A practical guide for connected agents

TurnScope for agents

Analyze recorded agent runs, token and cache accounting, and comparative outcome evidence.

Choose this tool

Use TurnScope for analysis of recorded data the user supplies. It does not execute an agent, fetch a trace URL, infer provider prices, or verify outcomes independently.

Connect an AI client

A model needs a configured HTTP tool connection or browsing capability. A prompt by itself does not grant API access. This is an OpenAPI import; no MCP server is required.

1. In Open WebUI, add an OpenAPI tool server in Settings → Integrations, or Settings → Admin → Integrations for a shared backend connection.
2. Set the API base URL to https://turnscope.agentextual.com and the OpenAPI spec URL to https://turnscope.agentextual.com/openapi-tools.json (or set its spec path to /openapi-tools.json).
3. Choose Authentication: None. Do not forward cookies or session credentials for this public profile.
4. Save, verify the tool list, then enable the tool in the chat's + → Tools menu or the model's tool configuration. Use a model with reliable native tool calling.

User connections run in the browser and require Direct Integrations and the relevant permission to be enabled by an administrator. Shared connections run in the WebUI backend, which must reach the API. Public browser imports support credential-free CORS; backend connections do not depend on browser CORS.

The profiles use OpenAPI 3.1 and preserve real input constraints. Tool clients differ in support for schema references, nullable values and composition. Import and execute a small request in your installed client before relying on it; no universal model or Open WebUI interoperability is claimed.

First workflow

1. Choose analyze for a recorded sequence of turns, accounting for normalized usage and cache charges, compare for paired measurements, or studies/compare for a declared trial manifest and records.
2. Prepare JSON using the matching published schema. Preserve unknown measurements as unknown; do not replace missing values with zero. Use fingerprints where raw context is unnecessary.
3. POST the JSON object with Content-Type: application/json. Read the report's limitations and coverage before interpreting costs, comparisons, or possible savings.

Try the HTTP API

curl --fail 'https://turnscope.agentextual.com/api/v1/example' -o trace.json
curl --fail 'https://turnscope.agentextual.com/api/v1/analyze' \
  -H 'Content-Type: application/json' --data-binary @trace.json

Limits and interpretation

• The request limit is 2 MiB. Only one analysis runs at once; honor Retry-After on 429. Malformed JSON/content type returns 400, oversized bodies 413, and invalid or inconsistent input 422.
• Data is processed statelessly without persistent trace storage or outbound analysis network access. Request bodies are not logged by the analysis service. Only send data you intend to provide for this analysis.
• Full recorded runs and reports may exceed a model's context budget. Use the local Python clients for larger inputs instead of asking the model to reconstruct a trace. There is no server-side summary-only mode or trace-URL import.
• Caller-supplied measurements, rates, scope, provenance and outcomes remain claims. Observations overlap and do not establish guaranteed savings, statistical equivalence, or causality.

Browser agents

This page remains readable without JavaScript. Where a browser supports the experimental WebMCP document.modelContext API, the page can register a small set of same-origin tools. Unsupported browsers keep the ordinary page and HTTP API. WebMCP is browser-local capability after visiting a page, not remote MCP or global tool registration. Read the tool's input and error messages; do not fabricate identifiers or measurements.

Documents

• Open this tool
• Public tool import
• Full API contract
• Markdown guide
• Documentation index (llms.txt)
• Trace schema
• Comparison schema
• Accounting schema
• Study schema
• Local Python client
• Evidence client

Discovery and evidence

llms.txt is a documentation proposal, not an indexing guarantee. OpenAPI describes callable operations after a client imports the service. Open WebUI Tool Search searches enabled tools, not the public internet. Clear purpose, parameter descriptions, bounded outputs and real request evaluations help a model use tools; they do not guarantee selection or ranking.

• Open WebUI integration
• OpenAPI spec support
• Tool discovery
• Writing effective tools
• llms.txt proposal

Privacy

These guides use the same aggregate usage measurement as the service. See privacy information. Request bodies and credentials are not collected by the page tracker.