Skip to main content
You’ll need access to an Impact product and an app with a supported JavaScript or Python integration.
1

Install the setup plugin

Choose your coding tool. Replace the example path with the full marketplace directory or Git URL supplied by Impact. Plugin distribution is currently in preview; a public install URL is not available yet.
Start a new Codex session in your app and invoke sdk-setup. In the desktop app, select Impact SDK setup from the registered marketplace. This plugin targets Codex CLI and desktop; the separate Impact IDE extension has its own installation.
Use the Git URL supplied by Impact, or a local marketplace directory containing .agents/plugins/marketplace.json and .claude-plugin/marketplace.json. The standalone plugin ZIP does not contain those manifests and cannot replace the full marketplace directory.
2

Ask it to instrument your app

The plugin checks your stack, installs the SDK and helps apply the startup and shutdown changes. Installing the plugin alone does not change your app.
3

Connect your Impact product

Create an ingest key in Impact and put the endpoint and raw key in your app’s secret configuration. Tell your agent where the configuration lives; keep secret values out of chat.
Ask a workspace owner or admin to open Ingest keys → Create ingest key in the intended product. Use the OpenTelemetry exporter settings URL ending in /v1/traces and the raw ingest key. Impact adds the Bearer header. Your model-provider credentials stay as they are.
Choose your content settings before sending traffic. Available inputs, outputs, tool, retrieval and media content is captured by default, within SDK limits and stricter native settings. Ask the plugin to preserve your application’s privacy controls. JavaScript controls · Python controls
The plugin uses the installed SDK’s version and support facts and reports missing or unverified capture. Simulation, Protect, presence and Source upload are configured only when you request them.After setup, run your app and find its Trace in a separately authorized Impact session. The ingest key allows delivery; it does not grant read access.
When replacing a supplied local review bundle, remove its previous registration with the commands below, then install the new bundle using the steps above.
Removing the plugin stops future helper discovery. It does not uninstall the SDK or undo application edits.

Run your app. Open its Trace.

Make one ordinary request using your app’s existing interface.
  1. Open Traces in the same Impact product.
  2. Find your service and the request you just ran.
  3. Inspect the model call, available usage and captured input and output.
Check the endpoint and key, initialization order, supported integration and runtime.status(). A successful startup or shutdown alone does not confirm delivery. Use a signed-in Impact session to view the result; an ingest key does not grant read access. JavaScript troubleshooting · Python troubleshooting

Try a complete example

These standalone examples make one billable OpenAI Responses call. Use an OpenAI project key separately from your Impact ingest key. Set OPENAI_MODEL to a model your project can use and adjust the model-specific options if needed. Imports are passive; initialize Impact before constructing the client.
Use a new directory with npm init -y, then install the exact prerelease and the example’s client:
Use Node.js 22.23.5 or 24.22.3, or a later security patch in either line, with ESM. The documented npm version is 11.21.2. Keep an existing application’s compatible clients and lockfile instead of replacing them for this example.Add .env.impact-quickstart to .gitignore before creating it with your values:
Save this as quickstart.mjs:
If the environment variables are already set, run node quickstart.mjs. In the same Impact product, find impact-quickstart and match the printed runtime ID to impact.runtime.id. Check Recorded fields for the model and available usage, then compare the prompt and printed reply when your product retains content. Confirm the selected SDK version in the output. See installation for supplied archives and troubleshooting for zero exported spans or missing Traces.
Use Python 3.11 or later. These commands use a POSIX shell:
Set IMPACT_ENDPOINT, IMPACT_API_KEY, IMPACT_SERVICE_NAME=impact-quickstart and OPENAI_API_KEY through your application’s secret configuration as described above. Keep the complete prerelease pin. In an existing uv project, use uv add 'impact==1.2.0rc10' alongside its compatible clients. See installation for supplied wheels and existing OpenTelemetry.Save this as quickstart.py:
In the same Impact product, find impact-quickstart and match its impact.runtime.id to the printed runtime ID. Check the model and available usage in Recorded fields, then match the prompt and printed reply when your product retains content. Verify the selected SDK version. Cleanup, local export and a visible Trace are separate checks; see delivery diagnosis for unknown, incomplete or exported-not-readback status. Zero exported traces leaves capture unverified.
OpenAI API keys · OpenAI model reference

Next steps

Choose your integration, add application context, or choose what content to capture in JavaScript or Python.