Documation - Step - by Step Guide

01 / THE WORKFLOW

One place to send and track work.

LTECP means Low Traffic Executable Command Protocol. The gateway accepts your request, gives it a tracking ID, schedules it within configured limits, and stores the result. You can follow that work in your browser or terminal.

The provider's answer returns through the gateway to your application. Type a prompt and read the answer under Your result. Advanced JSON and Technical details provide native request and response control when you need it. Provider keys stay on the server; the browser uses a separate EDS access token.

02 / FIRST REQUEST

See OpenAI in action.

1. Open the gateway in another tab. Keep this guide beside it.
2. Obtain your EDS gateway access token from your EDS administrator. Paste it into EDS access token. This is not your OpenAI API key.
3. Leave Gateway URL set to the gateway's address and click Connect gateway. The token field clears; the connection keeps it only in page memory.
4. Under Configured model, select openai / gpt-4.1-mini. If the list is out of date, disconnect and reconnect using your EDS token.
5. For this demonstration, leave LTECP ON and choose Urgent in the traffic monitor. These controls update the settings for your gateway client, including other sessions using that client.
6. Keep Prompt selected. In Your prompt, type Give me a five-item checklist for preparing an EDS Mobile software release. Leave Maximum output tokens at 512 or choose a smaller limit.

Click Submit request once. The gateway returns a task ID. Under Your result, follow the status and read the answer when it is available. Use Copy answer to copy the text or expand Technical details — full response JSON to inspect the complete record.

Each submission is a separate task; earlier answers are not automatically included in a new prompt. Your text is sent without shortening it.

Optional: use the native request

Select Advanced JSON to edit the provider payload directly. This example sends the same request:

{
  "model": "gpt-4.1-mini",
  "input": "Give me a five-item checklist for preparing an EDS Mobile software release.",
  "max_output_tokens": 512
}

The selected format determines what is submitted. Prompt and Advanced JSON keep separate drafts. Keep the payload’s model matched to the selected model; other supported native settings are preserved.

This sends a real request to OpenAI and uses API credits. The pilot limits output to 512 tokens. Short answers can finish before you notice the Running counter change.

03 / OTHER MODELS

Try Claude or a no-charge demonstration.

Anthropic Claude Haiku 4.5

Select anthropic / claude-haiku-4-5-20251001, type your prompt, and submit. Prompt mode creates the Anthropic request for you. This is a paid provider request. For direct payload control, select Advanced JSON and use this example with messages and max_tokens:

{
  "model": "claude-haiku-4-5-20251001",
  "messages": [
    {"role": "user", "content": "Give me a five-item checklist for preparing an EDS Mobile software release."}
  ],
  "max_tokens": 512
}

Read the answer under Your result. Technical details retain the original result.content, including other provider response blocks.

Mock demonstration

Select mock / mock-demo and type a prompt to exercise submission, tracking, and the traffic widget without calling an LLM. Its deterministic response is a demonstration, not a generated answer to your prompt. The following Advanced JSON example also selects the mock success scenario:

{
  "model": "mock-demo",
  "messages": [
    {"role": "user", "content": "Demonstrate LTECP task tracking."}
  ],
  "max_tokens": 64,
  "scenario": "success"
}

Keep the selected provider and the payload's model matched. LTECP does not automatically switch your request to another model or provider.

04 / UNDERSTAND THE RESPONSE

Your result comes with a record.

Task ID

The tracking ID for accepted work. Use Copy task ID to save it. After reconnecting with the same client’s access token, paste it into Open a task by its ID and click Open task. Viewing a task does not create another provider request.

QUEUED / RETRY_WAIT / RUNNING

The request is waiting, waiting for another allowed attempt, or executing. A queue can reflect local throttles, provider cooldowns, budget limits, or Low traffic mode.

SUCCEEDED / INCOMPLETE

SUCCEEDED means the provider completed the request. INCOMPLETE means it ended without a full completion, for example at an output limit. Inspect the returned content before deciding what to do next.

FAILED / EXPIRED / CANCELLED

Execution failed, the task's deadline passed, or cancellation was recorded. Read the error and attempt information to understand what happened.

EXECUTION_UNKNOWN

The gateway cannot establish the provider's final outcome. Preserve the task ID and ask an operator to investigate before submitting a fresh request that could be billed again.

Attempts and timing

Attempts record individual dispatches. first_output_ms measures time to the first provider output; total_ms measures the provider attempt. The span from task creation to completion also includes queue time. The browser currently displays the result after completion, rather than streaming each output token.

Tokens and cost

Provider usage reports input and output tokens. reserved_usd is the amount reserved before dispatch; accountedUsd is the gateway's accounted amount. Successful usage is reconciled using configured prices. These figures are not a provider invoice.

Mode and actual tier

mode: urgent describes LTECP scheduling. actualTier: standard describes provider processing. The pilot uses standard processing for both live providers; urgent scheduling does not grant paid provider priority.

05 / THE TRAFFIC MONITOR

Watch your own request conditions.

Select the matching provider/model route in the monitor, such as openai/gpt-4.1-mini · standard. Routes also distinguish input-size bands. The chart plots observed time to first output, while counters show queued work, running work, and delivery dead letters.

UNKNOWN

A usable baseline needs 20 successful observations within 24 hours for that route. Readings also become UNKNOWN when the last observation is older than 60 seconds. It does not mean the gateway is broken or that provider traffic is low.

Elevated conditions

Recent errors or latency above the route's threshold can mark conditions ELEVATED. Recovery uses successive improved readings. No paid background probes run just to populate the chart.

LTECP ON + Urgent

Dispatch at the earliest permitted opportunity within rate and budget limits. The usual urgent deadline is 10 minutes. Consecutive requests can still queue.

LTECP ON + Low traffic

Wait for suitable observed conditions, with controlled real-task trials when observations are unavailable. The usual deadline is 24 hours. Use this for work that can wait.

LTECP OFF

New requests use standard execution while durable tracking, rate limits, and budgets remain active. Turning it off does not cancel previously accepted tasks.

Disconnect

Close the page's connection and clear its in-memory credential and displayed task result. Accepted tasks continue on the server. Reconnect and use a saved task ID to open the result, or retrieve it through the terminal.

06 / TERMINAL WORKFLOW

Keep a live monitor beside your work.

On the configured EDS workstation, open a terminal in the LTECP-Protocol-Gateway repository. The private wrapper already knows the cloud address and the access-token file. Run:

../.ltecp-local/ltecp-cloud status
../.ltecp-local/ltecp-cloud models
../.ltecp-local/ltecp-cloud watch

watch stays open as a terminal visualizer. Submit prompts from the browser or another terminal. Press Ctrl+C to detach; accepted tasks continue.

Submit a saved payload

Save the OpenAI JSON example above as openai-request.json in your working directory. Then run:

../.ltecp-local/ltecp-cloud submit --provider openai --payload openai-request.json --mode urgent --wait

This submits a real, paid request. The CLI prints the submission key and task ID, waits for completion, and prints the result. Save those identifiers if the connection is interrupted.

Retrieve a known task later by replacing TASK_ID with its actual ID:

../.ltecp-local/ltecp-cloud result TASK_ID

Another computer

The workstation wrapper is a local setup convenience, not a file supplied by this website. Install the repository's CLI and obtain an owner-only gateway token file from your administrator. From the repository, the equivalent monitor command is:

cli/.venv/bin/ltecp --url https://ltecp-pilot-oz2kfqxxfa-uc.a.run.app --token-file /path/to/private/gateway-token watch

CLI installation and local setup instructions. Keep credentials in the token file rather than placing their values in shell commands.

07 / LIMITS AND HELP

Know what the pilot allows.

EDS cloud pilot settings, documented September 22, 2026. Operators can change these settings; the connected gateway is authoritative.

The local token-rate ceiling is 8,192 reserved tokens/minute per provider. The input bound conservatively uses payload bytes plus overhead, so a request may reach the bound before an exact tokenizer would. Provider credits and Google Cloud hosting charges are separate from gateway budgets.

I cannot connect

Check the gateway URL and use the EDS access token, not an OpenAI or Anthropic key. Remove accidental surrounding whitespace. If access is rejected, ask your administrator to verify the token and its permissions.

OpenAI or Anthropic is missing

Disconnect and reconnect to refresh the configured-model list. If it is still missing, ask the operator whether that provider is enabled.

My request is queued

Check the mode and rate/budget limits. Allow about a minute between live requests at the pilot's local throttle. Provider cooldowns or Low traffic mode can require longer; the task's deadline still applies.

Credits or budget exhausted

PROVIDER_QUOTA_EXHAUSTED requires the operator to check provider billing or quota. DAILY_BUDGET_WAIT means the gateway's daily budget is unavailable. Adding provider credits does not increase the gateway's budget.

The payload or output limit is rejected

In Prompt mode, choose an output limit within the displayed model limit. In Advanced JSON, use valid JSON, match the selected model, and use the provider's correct output field: max_output_tokens for OpenAI or max_tokens for Anthropic. Keep it at or below 512 for live models. Reduce a long prompt yourself or ask the operator to review its bounds; LTECP does not silently truncate it.

I lost the connection after submitting

If a task ID arrived, use Open task to retrieve it. If the receipt was not confirmed, keep the page open and use Retry same submission; it reuses the original body and recovery key. The draft is locked until that receipt is resolved. Before closing or disconnecting, preserve the recovery key for an operator to investigate. Refreshing the page clears that in-memory recovery information. A fresh submission can create duplicate paid work.

The graph says UNKNOWN

Select the correct route and check sample age and count. See Traffic and controls; creating extra paid requests solely to fill the chart is unnecessary.

08 / WHAT COMES NEXT

A working pilot with a production path.

The gateway currently runs on Google Cloud Run with Cloud SQL PostgreSQL and Memorystore Redis. OpenAI and Anthropic have passed direct, gateway, and browser generation checks. Deployment verification also covers authentication, telemetry, and preservation of task records.

Hosted web application — available as a pilot

Use the prompt box and readable answers from your browser, with optional native JSON controls, technical details, copy buttons, and saved-task lookup. An EDS access token is still required.

EDS product integration — API available

Mobile, web, and other EDS applications can submit work through the authenticated API. Each product needs its own integration and access configuration.

Terminal — available

Use the CLI for submissions, result retrieval, and the live HUD.

Desktop application — future packaging

A native or packaged desktop interface could connect to the hosted gateway and show the visualizer. The current web widget is not an installed desktop application.

Self-hosting or another cloud — requires setup

The Docker setup can run the gateway, PostgreSQL, and Redis locally. An organization can evaluate its own hosting or a container platform such as Railway, with deployment, database, secrets, networking, and operational configuration. Moving hosts does not change provider billing, rate limits, or answer quality.

Engineering references: API integration, deployment, operations, and validation status.