DEV Community: AgentOne

AgentOne 1.0 Is Here

BestCodes — Mon, 01 Jun 2026 21:06:57 +0000

AgentOne is finally out of beta.

With the 1.0.0 release, AgentOne has reached a milestone I have been working toward for a long time: a stable, everyday AI workspace that feels fast, flexible, and reliable enough to recommend as more than an experiment.

This does not mean AgentOne is finished. It means the foundation is solid. The core workflow is ready, the roughest edges have been smoothed out, and the app is ready to move from "try this beta" to "use this as your daily AI workspace."

What AgentOne Is

AgentOne is a desktop AI agent app built around a simple idea: you should be able to use the best models, tools, and providers without having to rebuild your workflow every time the AI ecosystem changes, and you should be able to do this easily without any technical knowledge.

Instead of locking you into one provider or one kind of model, AgentOne is designed to be flexible. You can use hosted models, local or OpenAI-compatible providers, text generation, voice tools, and MCP-powered extensions from one place. The goal is not to be another chat box. The goal is to be a useful AI workspace.

What 1.0 Means

The beta phase was about proving the direction: could AgentOne be useful across different providers, workflows, and model types without becoming complicated? Could non-technical users figure it out easily? Could AgentOne outperform other AI agents for a lower price?

The 1.0 release is about stability and confidence. It means the app has matured enough that the main experience is no longer experimental.

In practice, 1.0 focuses on:

A more stable chat experience for everyday work.
Better model and provider support by adding more providers and models to the AI Model Directory.
Cloud sync for chats so you can access your chats from anywhere.
Better MCP support for connecting AgentOne to external tools and services (AgentOne already supports MCP servers, but MCP Apps support is coming soon)
Voice and media workflows through providers like OpenAI, ElevenLabs, LMNT, and Hume.

Most importantly, 1.0 gives AgentOne a stable baseline to build on.

Why This Matters

The AI tooling world changes constantly. New models appear, pricing changes, provider APIs shift, and entire categories of tools become useful or obsolete almost overnight.

That is exactly why AgentOne exists.

I do not want an AI app that only works well with one provider. I want a workspace where I can compare models, switch providers, use local options when they make sense, easily connect external tools, and keep moving without rewriting my setup.

AgentOne 1.0 is the first release that really feels like that vision is ready.

From Beta to Daily Driver

During beta, AgentOne changed quickly. Features moved around, provider support expanded, and the app evolved as real workflows exposed what mattered and what did not.

Getting to 1.0 meant tightening the basics: chats should load and save reliably, model selection should be fast, provider configuration should be flexible, the desktop app should feel native enough to leave open all day, and advanced features should not get in the way of the simple act of asking a model for help.

That last point is important. AgentOne can connect to a lot of things, but the core experience still needs to feel simple.

What's Next

Now that AgentOne is out of beta, development can shift from proving the product to expanding it.

The next phase is about making AgentOne more capable without making it harder to use: better agent workflows, deeper tool integrations, stronger model discovery, improved local-provider support, and more polish across the desktop experience.

There is still a lot to build, but 1.0 is the line in the sand. AgentOne is no longer just a beta project. It is a real app with a stable foundation.

Thank You

If you tried AgentOne during beta, gave feedback, tested weird provider setups, reported bugs, or just followed along, thank you.

AgentOne 1.0.0 is here, and I am excited for what comes next.

You can try it at agent-one.dev.

4 Best AI TTS APIs in 2026

BestCodes — Thu, 28 May 2026 17:24:18 +0000

Text-to-speech has gotten good enough that it is no longer just an accessibility feature or a novelty. If you are building an AI app, voice agent, audiobook tool, customer support bot, or content workflow, your TTS provider now has a huge impact on how polished the final product feels.

In this post, I am comparing four of the best TTS services worth considering right now:

Skip to the rankings

How I Am Comparing Them

For this comparison, I want to look beyond just "which one sounds best." Voice quality matters, but the best TTS service depends on what you are building.

I'll be measuring these criteria:

Overall voice quality: How natural, clear, and human the generated speech sounds.
Latency: How quickly audio starts and finishes generating.
Customization: How easy it is to customize the voice and speech style.
Pricing: Cost per character, token, minute, or request, plus how predictable it is at scale.

For each criterion, I'll assign a score from 1 to 5, with 5 being the best. I'll be testing the TTS models on their native platforms and in AgentOne.

1. OpenAI

I tested OpenAI's latest TTS model, GPT-4o mini TTS, on their official site, https://www.openai.fm/. There are 13 voices available.
This TTS model works with an input, which is the text you want to convert to speech, and instructions, where you can tell the model how to speak it. Instructions are very useful for customizing how the voice sounds - for example, you can tell the model to speak faster or slower, or express a certain emotion or tone.

Here are a couple of samples:

Sarcastic (Male):

You can find the samples in the original blog post.

Actioneer (Female):

You can find the samples in the original blog post.

As far as latency, it does take a bit longer to generate audio compared to other services. Pricing details are a bit confusing. It looks like the model costs $12.00 per 1 million output audio tokens, with text input costing $0.60 per token. You can find the details here:
https://developers.openai.com/api/docs/pricing

2. ElevenLabs

I tested ElevenLabs' newest model, v3, on https://elevenlabs.io/app/speech-synthesis/text-to-speech. It's honestly really impressive! I think it's more realistic than OpenAI, though you style the voice by including instructions inline as brackets, for example [whispering] What is that noise? [screaming] Ah, a ghost!.
The voice quality is great, the expressiveness is great, and the latency is acceptable.

Here's a sample audio output:

You can find the samples in the original blog post.

The pricing is pretty straightforward. For the latest pricing details, visit https://elevenlabs.io/pricing/api. You're billed per character - as of the time of writing, 1000 characters cost $0.05 with the Flash model. I think this is more expensive than OpenAI, but it's still a great option.

3. LMNT

What really stood out to me with LMNT was how fast it is. The audio begins streaming so quickly! I tested LMNT in AgentOne and on their website at https://app.lmnt.com/. Both LMNT and ElevenLabs support voice cloning, but I found it to be a smoother experience on LMNT.
At the time of writing, there are 24 built-in voices and a generous free tier. The voice expressiveness and quality are okay, but not as good as ElevenLabs or OpenAI in my opinion. Here are some samples:

Male Voice:

You can find the samples in the original blog post.

Female Voice:

You can find the samples in the original blog post.

Pricing is simple. Check out https://www.lmnt.com/pricing for the full details.

4. Hume

Hume is a great general-purpose TTS service. The voices are expressive and natural, though they sometimes struggle or pronounce words incorrectly. Many voices are available (too many to count!). Rather than passing instructions to the model, the model relies on context to determine how to read your input aloud. This is a cool approach, but it does have downsides, and I found it harder to customize the voices compared to ElevenLabs for example.
Hume is pretty fast. It also has a free tier. More details are available on their website, here: https://www.hume.ai/pricing

Two samples:

Male English Actor:

You can find the samples in the original blog post.

Female Voice:

You can find the samples in the original blog post.

Rankings

Best overall: ElevenLabs

ElevenLabs has the highest ceiling for voice quality and customization. It is not the cheapest option, and its latency is not the fastest in this group, but if the final audio needs to sound polished, expressive, and production-ready, ElevenLabs is my top pick.

Best for low latency: LMNT

LMNT ties ElevenLabs on total score, but it wins the latency category clearly. If you are building a voice agent, conversational interface, or any product where response time matters, LMNT is the easiest recommendation. The tradeoff is that customization and expressiveness are not quite as strong as ElevenLabs.

Best developer-friendly general option: OpenAI

OpenAI is the most balanced option here. The quality is strong, the pricing is competitive, and the instruction-based customization model is convenient if you already use OpenAI APIs. I would choose OpenAI when I want solid TTS without adding another specialized provider.

Best budget pick: Hume

Hume scores best on price and has a generous free tier, which makes it a good option for experiments, prototypes, and projects where cost matters most. The main downside is control: the context-driven style system is interesting, but I found it less predictable than direct instructions or voice settings.

That's all for now! Thanks for reading.

Introducing the AI Model Directory

BestCodes — Mon, 04 May 2026 14:33:33 +0000

Today we're open-sourcing the AI Model Directory, the most comprehensive, automatically updated list of AI models and their metadata available today. It's the data layer that powers model selection in AgentOne, and now it's free for anyone to use, fork, or contribute to.

If you'd rather just look at models, we also built a browser for the directory at models.agent-one.dev where you can search, sort, and compare every model in the directory.

Why Does This Exist?

When building AgentOne, I needed a comprehensive list of AI models and their metadata - costs, context windows, supported features, modalities - so AgentOne could give users easy access to every model an AI provider had to offer.

I was frustrated with the existing options:

Models.dev is not comprehensive (it's opinionated), and it often takes anywhere from a few days to weeks for frontier models to be added across all providers
LiteLLM is more comprehensive for some providers, but the data is fragmented and harder to work with
Portkey Models doesn't list as many models as alternatives do
Other catalogs are often developed with a certain product or service in mind, so they wind up being non-agnostic, not comprehensive, or not always up-to-date

The AI Model Directory aims to be easy to use (like Models.dev), truly comprehensive across every provider it includes, and automatically updated with security in mind.

How Does It Work?

A GitHub Actions workflow runs every 24 hours and re-fetches model metadata from every supported provider. Each provider has its own small adapter that knows how to talk to that provider's API or read its docs, and normalizes the response into a single shared schema covering things like:

Pricing: input, output, reasoning, cache read/write, audio in/out
Limits: context, input, and output token limits
Modalities: text, image, audio, video, file (in and out)
Features: attachments, reasoning, tool calls, structured output, temperature
Metadata: knowledge cutoff, release date, last updated, open weights

Every model gets its own folder under data/providers/<provider>/<model-id>/index.toml, so the directory is just a tree of TOML files. This makes it easy to read, easy to diff, and easy to consume from any language. If a provider's data is wrong or missing something, you can drop a metadata.toml (with data overrides) next to the generated file and the next refresh will merge your overrides on top of the fetched data instead of clobbering them.

To provide an experience similar to models.dev/api.json, a data/all.json file is automatically generated as well, so you can pull the entire directory in one fetch. We also provide a data/all.min.json file for less bandwidth consumption:

https://raw.githubusercontent.com/The-Best-Codes/ai-model-directory/refs/heads/main/data/all.min.json

What's In the Directory?

At launch, the directory tracks models from 50+ providers, including OpenAI, Anthropic, Google, xAI, Mistral, DeepSeek, Cohere, Perplexity, OpenRouter, Vercel, GitHub Copilot, GitHub Models, Hugging Face, Groq, Cerebras, Fireworks, Together, DeepInfra, Baseten, Novita, Alibaba, Inception, Venice, Chutes, Friendli, and many more... and that list keeps growing. If your favorite provider isn't there, open an issue or send a PR; adding a new provider is usually a single small adapter file.

Browse It at models.agent-one.dev

Reading TOML files is great for machines, but not always great for humans. So we built a frontend for the directory at models.agent-one.dev.

It's a fast, sortable, searchable table with a column for everything in the schema. You can search across providers, model IDs, features, and modalities at once, sort by any column, and click straight through to a provider's website. It's the easiest way to answer questions like "which models support reasoning and tool calls under $1 per million input tokens?"

The table loads directly from data/all.min.json in the directory repo, so it's always in sync with the latest run.

Using It in Your Own Project

Consuming the directory is easy. Hit the raw GitHub URL for the bundled file:

curl https://raw.githubusercontent.com/The-Best-Codes/ai-model-directory/main/data/all.json

Or:

curl https://raw.githubusercontent.com/The-Best-Codes/ai-model-directory/main/data/all.min.json

You get back a JSON object keyed by provider, with each provider's models nested inside. This is the easiest path if you just need to populate a model picker or a pricing table. Because everything is plain files, you can fork the repo, add your own provider adapters, drop in metadata.toml for models you've measured yourself, and run the same GitHub Actions workflow on your fork. Your fork stays in sync with upstream while keeping your overrides intact.

Security

Because the directory is updated automatically based on data fetched from third-party providers, the data here is only as trustworthy as the providers it comes from. If you're using this to make billing or routing decisions, treat it as a strong default and not as gospel. We have several measures in place to mitigate the obvious vulnerabilities:

Provider endpoints are hardcoded in source, so providers cannot redirect the updater to arbitrary user-controlled URLs
All fetched data is validated against a strict Zod schema before it's written to disk, which helps prevent malformed or unexpected fields from slipping through
Model IDs are normalized into safe directory names before writing, and entries whose normalized name would be empty are rejected
If multiple model IDs normalize to the same directory name, we resolve that deterministically instead of writing multiple conflicting directories
Terminal output is sanitized before logging, which reduces the risk of ANSI escape sequences or control characters spoofing the updater output
Every network fetch has a 60 second timeout so a slow or hostile provider can't hang the update job forever
IDs and names are length-limited and reject raw control characters, which helps defend against weird escapes, invisible junk in logs, and other malformed provider output
Generated model directories that no longer exist upstream are removed automatically on refresh
Overrides stay local: metadata.toml only applies to that model directory and is merged on top of fetched data
The updater does not execute provider-supplied code, shell commands, or HTML; it only fetches remote content, parses it, validates it, and writes normalized TOML files

That said, this is still provider-supplied metadata. A provider can lie about pricing, capabilities, limits, or release dates, and some providers expose better metadata than others. The goal here is to make the pipeline safe and robust, not to pretend third-party metadata is perfectly trustworthy.

What's Next

This is a beta release, so expect a few rough edges. Some of the things we're working on:

More providers (especially regional and self-hosted offerings)
A proper docs site
Programmatic SDKs for JS/TS, Python, and Go

If you want to help shape any of this, join us on Discord, open an issue, or send a PR.

Try It Out

Browse the data: models.agent-one.dev
Read or fork the source: github.com/The-Best-Codes/ai-model-directory
Use it in your app: AgentOne already does

Happy building!

How to Use GPT 5.5 for Agentic Coding

BestCodes — Sat, 25 Apr 2026 00:41:31 +0000

OpenAI's GPT 5.5 is one of the most capable models available for agentic coding - writing code, using tools, running commands, and iterating on complex tasks autonomously. In this guide, you'll learn how to set up GPT 5.5 in AgentOne and start using it for agentic coding workflows.

What Is Agentic Coding?

Agentic coding is when an AI model doesn't just suggest code, it actively writes, runs, debugs, and iterates on code using tools. Instead of copy-pasting snippets from a chatbot, you give the agent a task and it handles the implementation end to end.

GPT 5.5 excels at this because of its strong tool use capabilities, large context window, and ability to follow multistep instructions reliably.

Prerequisites

AgentOne desktop app installed
An OpenAI API key with access to GPT 5.5

If you don't have an OpenAI API key yet, head to platform.openai.com to create one.

Step 1: Open Provider Settings

Launch AgentOne and open Settings. Navigate to the Provider section. This is where you manage all your AI providers and API keys.

Step 2: Add a Custom OpenAI-Compatible Provider

GPT 5.5 isn't in AgentOne's built-in model list yet, so you'll add it as a custom provider. Click the Add Provider button in the top right, then select OpenAI Compatible from the dropdown.

A dialog will appear with fields to configure the new provider.

Step 3: Configure the Provider

Fill in the following fields:

Name: GPT 5.5 (or whatever you'd like to call it)
Base URL: https://api.openai.com/v1
API Key: Your OpenAI API key

You can leave the Custom Headers section empty, as it's only needed for providers that require extra authentication headers.

Step 4: Add the Model

In the same dialog, scroll down to the Models section. You have two options:

Option a: Auto-Fetch Models

Click the Auto button. AgentOne will call OpenAI's /models endpoint and pull in all available models. Find gpt-5.5 in the list and remove any models you don't need.

Option B: Add Manually

Click the Add button to open the model form. Fill in:

Model ID: gpt-5.5
Display Name: GPT 5.5 (optional, for a cleaner label in the UI)

Make sure both Supports Tools and Supports Images are toggled on… GPT 5.5 supports both!

Click Add Model, then click Add Provider to save everything.

Step 5: Select GPT 5.5 in the Model Selector

Go back to the main chat view. Click the model selector (the model name shown near the input area) and you should see GPT 5.5 listed under your custom provider. Select it.

That's it. You're now using GPT 5.5 for all your conversations in AgentOne.

Step 6: Start Agentic Coding

With GPT 5.5 selected, you can give AgentOne tasks like:

"Create a REST API with Express and PostgreSQL"
"Refactor this component to use React hooks"
"Write tests for the auth module"
"Find and fix the bug in the payment flow"

GPT 5.5 will use AgentOne's tool system to read your files, write code, run terminal commands, and iterate until the task is done.

Tips for Getting the Best Results

Be specific. Instead of "make it better," say "add input validation to the sign-up form and return proper error messages."
Give context. Mention the framework, language, and any constraints. "Use TypeScript, Prisma, and the existing database schema in prisma/schema.prisma."
Let it iterate. Agentic coding works best when you let the model run commands, see errors, and fix them on its own.

Adding Other GPT 5.5 Variants

OpenAI offers several GPT 5.5 model variants. You can add multiple model IDs to the same provider:

Model ID	Best for
`gpt-5.5`	General agentic coding tasks
`gpt-5.5-pro`	More expensive, but smarter

To add more models, go to Settings > Providers, expand your GPT 5.5 provider, and use the Add button in the Models section.

Wrapping Up

Setting up GPT 5.5 in AgentOne takes less than a minute. Once configured, you get a powerful agentic coding environment where GPT 5.5 can read, write, and run your code autonomously.

If you run into issues, make sure your OpenAI API key has access to the gpt-5.5 model and that your base URL is set to https://api.openai.com/v1.

Give it a try! Download AgentOne and start building with GPT 5.5 today. Need help? Join the AgentOne Discord server.