flompt
flompt is a free, open-source visual prompt builder for anyone who keeps rewriting the same prompt. It splits a prompt into typed blocks you edit on a canvas, then recompiles them as Claude-optimized XML. No account, no paid plan.
What is flompt?
flompt, a contraction of flow and prompt, describes itself as a Visual AI Prompt Builder, and its thesis is printed on the home page: structure beats a better model. The claim is that reorganizing what you already wrote lifts the output faster than reaching for a stronger model.
The loop is short. You paste a prompt, an AI pass decomposes it into typed blocks, you edit those blocks, and flompt recompiles them into XML. There are 16 block types, each mapped to a Claude XML tag: Document, Role, Tools, Audience, Context, Environment, Objective, Goal, Input, Constraints, Guardrails, Examples, Chain of Thought, Output Format, Response Style and Language. Blocks are automatically ordered following Anthropic's recommended prompt structure, so the assembly rule is not yours to remember. The editor offers two views you can switch between at any moment, a drag-and-drop flowchart canvas built on React Flow or a plain list of cards. Any block can be hidden, which excludes it from the assembled prompt, or duplicated in one click.
Three tools sit on top. Audit flags what is missing, such as no role or no examples; Score rates the compiled prompt from 0 to 100 with advice; Version History stores labeled snapshots you can restore or diff side by side. Context Memory holds per-project facts, your stack, your tone, your rules, and loads them into every new prompt. A library of 100+ templates covers ten categories, from code and writing to sales and education.
flompt exists on three surfaces: the web app, a Chrome and Firefox extension that opens a sidebar inside ChatGPT, Claude and Gemini, and an MCP server that plugs into Claude Code. It is free, requires no account, ships under the MIT license and can be self-hosted with Docker. Hosted decomposition runs through Anthropic or OpenAI; without an API key it falls back to a keyword heuristic that stays entirely on your machine. The interface is translated into ten languages.
Behind it is a solo project by Nyrok (Hamza Konte), with 146 stars and 16 forks on GitHub as of 23 August 2026, a Vite/React app, a FastAPI backend and a Caddy reverse proxy. There is no SLA.
What it does
- Decompose a raw prompt into typed blocks
- Edit those blocks on a drag-and-drop flowchart canvas or in list view
- Recompile everything as Claude-optimized XML, in Anthropic's canonical block order
- Audit a prompt for missing blocks and score it from 0 to 100
- Save labeled versions and compare two of them in a side-by-side diff
- Inject the finished prompt straight into ChatGPT, Claude or Gemini from the browser extension
- Call flompt from Claude Code over MCP, push a prompt to a Make.com scenario by webhook, or export it as TXT or JSON
When to use flompt / When not to
A quick filter to help you decide if flompt is the right fit.
When to use flompt
- Daily ChatGPT, Claude or Gemini users who rebuild the same prompt from scratch every time
- Claude Code users who want decompose_prompt and compile_prompt available as native MCP tools
- Teams that need a reproducible prompt format, built on 16 blocks in a canonical order
- Non-technical users who want a running start: 100+ ready-to-load templates and nothing to install
- Privacy-conscious users, who can self-host with Docker or stay on the fully local heuristic decomposer that needs no API key
When not to use flompt
- Organizations that require an SLA or an availability guarantee: this is a best-effort indie project
- Anyone expecting multi-device sync, since there is no account and work is persisted only in the browser
- Teams looking for a shared workspace or real-time collaboration, neither of which is documented
- Users whose target model is not Claude: the compiled output is XML shaped for Anthropic's guidance
- People who want the prompt executed, or a native mobile app: flompt builds the prompt, it does not generate the final answer, and it ships only as responsive web plus browser extensions
How to use flompt
A typical end-to-end flow, from setup to results.
- Open flompt.dev/app in your browser: no account and no installation are required
- Paste an existing prompt, or start from one of the 100+ templates
- Let the AI decomposition produce the blocks, or build them by hand
- Reorder, hide or duplicate blocks on the flowchart canvas or in list view
- Run Audit to spot what is missing, then Score to rate the compiled result
- Save a labeled version before making changes, so you can diff it later
- Copy the compiled XML, or export it as TXT or JSON
- For everyday chats, install the extension from the Chrome Web Store or addons.mozilla.org, then use the button it adds to the ChatGPT, Claude or Gemini toolbar: the sidebar opens inside the chat, picks up the input field and reinjects the result in one click
- For Claude Code, run claude mcp add --transport http --scope user flompt https://flompt.dev/mcp/ or add the equivalent mcpServers block to ~/.claude.json, then call decompose_prompt, compile_prompt or list_block_types
- To automate, paste a Make.com webhook URL into flompt and send the prompt (payload: prompt, format, blockCount, source, sentAt); to self-host, run the stack with Python 3.12+ and Node 18+, app on :5173, backend on :8000, MCP on :8000/mcp/, with an Anthropic or OpenAI key optional
Pros & Cons
Pros
- Free with no account: nothing stands between you and a first test
- MIT-licensed and self-hostable, so the code is public and the exit path is real
- An explicit prompt model, 16 named blocks, instead of one opaque text box
- Three surfaces covered: web app, browser extension and MCP inside the agent
- Output aligned with Anthropic's published prompt-structure guidance
- Audit and Score give a numeric read on prompt quality, which is rare in this category
- Privacy-lean by design: prompts are not retained after processing, there is no account data to collect, analytics are hosted in the EU region with Do Not Track honored, and the local heuristic fallback makes no external call
Cons
- A solo project with no SLA and no availability guarantee; the publisher reserves the right to discontinue the hosted service
- Very young: the domain was registered on 24 February 2026, so there is little track record
- No account means no sync across devices and nothing shared with collaborators
- In hosted mode, the prompt is sent to Anthropic or OpenAI for decomposition
- No DPA is published, and no training opt-out is documented anywhere
- The XML output is designed for Claude; other models are not what the format targets
- Rough edges around the product: /docs is a JavaScript-rendered GitBook proxy whose navigation link returns 404 without a trailing slash, Google Analytics loads on the home page although the privacy policy mentions only PostHog, there is no native mobile app, and the publisher is identified by a pseudonym with no registered company name or postal address
Pricing & Plans
flompt is free of charge. There is no paid plan, no pricing page and no payment is ever collected: the home page states Free for everyone. No account required., the project is distributed under the MIT license and advertised as free forever, and the site's structured data declares an offer price of 0 USD. No account is needed to use the hosted application, and there is consequently no lowest paid price point to report. The only indirect cost arises in self-hosting, and it is optional: supplying your own Anthropic or OpenAI API key means that provider bills you for the calls, whereas the built-in local heuristic decomposer runs without any key and at no cost.
- the single access tier
- covering the web app
- the browser extension and the MCP server
- with no account required
- no paid upgrade and no commercial plan naming on the site
Data, GDPR & hosting
A consolidated view of how flompt handles your data.
GDPR overview
GDPR is addressed directly. The privacy policy, effective 25 February 2026, lists the six data-subject rights (access, rectification, erasure, restriction, portability and objection), names the legal bases used, legitimate interest for analytics and performance of the requested service for prompt processing, and points to the CNIL as supervisory authority; French law governs the terms. Transfers to Anthropic and OpenAI outside the EEA rely on standard contractual clauses under Article 46. In practice, exercising your rights means asking for your PostHog analytics to be deleted, via contact@flompt.dev. Do Not Track is honored by PostHog, and a content blocker stops analytics entirely. Minimum age is 13, or 16 where the GDPR threshold applies. Two gaps: no DPA is published or mentioned, and no Article 27 representative is designated, which is expected since the publisher is established in France.
Who owns the data?
flompt's terms are explicit: the prompts you write are yours, and the publisher claims no ownership over them (Terms, section 4). In exchange, you confirm that you have the right to share what you submit, because a hosted decomposition sends that prompt to a third-party AI provider, Anthropic or OpenAI, whose own terms then also apply to it. Nothing else about you is claimed either: the extension keeps its sidebar preferences in chrome.storage.local, on your device, and never sends them to flompt's servers. Ownership is therefore uncomplicated; the question worth asking is what you hand over to the model provider.
Reuse rights
Because you keep ownership of your prompts, you are free to reuse, publish or sell whatever you build in flompt without asking permission. On the publisher's side, collection is deliberately thin. Anonymized product analytics go to PostHog in its EU region (pages viewed, features used, session length), on a legitimate-interest basis. A prompt sent for AI decomposition travels to the backend and on to Anthropic or OpenAI purely to perform the service you requested; the content is held in memory only long enough to return the result, then discarded. What is never collected: no email address, no name, no user account, no payment information, and no browsing history beyond those analytics. There is no advertising and no user profiling. If you would rather send nothing outside, the local keyword-based heuristic decomposer works without an API key and makes no external call.
Data retention & training
Hosting summary
Data lands in three places, depending on what it is. Product analytics go to PostHog in its EU region, hosted in Frankfurt (eu.i.posthog.com). Prompts sent for hosted AI decomposition go to Anthropic, with OpenAI as fallback, both in the United States; those transfers outside the EEA rely on standard contractual clauses under GDPR Article 46, and the content is not kept once the job is done. The extension's own settings, whether the sidebar is open or closed, stay in local browser storage on your device and are never sent to a server. On the network side, measured rather than announced by the site, flompt.dev resolves to an address in Paris, France, on AS47583 (Hostinger International Limited). Traffic runs over HTTPS with HSTS, and the reverse proxy sets a Content Security Policy and X-Frame-Options. If none of this fits your jurisdiction requirements, the MIT license allows a self-hosted deployment on infrastructure you choose, in which case nothing leaves it unless you supply a provider API key.
Things to keep in mind
Risks and trade-offs to weigh before adopting flompt.
- Hosted AI decomposition sends your prompt to Anthropic or OpenAI: do not paste confidential, personal or client data into it without checking those providers' policies first
- No DPA is available; in a regulated or contractual setting, that gap has to be assessed before real work goes through the hosted service
- Nothing in the privacy policy addresses model training, and no opt-out is documented, so the absence of a statement should not be read as a guarantee
- There is no server-side backup: your work lives in your browser's local storage, so clearing browsing data or switching machines loses it
- No SLA, the service may be modified or discontinued, and access can be blocked without notice if abuse is detected, so keep exports of anything you depend on
- Two rough edges observed on 23 August 2026: Google Analytics loads on the home page although the privacy policy lists only PostHog, and the /docs navigation link returns a 404 without its trailing slash; the publisher is also identified only by a pseudonym, with no registered company name or postal address
- A structural score can become a crutch: a high rating measures the shape of your prompt, not whether the instruction itself is right, and delegating that judgement long enough dulls the very skill the tool is meant to build
Setup & Integrations
Technical difficulty
It depends entirely on the surface. Web app: none, open flompt.dev/app, no account, no install. Extension: one click from the Chrome Web Store or addons.mozilla.org. Claude Code over MCP: one CLI command or four lines of JSON, with a hosted server that needs no authentication, so nothing is installed locally. Make.com: paste a webhook URL. Self-hosting is the only developer-level path, with Python 3.12+, Node 18+, a virtual environment and npm, an optional provider API key, and Caddy plus supervisord in production. Most users never leave the first two levels.
Deployment
Integrations
Supported languages
Behind flompt
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
Is flompt free?
Do I need to create an account?
Which models and platforms does it cover?
Are my prompts stored?
Can I use flompt without sending my prompts to a third party?
How do I add flompt to Claude Code?
Which MCP tools are exposed?
How many block types are there?
Is there a mobile app?
Who publishes flompt?
Should you pick flompt?
flompt is a well-executed niche tool. Its whole idea is to make the structure of a prompt explicit, sixteen named blocks in a fixed order, rather than leaving it implicit in a paragraph you rewrite each time. That is a narrow ambition, and it is met.
The adoption risk is close to zero. It costs nothing, asks for no account, and the MIT license plus a documented Docker install mean you can take the whole thing in-house if you ever need to. Few tools in this category let you evaluate them with so little at stake.
The counterweight is maturity. This is a solo project whose domain was registered in February 2026, with no SLA and no commitment to keeping the hosted service running. The open code softens that considerably, since self-hosting remains available if flompt.dev stops answering, but it does not replace a support contract, and there is no legal entity or postal address behind the pseudonym.
The natural audience is people who already spend their day inside an LLM: daily ChatGPT, Claude or Gemini users, who get the browser sidebar, and Claude Code users, who get decompose_prompt and compile_prompt as native tools. If you write a prompt once and move on, the block model will mostly feel like overhead.
One reservation to settle before any sensitive use: in hosted mode the prompt you decompose is sent to Anthropic or OpenAI, no DPA is published, and nothing in the privacy policy addresses model training. For confidential material, the local heuristic decomposer or a self-hosted instance is the honest answer, not the hosted app. Test it on a prompt you rewrite often, and judge the compiled output on its own merits.
- Choosing a selection results in a full page refresh.
- Opens in a new window.