# Revalvo

> A local-first prompt workbench that runs the same prompt against multiple AI models simultaneously, scores results in real time, and versions prompts like code — all in your browser with your own API keys.

Revalvo is a local-first prompt evaluation workbench that runs entirely in your browser, using your own API keys with no Revalvo account required. It lets you write a prompt once and dispatch it to multiple models in parallel, score the outputs automatically, and keep a full version history — treating prompts like code with diffs and rollback. The tool positions itself between lightweight chat playgrounds (fast but no rigor) and hosted eval platforms (rigorous but slow to set up and server-side).

## What It Is

Revalvo is a browser-based prompt engineering and evaluation tool designed for developers and prompt engineers who need to compare model outputs systematically before shipping. It covers four core steps: connecting providers, running prompts in a side-by-side playground, versioning saved prompts with GitHub-style diffs, and batch-testing against datasets with 40 built-in evaluators. All data — keys, prompts, run history — lives in IndexedDB in the user's browser and is never uploaded to Revalvo's servers.

## Local-First Architecture

The "0 Servers of ours" tagline reflects a deliberate design choice: Revalvo has no hosted database, no user accounts, and no markup on API calls. Keys are stored in IndexedDB only. The tool proxies direct API calls for providers like OpenAI and Anthropic only to work around browser CORS restrictions, not to log or store data. Multiple workspaces per browser install each scope their own prompts, datasets, keys, and run history. A GitHub sync feature lets users push and pull prompt recipes as plain YAML bundles via a fine-grained token.

## Provider and Model Coverage

Revalvo supports 12+ providers across three categories:
- **Aggregators**: OpenRouter (300+ models, OAuth sign-in), Ofox.AI, Vercel AI Gateway
- **Direct APIs**: OpenAI, Anthropic, Groq
- **Local / offline**: Ollama (localhost:11434), LM Studio (localhost:1234), plus any OpenAI-compatible custom endpoint

This means users can run fully offline against local models or mix cloud providers in the same comparison run.

## Playground and Versioning Workflow

The playground dispatches a single prompt to all selected models in parallel, rendering outputs in synced columns with per-result latency, token count, and cost metrics. Parameters (temperature, max tokens, top P, frequency penalty, reasoning effort) are tunable per model. Every run is stored in run history and can be pinned. The `{{variable}}` syntax enables prompt templating, and 14 built-in prompt templates cover common use cases like onboarding copy, support replies, and release notes.

Saving a run creates an immutable, scored snapshot. Version history supports GitHub-style word diffs and model config diffs between any two versions. Any version can be restored as the active one. Saved versions can be exported as API code snippets (Python, TypeScript, cURL, Node.js, Go) or formatted as agent prompts for Cursor, Copilot, ChatGPT, or Claude.

## Batch Evaluation and Reporting

The batch eval layer runs a saved prompt version against a full dataset — built by hand, uploaded as CSV, or AI-generated from a description using 12 templates. Forty evaluators span four categories:
- **Structure**: exact match, contains, regex, JSON schema, Levenshtein, refusal detection
- **AI judgment**: semantic similarity, faithfulness, hallucination, answer relevance, pairwise judge, factuality
- **Performance**: latency SLA thresholds, cost-per-run limits
- **Safety & advanced**: PII detection, toxicity, jailbreak, bias, JavaScript/webhook/tool-calling validators

Each batch run produces a structured report with pass rate, latency, cost per model, model ranking, per-case pass/fail, and a debug console, exportable as HTML, Markdown, JSON, or CSV.

## Audience and Positioning

Revalvo targets prompt engineers and developers who want evaluation rigor without the setup overhead of a full hosted eval platform. The homepage explicitly contrasts it with chat playgrounds (no rigor) and hosted eval platforms (no speed, server-side data lock-in). The tool is free and described as "local-first," with no signup required to run a first evaluation.

## Features
- Multi-model parallel prompt execution
- 40 built-in evaluators (structure, AI judgment, performance, safety)
- Local-first: all data in browser IndexedDB, no server storage
- Bring your own API keys, no markup
- Git-style prompt versioning with diffs and rollback
- Batch evaluation against datasets (CSV upload, AI-generated, or manual)
- Side-by-side model output comparison
- Per-result latency, token, and cost metrics
- Variable templating with {{variable}} syntax
- 14 prompt templates built-in
- GitHub sync via YAML bundles
- Multiple workspaces per browser install
- Monthly budget cap and usage breakdown
- Export as API code snippets (Python, TypeScript, cURL, Node.js, Go)
- Export as agent prompts for Cursor, Copilot, ChatGPT, Claude
- Voice dictation for prompts (browser-native speech-to-text)
- Batch report export (HTML, Markdown, JSON, CSV)
- Support for local/offline models via Ollama and LM Studio
- OpenAI-compatible custom endpoint support
- Abort in-flight runs

## Integrations
OpenRouter, Ofox.AI, Vercel AI Gateway, OpenAI, Anthropic, Groq, Ollama, LM Studio, GitHub, Cursor, GitHub Copilot, ChatGPT, Claude

## Platforms
WEB, API

## Pricing
Open Source

## Links
- Website: https://revalvo.com
- Documentation: https://revalvo.com/help
- Repository: https://github.com/revalvo/revalvo
- EveryDev.ai: https://www.everydev.ai/tools/revalvo