HarnessRouter is an infrastructure service that lets you embed “agent harnesses” such as Codex, Claude Code, and Hermes Agent (execution environments that give an AI model tools and working procedures so it can act autonomously) into your own product through a single API. It is operated by Lumentree Corporation in the United States and is backed by Y Combinator. Putting an agent into a product normally means building the surrounding machinery yourself: starting sandboxes, managing sessions, streaming intermediate output, passing generated files back and forth, and recording execution logs. HarnessRouter takes care of all of this in one place. It launched in July 2026, and in August of the same year a self-hosted edition (Community Edition) was released as open source under the Apache 2.0 license.
Key Features
- Unified API (Unified Harness Protocol): Call multiple harnesses, including Codex, Claude Code, Hermes, DeepSeek Harness, and Pi, through a common request specification. Switching harnesses only requires a configuration change, with no need to rewrite the calling code in your product
- Isolated sandbox per task: An isolated sandbox is started automatically for each task. As parallel tasks grow, you can keep running them without thinking about infrastructure operations
- Execution tracing and cost control: Tool calls and elapsed time are recorded step by step, and each task shows a breakdown of model usage fees, agent run time, and idle time. You can also set budget limits and usage caps
- Benchmarking with Harness Arena: Run the same task across multiple harness and model combinations, compare speed, cost, and accuracy, and then decide on your production setup
- Community Edition (self-hosted): Published under the Apache 2.0 license, it lets you run the same API on your own infrastructure with a single Docker container. It is an option when you do not want API keys, state, or files to leave your environment
- Starter kits: Open-source templates for agent-powered apps such as slides, spreadsheets, dashboards, and video
Pricing
Using the console and configuring agents is free; billing is usage-based on the time agents actually run. The monthly fee of a paid plan can be used directly as usage credit.
| Plan | Monthly | Highlights |
|---|---|---|
| Free | $0 | 2 concurrent tasks, 1 workspace, Agent work $0.020/min, 7-day data retention |
| Developer | $20 | Unlimited concurrent tasks, 1 workspace, Agent work $0.014/active min |
| Production | $200 | Unlimited concurrent tasks, 5 workspaces, Agent work $0.010/active min |
| Scale | $1,000 | Unlimited workspaces, Agent work $0.006/active min, tracing (Beta) |
| Enterprise | Contact sales | Volume discounts, invoice and bank transfer billing |
If you use your own API keys, HarnessRouter does not charge for model usage; you pay each model provider directly. If you use HarnessRouter’s keys, catalog prices apply as-is with no markup (a 5% routing fee applies to spend above $25,000 per month).
Pricing is as of September 2026. Please check the official site for the latest pricing.
Pros & Cons
✅ Pros
- You skip the work of building agent execution infrastructure and can spend that time on product features
- Because you are not tied to a specific harness, it is easier to switch when a better agent comes along
- Per-task cost breakdowns and budget limits make it easier to contain spending when an agent runs longer than expected
- You can try it on the free plan, and with a self-hosted edition available, you can choose between the cloud version and on-premises-style operation depending on the use case
⚠️ Cons
- It is a new service launched in 2026, so if you build it into core features you need to consider the risk of depending on the provider
- The pricing structure has changed within a short period, and pricing information in older review articles may be out of date
- Compared with building your own execution infrastructure, fine-grained control over the inside of the sandbox is limited
- The free plan limits task run time to 30 minutes (up to 1 hour after a top-up) and retains data for 7 days
- As an API platform for developers, it is not suited to individuals who just want to use an agent on its own
Comparison with Similar Services
| Service | Difference from HarnessRouter |
|---|---|
| OpenRouter | Specializes in API routing per AI model and does not handle sandboxes or agent execution itself |
| E2B | A component that provides isolated sandboxes for code execution; it has no mechanism for switching between multiple harnesses through a unified API |
| Daytona | A container-based sandbox platform for agents focused on providing execution environments; it does not act as a routing layer across harnesses |
Who Is It For
- Development teams that want to embed agent features that generate documents, code, video, and more into their own product
- People who want to compare Codex and Claude Code on real data before deciding which fits their tasks
- Teams already integrated with one harness who want to move to a setup that is ready for a future switch
- People who want to run large-scale parallel evaluations and tests without building infrastructure
- Organizations that want to use an agent platform while keeping API keys and generated data in their own environment
Summary
When you put an agent into a product, the hard part is less the model itself than the execution infrastructure around it. HarnessRouter carves that part out as an API and is designed with comparing harnesses, choosing one, and switching later in mind. A practical approach is to run your own tasks on the free plan or the self-hosted edition first, then look at the benchmark results and cost breakdown before deciding on your production setup.