI.
I wanted something that feels like a proper development workspace rather than a collection of terminal screens.
The redesign includes:
- Card-based boards for tasks and issues, with previews and clearer status information.
- A redesigned task workspace with a step timeline, unified output feed, improved diffs, and contextual overview cards.
- A dedicated chats tab for managing interactive agent sessions.
- Simplified task and chat creation through streamlined forms.
- Improved navigation across projects, workflows, pull requests, triggers, and settings.
- Optional Nerd Font icons for a richer terminal experience.
The goal wasn't just to make vincent look better. It was to make managing multiple concurrent AI workloads significantly easier.
AI conversations you can rewind, fork, and compare
This is probably the most exciting technical addition.
Every chat turn now creates a checkpoint of its working tree.
That means you can:
- Rewind a conversation to any previous turn.
- Restore the code, the conversation, or both.
- Move forward again after rewinding.
- Fork a conversation onto a separate branch.
- Hand the conversation over to a different AI agent.
- Compare two agents responding to the same prompt and keep the better result.
Imagine trying one implementation with Claude Code, exploring an alternative with Codex, and keeping whichever works better.
No manually reconstructing context. No copying conversations between terminals.
Just a controlled development workflow.
Review AI-generated changes with precision
Another important improvement is hunk-by-hunk code review.
You can now inspect the changes produced by an agent, including untracked files, and accept or reject individual hunks.
Rejecting a hunk reverts that change and tells the agent not to reapply it.
This addresses something I've always considered essential in AI-assisted development: the AI can write the code, but the developer must remain in control of what gets accepted.
Better conversations and less friction
There are plenty of improvements to everyday development workflows, too.
You can now:
- Queue messages while an agent is working.
- Change the model, reasoning effort, or permission mode between turns.
- Execute shell commands directly from a chat.
- Paste images for agents that support them.
- Search conversations, including archived chats.
- Preserve unfinished drafts across restarts.
- Track context usage, token counts, and costs.
- Export conversations as Markdown, HTML, or JSON.
- Switch between chats without returning to the chat list.
There are also improvements to daemon controls, automatic task archiving, Jira integration, reliability, and recovery.
You can find the complete list in the 0.14.0 changelog.
Building vincent with vincent
Something I find particularly rewarding about this project is that I use vincent every day to build vincent itself.
It's not a hypothetical tool designed around imagined development workflows.
As I work on the project, I encounter limitations, identify opportunities, and implement improvements that directly affect my own daily development experience.
That feedback loop has been invaluable.
My long-term goal is to make vincent a comprehensive, vendor-independent environment for AI-assisted software development, from managing issues and orchestrating agents to reviewing, verifying, and delivering code.
And I want developers to have the freedom to choose their own tools.
Today, vincent supports Claude Code, Codex, Cursor, and GitHub Copilot, with reusable workflows, Git worktree isolation, a durable background daemon, and human approval controls.
Give it a try
vincent is open source, MIT licensed, and available for Windows, macOS, and Linux.
Documentation and getting started:
https://lezli01.is-a.dev/vincent/latest/
GitHub repository:
https://github.com/lezli01/vincent
Full changelog:
https://lezli01.is-a.dev/vincent/latest/changelog.html
If you're working with multiple AI coding agents, or you're looking for more structure and control over your AI development workflows, I'd love for you to try it.
If you find the project useful, consider giving it a star on GitHub. It genuinely helps an independent open-source project gain visibility.
And most importantly, I'd love to hear your thoughts.
What's the biggest frustration you encounter when developing software with AI agents today?
I'm actively shaping vincent around real development challenges, and your feedback could influence what comes next.


