What it does

Coding agents are brilliant for ten minutes, then they start guessing. They cannot see the four hundred functions already in your repository, so they write a fifth version of the one you needed, in a style that matches nothing, and confidently tell you it is done.

FlowKy gives the agent the context it is missing.

• Be SOTA — a library of 46 engineering standards (prompt security, login flows, SEO, migrations, technical debt, MCP, and more), written as prompts your own agent executes: verify the repo rule by rule with file-level evidence, make it conform, or build the area right from the first line. Ctrl+Alt+S, or the Be SOTA row in the panel.
• Project kickoff — open an empty folder, describe the tool in one sentence, and your agent scaffolds it to online: stack chosen with reasoning, deploy and CI in the first slice, a landing page with real copy. The kickoff bundle is free on every connected account.
• Project map — a symbol index of what already exists, so a suggestion reuses your code instead of duplicating it.
• Duplicate guard — catches the moment the agent reimplements something you already have, and says which file to use instead.
• QuickFix — reads your real diagnostics and fixes them, across files, with the errors as evidence rather than a guess.
• Health score — an honest read on the state of a codebase: what is stale, what is duplicated, what has no tests.
• Claude configuration generator — writes CLAUDE.md, project rules and a project-map skill from what your repository actually contains.
• Anti-loop circuit breaker — after three identical failed commands, it stops and shows you the last error instead of burning your budget retrying.
• Bring your own key — Google, OpenAI, Anthropic and any OpenAI-compatible endpoint. Your key, your account, your usage.

Getting started

1. Install the extension.
2. Open the FlowKy panel from the activity bar.
3. Connect to your free flowky.ai account, or set an API key directly with FlowKy: Set AI Provider API Key.
4. Run FlowKy: Analyze Current Project once, so it knows what you have.

Everything is in the command palette under FlowKy.

Configuration

The two settings that matter most, and they must agree with each other:

If you switch provider, switch the models too — a Gemini model name will not validate against OpenAI, and every AI call fails until they match.

The rest is under Extension Settings, grouped by area.

Privacy

• Your code is not sent anywhere except your chosen AI provider, using your own key.
• The project map is built locally and stored in your workspace.
• Repository analysis reads metadata — file names, symbols, languages — not the contents of files it has no reason to open.
• Connecting your flowky.ai account exchanges a signed token for your plan and quota. It does not grant access to your code.

Requirements

VS Code 1.88 or later, Node 18 or later.

Links

flowky.ai · Report an issue · Changelog

Be SOTA: standards your agent enforces

Every playbook is a set of numbered rules — the do, the don't, and what breaking it costs — served by flowky.ai and written into your repository at .flowky/sota/<id>.md, where it is yours to edit and version. Pick one (Ctrl+Alt+S), pick the mode:

1. Verify conformance — measure only. Per rule: pass / fail / not-applicable / unknown, each with the files it was judged from. An unknown is an honest answer; a guessed pass is not allowed.
2. Verify and make conform — fix every failing rule, re-verify, show the evidence.
3. Integrate the SOTA way — the area does not exist yet; the agent plans the smallest complete slice and builds it conforming from the start.

Reports land in .flowky/sota/reports/, the panel shows each standard's standing, and when your agent learns a codebase-specific lesson it writes the lesson back into the playbook — the library improves as you work. The standards whose domain FlowKy detects in your repository lead the picker.

The catalog is browsable at flowky.ai/sota. The project-kickoff bundle is free; the full library is part of the Pro plan.

Fixing tasks on your plan, not on tokens

FlowKy can talk to Gemini, OpenAI and Anthropic with a key you paste in. Task resolution deliberately does not use that path.

Most people running this already pay for Claude Code, Cursor or Copilot. Sending the same task to an API on top of that subscription bills the work twice, and the second time at the worse rate. So FlowKy: Fix with your coding agent runs your agent's own CLI, in a terminal you can see, against the task files the sync wrote:

claude 'Read `.flowky/TASKS.md` and do what it says. …'

Detected automatically: claude, cursor-agent, codex, opencode, aider, gemini. FlowKy sends nothing to a model and holds no key for this — the token cost of the whole path is zero.

The trap, and the warning

claude bills your subscription when it is logged in with your account. If ANTHROPIC_API_KEY is set in the environment it silently uses the API instead, and you are back to per-token billing with nothing saying so — the difference is invisible until the invoice. FlowKy checks for exactly that before running and names the variable responsible.

Order of preference

1. The agent CLI, on your plan. Always tried first.
2. A Claude Code conversation already open — the task is @-mentioned into it, so the work joins what you are doing.
3. A fresh Claude Code tab, prompt pre-filled.
4. Clipboard, with a message saying that is what happened. A silent no-op reads as a broken feature.

The metered API path is not in that list at all.

Slash commands

FlowKy: Add /flowky slash commands writes .claude/commands/, so you can type /flowky-fix critical in the composer and your own agent works from the same files. Written only on request — putting files in your repository should never be a side effect of pressing Sync.

Nothing runs unattended

The command is written into a visible terminal. Whether it presses Enter for you is flowky.agent.autoRun, off by default, and it never runs on a timer. An agent let loose on a repository unsupervised is an expensive way to be surprised.