Flash Agents

Flash Agents is an MIT-licensed Claude Code plugin by Thomas Wiesner (tomw1808). It hands bounded coding work such as implementing a slice, porting tests, reviewing a diff or mapping a codebase to DeepSeek V4.1 Flash workers instead of Claude subagents. Claude keeps the architecture, acceptance criteria and final review.

What It Is

Flash Agents is an MCP server and Claude Code plugin that splits coding work between a steering model and cheaper worker models. Claude frames and slices the task, then dispatches it through the flash_task or flash_batch tools. Each writing job runs in a disposable copy of the repository and comes back as a patch with a diffstat computed by git, so nothing touches the working tree until Claude or the user applies it.

How the Delegation Loop Works

The /flash-agents:delegate skill encodes the everyday loop: frame, slice, dispatch, verify, review, fix and apply. flash_batch runs up to 16 independent tasks in one call, four at a time, and returns results in task order. flash_apply lands a patch from an earlier call, with a dry-run option. A gauntlet skill runs several competing builders and cheap critics, with a blind A/B judged by Claude. Read-only mode is available for reviews and codebase maps, running in a process whose sandbox refuses writes.

Safety Model

The README describes three layers: the copy-on-write clone, a sandbox provided by DeepSeek Harness, and a deterministic model-free filter. The filter refuses reads of secrets, writes to .git, destructive git commands, catastrophic deletions and writes outside the workspace.

Setup Path and Limits

It requires Claude Code, Node 20+ and Ollama. Setup installs @deepseek-ai/dsh, signs in to Ollama and pulls the deepseek-v4.1-flash:cloud model, then runs /flash-agents:setup to check the environment and run a smoke test. Documented limits include that a running worker cannot be steered or cancelled, that members of one batch share a copy, and that gitignored output is not carried in the patch. macOS and Linux are supported; Windows is untested. The server speaks plain MCP over stdio, so other MCP clients can use it.