A two-person startup emerging from Y Combinator launched Tuesday a service designed to eliminate one of the more tedious headaches in developing AI coding agents: testing integrations without risking production systems or burning through API rate limits.
Archal, based in San Francisco, provides what it calls API sandboxes—isolated environments that behave like GitHub, Slack, Stripe and more than 20 other widely used services, but run entirely as simulations. Developers can test their agents against these clones, examine what changes occur, and reset everything to a clean baseline without creating a single test account.
The idea grew from a frustration co-founder and CEO Aidan Tiruvan knew well. "When building an agent, it always bothered me that I could not evaluate exactly what actions it could take before dropping it into production," he wrote in a LinkedIn post last month. The solution his team built aims to bridge that gap between development and deployment, offering a controlled proving ground where mistakes don't cascade into real data.
Here's how it works in practice. A developer spins up a sandbox and receives REST and Model Context Protocol endpoints for each cloned service. The agent runs its tests. The developer inspects what changed. Then the whole environment resets to its starting state. Those clones maintain state across calls, enforce referential integrity, and throw the same errors you'd encounter with production APIs, per the company's documentation.
The platform emphasizes speed in its design. According to the product page, current timing metrics for sandbox operations reflect ongoing optimization efforts, with provisioning, starting, and resetting operations all completing in under a second. Those figures matter in continuous integration environments where every second compounds across hundreds of test runs.
The platform currently supports a sprawling list: GitHub, Slack, Linear, Datadog, GitLab, Stripe, Supabase, Discord, Google Workspace, HubSpot, Jira, ClickUp, Customer.io, Cal.com, Apify, Sentry, Ramp, Tavily, Webflow, OwnerRez, PriceLabs and Unipile. That coverage reflects the reality that modern applications rarely depend on just one or two external services.
Traditional testing approaches, as Tiruvan outlined in the company's Y Combinator launch post in late August, involve creating test accounts for each service, juggling authentication credentials, navigating API rate limits, and scrubbing dirty data between runs. It's not complicated work, exactly, but it's friction that accumulates. "Coding agents are getting much better at writing integrations, but testing those integrations is still painful," he wrote.
Archal tries to collapse that overhead into a single API key. Developers use it to manage every sandbox and environment. The platform integrates with CI systems, outputs JUnit XML for existing pipelines, and returns non-zero exit codes when tests fail defined thresholds.
Tiruvan's background includes stints in machine learning research at Scale AI and NASA. He studied at the University of Colorado Boulder, where he competed in the Putnam mathematics competition and reached USACO Platinum level—credentials that suggest comfort with both theoretical rigor and practical coding challenges. Co-founder Noah Song also attended Colorado Boulder, overlapping with Tiruvan for part of that period.
The company has identified three core workflows for its product. First, harness testing: comparing different models, prompts or tool configurations from identical starting states. Second, integration and quality assurance testing, where CI jobs run against isolated service data. Third, reinforcement learning or post-training scenarios that need parallel rollouts from declared states. That last category hints at the more experimental edges of agent development, where developers iterate rapidly on behavior patterns.
Archal ships a command-line tool and SDK through npm. The "archal" package hit version 0.11.3 recently, and the company also maintains an "@archal/vitest" plugin for JavaScript testing frameworks. Documentation covers state contracts, lifecycle semantics, CLI commands, MCP tools and an OpenAPI reference.
New developers receive $20 in usage credits when they sign up, no credit card required. Current pricing and credit details are available on the company's official site.
The broader landscape for agent testing includes several established players, each tackling the problem from different angles. LangSmith, developed by LangChain, provides evaluation and tracing tools focused on agent behavior and outputs. Braintrust markets similar evaluation capabilities. But neither offers stateful clones of actual SaaS APIs, according to a comparison HokAI published in July.
E2B runs a cloud platform for executing agent code in Linux-based sandboxes. Cloudflare introduced persistent code interpreter sandboxes in April that let agents execute code in isolated environments. Those approaches handle execution security but don't specialize in mimicking third-party service APIs.
Tiruvan's pitch distills to a straightforward proposition: "Create an API sandbox with the environments your CI, tests, and evals need." For developers wrestling with the mundane complexities of integration testing, that promise of reduced friction might be enough to warrant the experiment.
