Start saving on AI spend without changing how your team works.

This is the practical setup guide. Pick the path that matches what you are doing, follow the steps in order, then use the receipt to confirm what Multra.ai optimized and what evidence supports the savings.

1. Choose your path

2. Chat walkthrough

Chat is the fastest way to prove value because it feels like GPT or Claude but adds routing, context, caching, output controls, and a savings receipt.

1. Sign in. Open Login or create a free account.
2. Open Chat. Use the top nav item Chat.
3. Pick a project. Use the default project unless you already created a separate environment.
4. Check the mode. Paid projects default to Auto; free projects default to Audit. Use Pinned only when you want one specific model.
5. Send a realistic prompt. Use an actual support ticket, contract summary, coding request, meeting notes summary, or analysis task.
6. Review the answer. If it looks right, open the receipt to see what changed.
7. Refresh the page. The chat should still be there; history is server-backed.

Good first test prompts

Summarize this customer email in 5 bullets.
flag churn risk
suggest the next best reply
keep it under 120 words

Find the likely cause of this error.
Explain the fix in plain English.
Give me the smallest safe patch.

Review this NDA.
Flag renewal, confidentiality, indemnity, and assignment risks.
Give me a short executive summary.

Compare these two vendor proposals.
Rank by total cost, implementation risk, and support risk.
Show your assumptions.

3. Connect provider keys

Multra.ai uses your provider account. You paste an API key once, Multra.ai stores it in the provider credential store, tests it, and never shows it back to you.

4. Projects and Multra.ai API keys

A project is the boundary for API keys, default mode, receipts, optional repository context, and policy. API keys do not have their own independent mode; they inherit the project default unless a request explicitly passes an allowed mode.

1. Open APIs in the top nav, or use Account menu → Projects.
2. Use the default project unless you need separate environments.
3. Choose a project mode: Audit, Auto, or Pinned.
4. Click Create API key under that project.
5. Copy the key once and store it in your app or agent secret manager.
6. Use https://optimize.blueskyatg.com/v1 as the base URL.

Modes in plain English

5. OpenAI SDK examples

Use your Multra.ai project API key where your app normally uses an OpenAI key, and change the base URL.

Python: Chat Completions

import os
from openai import OpenAI
client = OpenAI(
  api_key=os.environ["BLUESKY_API_KEY"],
  base_url="https://optimize.blueskyatg.com/v1"
)

response = client.chat.completions.create(
  model="bluesky/auto",
  messages=[
    {"role": "system", "content": "You are concise."},
    {"role": "user", "content": "Summarize this support ticket in 5 bullets."}
  ],
  max_tokens=400
)

print(response.choices[0].message.content)

Python: Responses API

import os
from openai import OpenAI
client = OpenAI(api_key=os.environ["BLUESKY_API_KEY"], base_url="https://optimize.blueskyatg.com/v1")

response = client.responses.create(
  model="bluesky/auto",
  input="Draft a short customer renewal follow-up email.",
  max_output_tokens=300
)

print(response.output_text)

TypeScript

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.BLUESKY_API_KEY,
  baseURL: "https://optimize.blueskyatg.com/v1"
});

const response = await client.chat.completions.create({
  model: "bluesky/auto",
  messages: [{ role: "user", content: "Summarize this support ticket." }],
  max_tokens: 400
});

console.log(response.choices[0].message.content);

Raw curl

curl https://optimize.blueskyatg.com/v1/chat/completions \
  -H "Authorization: Bearer $BLUESKY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bluesky/auto",
    "messages": [{"role":"user","content":"Summarize this ticket."}],
    "max_tokens": 400
  }'

6. Anthropic SDK examples

Use the same Multra.ai project key and point the Anthropic SDK at Multra.ai.

Python

import os
from anthropic import Anthropic

client = Anthropic(
  api_key=os.environ["BLUESKY_API_KEY"],
  base_url="https://optimize.blueskyatg.com/v1"
)

message = client.messages.create(
  model="bluesky/auto",
  max_tokens=400,
  messages=[
    {"role": "user", "content": "Summarize this customer email."}
  ]
)

print(message.content[0].text)

TypeScript

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({
  apiKey: process.env.BLUESKY_API_KEY,
  baseURL: "https://optimize.blueskyatg.com/v1"
});

const message = await client.messages.create({
  model: "bluesky/auto",
  max_tokens: 400,
  messages: [{ role: "user", content: "Summarize this customer email." }]
});

 console.log(message.content[0].text);

7. Coding agents

Coding agents should use a Multra.ai project API key, not provider keys. Codex, Cursor, Windsurf, Cline, Roo, OpenAI SDK, Anthropic SDK, CI agents, and Claude tools receive bounded large-context handling when Multra.ai can preserve quality safely. Paid projects use Auto by default; free projects use Audit.

Codex CLI and IDE

1. Create a Multra.ai API key. Open APIs → Projects & API keys, choose the project, and create a production key. Copy the generated key once.
2. Connect provider keys in Multra.ai. Open Provider connections and connect OpenAI, Anthropic, or both. Codex will use the Multra.ai key; Multra.ai uses these provider keys behind the gateway.
3. Open the real Codex config file. Use the macOS, Linux, or Windows PowerShell command on the Develop page to create and open ~/.codex/config.toml.
4. Paste the provider config. Add the Multra.ai provider block below and save the file.
5. Export your Multra.ai key. Put the key in your shell profile, password manager, or CI secret. Do not paste provider keys into Codex.
6. Run one small task. Start Codex in a repo and ask for a low-risk read-only task first. Then check Analytics → Calls for the receipt.

# ~/.codex/config.toml
model = "bluesky/auto"
model_provider = "bluesky"

[model_providers.bluesky]
name = "Multra.ai"
base_url = "https://optimize.blueskyatg.com/v1"
env_key = "BLUESKY_API_KEY"
wire_api = "responses"

# macOS/Linux shell
case "${BLUESKY_API_KEY:-}" in
  bs_live_*|bs_test_*) ;;
  *) echo "Create and copy a Multra.ai API key before running this command." >&2; exit 1 ;;
esac
codex

# First test prompt:
# "Read this repo and tell me the test command. Do not edit files yet."

If you already use Codex with OpenAI directly

Do not replace your provider API key with an OpenAI key here. Codex should receive only the Multra.ai project key. Multra.ai then routes to the provider connections in your Multra.ai account and records the savings receipt.

Cursor, Windsurf, Cline, Roo, Claude, or SDK clients

1. Create a Multra.ai project API key.
2. Set the base URL to https://optimize.blueskyatg.com/v1 for OpenAI-compatible clients, or use the Anthropic-compatible route for Claude clients that support custom endpoints.
3. Set the API key to the generated Multra.ai project key, never the provider key.
4. Use model bluesky/auto unless the tool requires a specific value.
5. For Claude Code or Claude Desktop without a custom Anthropic endpoint, use Multra.ai MCP.
6. Run a small coding prompt first, then check Analytics → Calls.

8. GitHub and code context

GitHub is optional. Use it when developers want repository-aware context without pasting large files into every request.

1. Open Develop → Repositories.
2. Click Install GitHub App.
3. Choose the organization or repositories you want Multra.ai to access.
4. Return to Multra.ai and click Sync repositories.
5. Open Projects and choose a default repository for the project.
6. Click Build graph or Retry if the repository asks for indexing.
7. Use the coding agent normally. Multra.ai merges paths, open files, changed files, and natural prompt seeds; if graph context is unavailable, normal coding requests fail open instead of blocking the request.

9. MCP setup

/mcp is the human page. MCP clients should use https://optimize.blueskyatg.com/api/mcp. MCP optimize_run uses the same bounded large-context policy and project-default repository graph behavior as coding gateway requests.

MCP server URL:
https://optimize.blueskyatg.com/api/mcp

OAuth protected-resource metadata:
https://optimize.blueskyatg.com/.well-known/oauth-protected-resource/mcp

Authorization server metadata:
https://optimize.blueskyatg.com/.well-known/oauth-authorization-server

Available MCP capabilities

• Tools: optimize_estimate, optimize_run, receipt_get, models_list, skills_list.
• Resources: receipts, projects, Skills, and models.
• Prompts: audit-this-task, optimize-this-task, and run-with-skill.

Safe write behavior

Read-only actions may run when authorized. Write actions require exact-action approval. Destructive actions are never silently run.

10. Skills walkthrough

Skills package repeatable instructions, references, and assets so users do not have to retype the same setup every time.

1. Create a folder with SKILL.md.
2. Add optional references/ for supporting docs.
3. Add optional assets/ for reusable files.
4. ZIP the folder.
5. Upload the ZIP in Multra.ai.
6. Multra.ai validates the package without executing scripts.
7. Enable the Skill for yourself, a project, or the organization.
8. Use @skill in Chat, or pass the Skill ID through API/MCP.

my-renewal-skill/
  SKILL.md
  references/
    renewal-policy.md
  assets/
    response-template.md

---
name: Renewal reviewer
description: Review renewal risk in customer contracts.
---

Use this Skill when the user asks about contract renewal risk.
Read only the relevant renewal and termination clauses.
Return: summary, risk level, dates, and recommended action.

11. Read receipts

Open Analytics → Calls to inspect each request. Receipts are the proof layer: they show selected route, token counts, cache behavior, output controls, and evidence labels.

What each label means

• MEASURED: directly reported by the provider or observed from the completed request.
• ESTIMATED: based on public pricing or known rates, but not directly billed by the provider in the receipt.
• MODELED: a counterfactual comparison that was not separately executed.
• UNKNOWN: pricing or usage evidence was missing. Multra.ai does not turn unknown into zero.

Why savings can be negative

Sometimes the safest or highest-quality path costs more. Multra.ai shows negative savings instead of hiding them because the receipt should be trusted.

12. Plans, billing, and cancellation

1. Free Audit: no Multra.ai credit card required. Provider usage may still bill through OpenAI or Anthropic.
2. Paid plans: open Account menu → Subscription.
3. Checkout: paid checkout goes through Stripe.
4. Cancellation: manage cancellation through Stripe. Access continues through the paid period.
5. Provider billing: your provider account remains separate from Multra.ai subscription billing.

13. Troubleshooting