# Agent Toolbox

> Free in-browser tools and a curated directory for AI agent builders. Machine-usable via REST, OpenAPI and MCP.

Agent Toolbox is a small, independent site for developers building AI agents. Web tools run in the browser with no signup; the same tools are available to agents over a JSON API and a hosted MCP server. It also keeps a hand-curated directory of agent tools and an LLM API price list (14 models, verified 2026-10-29). Full text: https://agenttoolbox.net/llms-full.txt

## Tools
- [Token & Cost Estimator](https://agenttoolbox.net/tools/tokens): Estimate tokens and API cost for a prompt across popular LLMs.
- [JSON Schema → Tool Definition](https://agenttoolbox.net/tools/schema): Convert a JSON Schema into OpenAI, Anthropic and MCP tool formats.
- [System Prompt Linter](https://agenttoolbox.net/tools/lint): Catch vague, conflicting or risky instructions in system prompts.
- [MCP Server Config Generator](https://agenttoolbox.net/tools/mcp): Generate mcpServers JSON for Claude Desktop, Cursor and others.
- [Cron Expression Explainer](https://agenttoolbox.net/tools/cron): Explain cron schedules for agent jobs and preview next runs.
- [Regex Tester for Agent Outputs](https://agenttoolbox.net/tools/regex): Test regexes that extract structured data from LLM output.

## For agents
- [MCP endpoint](https://agenttoolbox.net/api/mcp): Streamable HTTP MCP server. Protocol 2026-07-28 (stateless, no initialize; call server/discover for versions) and 2025-03-26 to 2025-11-25 (initialize handshake)
- [OpenAPI](https://agenttoolbox.net/openapi.json)
- [Tools API](https://agenttoolbox.net/api/tools)
- [Directory API](https://agenttoolbox.net/api/directory?q=)

## Plans
- Free: 60 requests/hour per IP, no key.
- [Pro](https://agenttoolbox.net/pricing): $19/month or $190/year. Send your key as the x-api-key header. Keys are shown instantly after checkout.

## Pages
- [Directory](https://agenttoolbox.net/directory)
- [AI model API prices](https://agenttoolbox.net/models): input/output price per 1M tokens and context windows, from official provider pages
- [Best AI models for agents by budget](https://agenttoolbox.net/models/best-for-agents)
- [Compare two models](https://agenttoolbox.net/compare)
- [Best agent tools by category](https://agenttoolbox.net/best)
- [Tool builder & MCP server generator](https://agenttoolbox.net/build)
- [Community Games](https://agenttoolbox.net/games): mini games built with AI, submitted by the community and reviewed by hand
- [Submit a game](https://agenttoolbox.net/games/submit)
- [All tools](https://agenttoolbox.net/tools)
- [About & contact](https://agenttoolbox.net/about)
- [Refund policy](https://agenttoolbox.net/refunds)

## Blog (The Loop)
- [Agent memory design: what to remember across sessions, where to keep it, and how to stop it rotting](https://agenttoolbox.net/blog/agent-memory-design): How Claude's memory tool, OpenAI Agents SDK sessions and LangGraph stores split short-term and long-term memory, and the design decisions they leave to you: what to write, when to write it, who can read it, and when to forget.
- [Tracing agent runs with OpenTelemetry: the GenAI span names, attributes and privacy defaults](https://agenttoolbox.net/blog/tracing-agent-runs-with-opentelemetry): OpenTelemetry now has a dedicated set of conventions for agents, model calls, tool calls and MCP traffic. Here is how the spans nest, which attributes matter, and why prompts stay out of your traces by default.
- [Parallel tool calls: how OpenAI, Anthropic and Gemini batch them, and how to run the batch safely](https://agenttoolbox.net/blog/parallel-tool-calls-for-agents): All three major model APIs can return several tool calls in one turn. Here is how each one signals a batch, how to turn it off, and how an agent harness should execute and answer the calls without breaking the loop.
- [Budget caps for agent runs: turn limits, token budgets and provider spend limits](https://agenttoolbox.net/blog/budget-caps-for-agent-runs): A runaway agent loop can cost ten times a normal run before anyone notices. Here is how to cap a single run with turn limits and token budgets, how provider spend limits behave when you hit them, and why your retry logic needs to know the difference.
- [A2A 1.0 for agent builders: when your agent should call another agent instead of a tool](https://agenttoolbox.net/blog/a2a-protocol-for-agent-builders): The Agent2Agent protocol reached its first stable release this year. Here is how its Agent Cards, tasks, and update channels work, how it differs from MCP, and when it is worth adding to a system you already run.
- [Prompt injection in tool results: designing agents that stay safe when they read untrusted text](https://agenttoolbox.net/blog/prompt-injection-defenses-for-tool-using-agents): Filters and stern system prompts don't stop indirect prompt injection. Here is how to reason about the risk with the lethal trifecta, and the design patterns that actually limit what injected text can do.
- [Structured outputs for agents: what the JSON schema guarantee actually covers](https://agenttoolbox.net/blog/structured-outputs-for-agents): OpenAI, Anthropic and Google all offer schema-constrained JSON output. Here is what each one enforces, what it quietly ignores, and the edge cases an agent still has to handle.
- [Batch APIs for agent workloads: when half price is worth the wait](https://agenttoolbox.net/blog/batch-apis-for-agent-workloads): OpenAI, Anthropic and Google all sell batch inference at 50% off. Here is which agent jobs fit a 24-hour window, which do not, and the details that bite.
- [OAuth for remote MCP servers: the 2026-07-28 auth flow](https://agenttoolbox.net/blog/oauth-for-remote-mcp-servers): How a remote MCP server advertises its authorization server, how clients register and get audience-bound tokens, and the mistakes the spec forbids.
- [Evaluating agent tool calls: outcomes, paths and consistency](https://agenttoolbox.net/blog/evaluating-agent-tool-calls): How to build evals for a tool-using agent: grade outcomes, check only required tool calls, track pass^k, and catch grader bugs.
- [Approval gates for agent tool calls: where to put the human in the loop](https://agenttoolbox.net/blog/human-approval-gates-for-agent-tool-calls): How OpenAI's Agents SDK, the Claude Agent SDK, LangGraph and MCP let an agent pause for a person before risky tool calls, and how to choose which calls need a gate, what the reviewer sees, and how to resume safely.
- [Context compaction for agent loops: when trimming history actually cuts the bill](https://agenttoolbox.net/blog/context-compaction-for-agent-loops): Tool result clearing and server-side compaction on Anthropic and OpenAI, with worked cost math showing when shrinking an agent's context saves money, when it costs more, and how it interacts with prompt caching.
- [When reasoning models help (and waste tokens) in agent loops](https://agenttoolbox.net/blog/reasoning-models-in-agent-loops): How Anthropic thinking, OpenAI reasoning effort, and Gemini thinking_level change cost and quality inside multi-turn tool-using agents, and when to keep them off.
- [MCP drops sessions: what the 2026-07-28 revision changes for agent builders](https://agenttoolbox.net/blog/mcp-2026-07-28-stateless-revision): The newest Model Context Protocol revision removes the initialize handshake, protocol sessions, and server-initiated requests. Here is what that means for the servers you host, the clients you build, and the migration path in between.
- [Handling tool errors in agent loops: retries, idempotency, and knowing when to stop](https://agenttoolbox.net/blog/handling-tool-errors-in-agent-loops): A practical craft guide to the failures that happen between an agent and its tools, how to report them so the model can recover, when your code should retry instead of the model, and how to keep retries from doing damage.
- [How to Choose an LLM for Your AI Agent (October 2026 Edition)](https://agenttoolbox.net/blog/choosing-an-llm-for-agents-october-2026): A practical, data-backed guide to picking a model for agent work in October 2026, using Artificial Analysis and LMArena scores plus official list prices.
- [What Is an MCP Server? A Plain-English Guide, With a Free One to Try](https://agenttoolbox.net/blog/what-is-an-mcp-server): What the Model Context Protocol is, how MCP servers work under the hood, and how to connect the free Agent Toolbox MCP server to your own agent in a few minutes.
- [Why Your Agent Costs More Than You Think: A Practical Guide to Tool-Calling Costs](https://agenttoolbox.net/blog/agent-tool-calling-costs): How tool definitions, growing histories, and multi-turn loops drive the cost of AI agents, with a worked example and the caching and batching discounts that actually move the bill.
- [Writing tool descriptions and schemas that agents use reliably](https://agenttoolbox.net/blog/writing-tool-descriptions-and-schemas): A practical craft guide to naming tools, writing descriptions, and designing JSON schemas so language models call the right function with the right arguments.