An editor where the spec is the contract, not the chat
DevFlow is a desktop app where AI agents read and write your real code — but instead of improvising from a loose request, they follow a six-step cycle with human review at every step. Everything you see here is a real screenshot of the app, not a mockup.
You just talk — the agent is already thinking in specs
Given a feature request, chat doesn't jump straight to code: it proposes a
Given/When/Then scenario and a diagram first — editable windows before
committing to anything. The right-hand panel shows the live technical trace: tool
calls, duration, and the turn's real token/cost usage.
- Send to Specs → hands the conversation to the execution engine without reprocessing what it already covers.
- Meta-prompt (wand icon) rewrites a vague ask with Role/Objective/Context/Restrictions before you send it.
- Sessions per tab — different agents, running in parallel, never colliding.
Six phases, a human checkpoint in every one
Every phase writes a real, reviewable file —
enriched.md, requirements.md, design.md,
tasks.md, review.md — before the next one starts.
Run all chains whatever is left and stops only if something fails to advance.
Review compares the real diff against the spec's own acceptance criteria and adds a security pass (secrets, string-built SQL, endpoints missing auth). Archive makes the spec immutable and distills what it learned into the project's memory — the next spec starts already knowing what this one solved.
The agent experiments, your project never finds out
Every environment is a real git worktree with its own branch — local or on a remote machine over SSH. Specs uses it automatically when "Isolate" is checked: code gets written there, not in the active project, until you decide to promote the diff or discard it.
Fan out sends the same prompt to several agents in parallel, each in its own worktree, so you can compare results side by side before choosing.
What doesn’t need an agent, you chain visually
A graph editor to automate processes with or without AI in the loop: an AI task, an expert agent, QA verification with a verdict, an HTTP call, an MCP tool, a terminal command, a condition, a loop/foreach — even a fully nested sub-workflow.
They trigger on a webhook or a cron schedule — nodes on the same agent reuse the session, so they share context between steps.
⌘K jumps anywhere
Files, workflows, agents, projects, views — the Command Palette indexes all of them. You never have to leave the keyboard to move around the app.
Free, local, or your own key — all three at once
DevFlow Code ships with free models from the first launch. Want more? Add Ollama/LM Studio (local) or your own key for Anthropic, OpenRouter, Gemini, Groq or Mistral — they’re not mutually exclusive, you can use all of them at once, chosen per agent or per task.
PRs and issues, without leaving the app
Via the already-installed/authenticated gh CLI — no tokens handled
by hand. "Open as environment" checks out someone else’s PR into an isolated
worktree for review, and Specs can create the PR directly once a feature is done.
This is already running
Every screenshot on this page was taken live from a real DevFlow session while this demo was being put together. Full detail on every feature is in the user manual.