Prompt Prism

Local Agent debugging

See the request.
Follow the trace.
Fix the agent.

Built for Agent developers, Prompt Prism reveals prompt changes, normalized model output, tool activity, and complete multi-request traces through a transparent local proxy.

View on GitHub Read the guide
npm install -g prompt-prism

Node.js 20+ · MIT licensed · zero runtime dependencies

Live inspection 127.0.0.1:1028/_pp/
Prompt Prism Dashboard showing a multi-request agent trace and Proxy URL button

01 / Inspect

The whole model loop,
in one place.

Move from the request list to the exact tool arguments without reconstructing events from raw logs.

01

Trace every call

Follow user input, reasoning, assistant text, tool calls, and tool results across captures. Jump between linked events and reveal the real parameters.

02

Find the input change

Input Diff locates inserted and removed messages, then compares System, Tools, and request options.

03

Understand tool use

Browse declared schemas and parsed parameters, then jump from usage totals to each actual call in Trace.

04

Read normalized output

Inspect text, reasoning, usage, errors, stop reasons, and provider-neutral tool arguments.

05

Keep the raw truth

Unsupported traffic still passes through and remains available as a redacted HTTP request and response.

02 / Transparent

Your traffic keeps moving.

SSE and JSON responses are forwarded immediately. Capture and analysis happen on a side channel.

01Your agentyour Agent stack
request
02Prompt Prismlocalhost:1028
SSE / JSON
03ProviderAnthropic / OpenAI
redacted local capture

03 / Compatible

One proxy.
Two normalized protocols.

Automatic detection resolves each capture independently, so one protocol never locks or affects the next.

AAnthropic Messages
  • JSON and SSE
  • Thinking blocks
  • Tool calls and results
  • Prompt cache usage
OOpenAI Chat Completions
  • JSON and SSE
  • Reasoning extensions
  • Function tool calls
  • Cached token usage
REverything else
  • Still forwarded
  • Redacted headers
  • Raw request and response
  • No guessed normalization

04 / Local by design

The inspector stays
on your machine.

Captures are stored under your local data directory. Authentication headers and cookies are redacted before storage, and the oldest captures are evicted when the configured limit is reached.

  • 01 No hosted Prompt Prism service
  • 02 No analytics or telemetry
  • 03 Configurable data directory and storage cap

05 / Start

Inspect your next request.

Start Prism, generate a Proxy URL, and configure it as your Agent's model API target through its framework, SDK, CLI, environment, or HTTP client.

# install
npm install -g prompt-prism

# start the local proxy
p2 start

# open the Dashboard and click Proxy URL
http://127.0.0.1:1028/_pp/

06 / Go deeper

Use it. Measure it. Extend it.

Guide

Provider setup, CLI options, protocol detection, embedding, and privacy.

 Agent Insights

Compare token use, cache reuse, timing, and tool behavior across runs.

 Development

Run the demo, understand the monorepo, and work on Input Diff.

Open source · MIT

Stop guessing what
the model saw.

Install from npm Explore the source