Overview
Bench for Claude Code is an agent benchmarking and debugging platform built by Silverstream AI that gives developers deep visibility into their Claude Code sessions. By installing a lightweight hook via the open-source claude-bench-autotrace library, Bench captures every tool call, prompt, and assistant response as an OpenTelemetry span and streams it to your Bench endpoint in real time.
The result is a continuous, structured trace of exactly what your AI coding agent did, why it did it, and where it slowed down or made mistakes — without any manual instrumentation or changes to your existing Claude Code workflow.
Designed for developers and AI engineering teams who rely on Claude Code for complex coding tasks, Bench fills a critical gap: making autonomous agent sessions observable, reproducible, and improvable.
Key Features
Full-session OpenTelemetry tracing — Automatically captures every tool call, prompt, and assistant response as structured spans, giving you a complete audit trail of each Claude Code session without writing a single line of telemetry code.
Real-time streaming to Bench endpoint — Telemetry data is streamed live to your Bench dashboard as the session runs, so you can watch your AI agent in action and catch issues immediately rather than after the fact.
Zero-friction hook installation — A single setup step on
bench.silverstream.aiconfigures the Claude Code hook automatically; there is no complex SDK integration or environment reconfiguration required.Benchmarking — Silverstream describes Bench as a benchmarking and debugging platform for Claude Code sessions. Public materials support task-level review and comparison, but do not yet document a dedicated regression-tracking workflow.
Open-source tracing library — The
claude-bench-autotracehook is MIT-licensed and publicly available on GitHub, making it auditable and forkable. Public materials reviewed do not document direct compatibility with self-hosted observability stacks.Agent debugging platform — Bench's UI is positioned as a session review and debugging interface. Public materials support tool-call, file-change, subagent, and sharing workflows, but do not fully document span-level diagnostics such as stalled or looped-session detection — capabilities borrowed from Silverstream AI's enterprise-grade autonomous web agent infrastructure.
How to Get Started
Bench integrates with Claude Code through a hook-based approach that takes only a few minutes to configure:
Visit bench.silverstream.ai — Sign up or log in to your Bench account to obtain your personal endpoint URL.
Install the autotrace library — Install the open-source Python package from the
silverstream-ai/claude-bench-autotraceGitHub repository.Configure the Claude Code hook — Follow the on-screen instructions at bench.silverstream.ai to register the hook in your Claude Code environment. Claude Code handles the rest of the setup automatically once your endpoint is configured.
Run a Claude Code session — Every subsequent session will automatically stream OpenTelemetry spans to your Bench dashboard in real time.
Explore traces in Bench — Use the Bench UI to browse session timelines, inspect tool calls, and compare runs side by side.
No changes to your existing prompts, tasks, or Claude Code commands are needed after initial setup.
Pricing & Plans
Bench is currently presented publicly as Free. No official pricing page or published paid tiers were found on Silverstream's site during review.
| Plan | Price | Details |
|---|---|---|
| Free | Free | Public materials describe Bench as free, but no official plan matrix or feature-allocation table is published. |
The underlying claude-bench-autotrace hook library is MIT-licensed and free to use, inspect, and modify. Silverstream AI does not publish official team, enterprise, or custom-deployment pricing in the public materials reviewed for this page.
Best For
- Claude Code power users who run long or complex sessions and want to understand what the agent actually did
- AI engineers and developers building automations on top of Claude Code who need regression tracking across runs
- Teams evaluating AI code checker tools and wanting native observability before committing to a production workflow
- Developers investigating bugs in agent behavior who need replay and span-level inspection rather than raw terminal logs
- Researchers and early adopters interested in open-source agent benchmarking infrastructure
FAQ
What is Bench for Claude Code?
Bench for Claude Code is a session tracing and benchmarking platform made by Silverstream AI. It works by attaching a hook to Claude Code that captures every tool call, prompt, and assistant response as an OpenTelemetry span and streams the data to a Bench dashboard in real time.
How does the Claude Code hook work?
The hook uses the claude-bench-autotrace open-source library, which integrates with Claude Code's hook system. Once your Bench endpoint is configured, Claude Code automatically sends telemetry to Bench whenever you start a session — no changes to your commands or prompts are needed.
Is Bench for Claude Code free?
Public materials list Bench as free, but no official pricing page or paid-tier breakdown was found during review. The claude-bench-autotrace hook library is MIT-licensed and free to use.
Does Bench support AI coding agents other than Claude Code?
Bench is currently documented specifically for Claude Code. No official roadmap or product announcement regarding additional agent integrations has been published by Silverstream.
Is the tracing library open source?
Yes. The claude-bench-autotrace library is published under the MIT license on the silverstream-ai GitHub organization. You can inspect the code, fork it, or integrate it with your own observability stack.
How does Bench compare to other observability tools like OpenObserve?
Bench is purpose-built for AI agent sessions in Claude Code, whereas OpenObserve is a general-purpose log and metrics storage platform. Bench provides agent-specific features like session replay, span-by-span tool call inspection, and cross-run benchmarking that general observability platforms do not offer out of the box.
What data does Bench collect from my Claude Code sessions?
Bench captures tool calls, prompts, and assistant responses as OpenTelemetry spans. These are streamed to your personal Bench endpoint. Public documentation is limited on Bench-specific telemetry retention and security practices. Confirm current data-handling details with Silverstream directly before use in sensitive environments.
Who is behind Bench?
Bench is built by Silverstream AI, a San Francisco-based startup founded in 2024 that raised a $1.2M pre-seed round led by Google's Gradient Ventures. The company specializes in infrastructure for autonomous agents and has publicly reported a 94% top-5 success rate on common ServiceNow tasks.
