Foundry: Train and evaluate browser agents reliably
Foundry is a platform for building, evaluating and improving browser-based AI agents through simulated browser environments, tracked evaluation and training data.
What is Foundry?
Foundry is a platform for teams working on browser-based AI agents. It provides simulated browser environments that are reproducible, so agent runs are not affected by page drift, IP bans or rate limits. Each agent action is tracked and classified, so click failures, layout shifts and misfires are recorded rather than passing silently. Beyond simulation, Foundry offers custom long-horizon datasets produced by expert annotators for supervised fine-tuning, plus reinforcement learning sampling that lets teams evaluate unlimited trajectories without anti-bot constraints. Integration is through a Python SDK, with an example evaluation loop that resets an environment, runs an agent against a task, then collects final state, state mutations and events for scoring against ground truth.
Focus areas: Training Data.
Who is Foundry for?
- Teams building browser automation agents
- AI engineers needing reproducible agent evaluation
- Companies fine-tuning agents on enterprise web platforms
Is Foundry worth it?
A focused simulation and evaluation layer for teams building browser agents, though pricing and general availability are still unconfirmed
Where Foundry is strong
- Reproducible browser environments without drift or rate limits
- Action-level tracking and error classification during evaluation
- Python SDK integration with an example evaluation loop
Where it falls short
- Pricing is not published and access requires a private beta application
- Sample evaluation dataset is listed as coming soon
How do you get started with Foundry?
- 01Apply for access to the private beta
- 02Review the published benchmark results
- 03Integrate the Python SDK and run the evaluation loop example
Foundry: frequently asked questions
What does Foundry actually do?
It provides simulated browser environments, action-level evaluation, annotated training datasets and reinforcement learning sampling for browser AI agents.
How do I integrate Foundry?
Through a Python SDK. The site shows an example evaluation loop that creates an environment with an API key and task ID, resets it, runs an agent, then reads final state, state mutations and events for scoring.
How much does Foundry cost?
Pricing is not published. The pricing page returns a 404 and the site directs users to apply for the private beta.
Can I try it without applying?
Not currently. Access is through a private beta application, and the sample evaluation dataset is listed as coming soon.
Foundry alternatives
The closest tools in our directory, matched on category, focus areas and audience.
Appen provides data and services to improve AI model performance and accelerate AI development.
| Tool | Best for | Free plan | Starting price | Focus areas |
|---|---|---|---|---|
| Foundry | Teams training and evaluating browser agents at scale | Unknown | — | Training Data |
| Scale AI | AI labs and enterprises needing training data and model evaluations | No | — | Training Data |
| Snorkel AI | AI teams needing specialized training data and evaluations | No | — | Training Data |
| SuperAnnotate | AI teams producing and managing training data at scale | No | — | Training Data |
| Unitlab Annotate | Teams labeling multimodal or specialized data at production scale | Yes | — | Training Data |
| Appen | Appen provides data and services to improve AI model performance and accelerate AI development. | Unknown | — | Training Data |
| Basicai Cloud | Free AI-powered training data platform with annotation tools and teamwork management. | Unknown | — | Training Data |
| Giskard | An open-source framework for testing all ML models, from tabular to LLMs. | Yes | Free | — |
| UpTrain | Teams evaluating, testing and monitoring LLM apps in production | Unknown | — | — |
| WhyLabs | Teams studying open source AI observability and data logging | Unknown | — | — |
| Cleanlab | An open-source framework for data-centric AI that automatically finds and fixes errors in datasets. | Yes | — | — |

