Anvil - Run a fleet of parallel Claude Codes
by•
The IDE for parallel agent work. MIT licensed, farm to table, fun.
With one click git worktree isolation, first class plan tracking, color coding for agent states, flexible layout arrangement, and so much more, Anvil is crafted to make developers extremely productive, minimizing context switching and maximizing agent parallelism.
Replies
Feels like dev tools are moving from “AI assistant”
to “AI team”.
Parallel agents sounds powerful
but also messy if not managed well.
Curious how you keep control over multiple agents working at once.
@new_user___2902025abb5753b18b341a5 a great plan and testing suite. Is super important, I rarely kick off agents fully blind
Is this forked from the VS Code as I am getting kinda similar vibe here. But the feature is really helpful considering running multiple claudes at a time and managing outputs becomes bit overwhelming.
@nayan_surya98 this is built from scratch in rust and typescript! No fork here
Product Hunt
@curiouskitty this is a great question, I don't see the two (orchestration and verification) as mutually exclusive, but if I had to choose one I'd pick verification. Anvil starts with orchestration features because in order to get to verified results it often takes more work than one agent's context can handle.
For instance, I'll often use the anvil repl to loop claude until it can fully self verify, something that in one context window (even with compaction) would go awry.
this looks super fun, congrats! the REPL thing where Claude can call other Claudes programmatically sounds wild, any examples of what people are actually using that for?
@jens_deryckere1 The most practical use case is calling claude in a loop to execute against a larger task or plan file. There is a skill /breadcrumb-loop which allows you to do this. Whats nice about this skill is it compacts when the agent hits 100k tokens in its context window, and leaves a "breadcrumb" markdown file for the next agent to read and continue with.
Other things I've seen are people orchestrating entire workflows like compound https://github.com/EveryInc/compound-engineering-plugin
Instead of running each step manually they will script the entire thing. I can't say I'd recommend this second one bc I think human in the loop is important for the planning phase to get things right and make sure the architecture is sound
How are you handling state isolation between parallel agents? When you run multiple Claude Codes on the same codebase, conflicts in shared files are the main issue I've hit. Do they each get their own branch or working copy?
@mykola_kondratiuk Anvil makes using git worktrees very simple. This creates an isolated copy of your codebase for agents to run in. Worktrees are different than branches because you can have multiple checked out at the same time (its a full copy)
Clipboard Canvas v2.0
As someone who's wrestled with the chaos of parallel coding, I’m curious how do you ensure effective collaboration between agents without conflicts in shared file states?
@trydoff Two approaches
1. If the two features are big or divergent enough, you can create multiple work trees for each agent, this gives them full isolation
2. If the agents should collaborate on the same branch, I usually coordinate through a plan file. This is a simple .md file that details the phases of what I'm trying to do sometimes with dependency mapping. From there agents can coordinate and tackle isolated components of the plan.
Claude is really good at creating the dependency graph of what needs to get done and what is parallellizable.
I've been running 3-4 Claude Code sessions in tmux and losing track of which agent is doing what is a daily struggle. The git worktree isolation alone would save me from merge conflict hell. Does Anvil handle cross-agent dependency — e.g., Agent B waits for Agent A's PR before starting?
@ilya_lee the way I would typically do this kind of thing is by writing a custom orchestration skill, and yes anvil is really great for this kind of thing.
You can start by creating a plan file initially which details all the PRs you want to create, then have a parent agent manage the different phases. These parent agents can actually create worktrees themselves should you need.
I'd be curious though what your typical use case is?
Is it possible to work with two or more licenses, for example Claude Code? If the limit on one is reached, can you switch to the second one without losing the project context?
Yo dude, I'm reaching out for a comment.
I’m a researcher for the H1Gallery newsletters (you can google us).
We’re featuring Anvil in our April 3 issue. H1Gallery highlights excellent homepage headlines, and we wanted to reach out to see if you wanted to share a quick comment. The clarity and directness in "The open source IDE for parallel agent work." is very refreshing. The animation is super cool as well. Love it
If you’re up for it, we’d love to get a short note on the copywriting strategy behind your headline and overall messaging. As well as where our readers can follow you on social media, I couldn't find an X account or Linkedin account of yours.
Totally optional to provide a quote for us. The feature is happening regardless, but it’d be great to include your perspective. It makes the newsletter more fun for our readers when we have comments from the creators behind the headlines.
Thanks, really appreciate your time.