Track customer spend

info

Without a LiteLLM Enterprise license, Customer Management is free for up to 5 customer organizations, 5 customer teams per organization, and 5 customer members per team. The auto-created default team counts toward the team limit. LiteLLM Enterprise removes these limits.

Use one application API key for your customers. Pass the customer's organization alias and member ID with each request. LiteLLM resolves the member's team and organization, checks their budgets, and records spend, tokens, and requests at all three levels.

Set up budget notifications, then import Acme, its Support team, and two members. Send a request for Alice and fetch usage and budgets to display in your product.

Local Customer Management after one controlled-provider request, with usage read from the accounting ledger.

Use a build containing the customer hierarchy APIs, identity headers, usage routes, and database migrations on main. Check /openapi.json for /management/v1/customer_members. Preview requirements appear beside the steps that need them; see the availability reference for the required changes.

Free tier limits

Management creates that would exceed a free-tier limit return 403. A member counts toward the limit of every team in its parent_team_ids.

Requests with a valid virtual key keep working, subject to the usual authentication, model access, budget, and rate-limit checks. If a limit prevents auto-registration, new organization aliases or member IDs are not auto-registered or tracked at that customer level. The proxy logs a CUSTOMER TRACKING OFF warning. Existing customer levels on the request remain tracked.

The Customers dashboard shows a banner when any free-tier limit is reached. If an Enterprise license lapses, existing rows keep being tracked, even when their counts exceed the free-tier limits. The limits restrict new creates, not existing tracking. Proxy admins can check current capacity with the limits API.

Before you start

You need a running LiteLLM Gateway, a connected PostgreSQL database with the customer migrations, your existing application API key, and a proxy admin key for management and usage reads.

Make these values available in your application's environment and the terminal where you run curl:

Variable	Value
LITELLM_BASE_URL	Gateway URL without /v1, for example http://localhost:4000.
LITELLM_MODEL	An existing gateway model alias your application key can use.
LITELLM_API_KEY	Your existing application API key. The same key serves every customer.
LITELLM_ADMIN_KEY	Proxy admin key for Customer Management, usage, and log reads.
START_DATE, END_DATE	Inclusive UTC dates in YYYY-MM-DD format. Use the current UTC date for both on your first run.

Keep keys on your server. Your backend authenticates the customer and selects the identity headers from that customer's session before it calls LiteLLM. It also authorizes customer access before it returns usage to your product.

1. Set up budget webhooks

Customer budget webhooks are on main through customer budget webhook support and the customer budget alert fix. Use a build containing both changes for the recurring-budget and member events below. Prepare a POST endpoint in your product that accepts JSON and returns a successful 2xx response within 10 seconds. The gateway must be able to reach it.

Replace the example endpoint below with your receiver URL. Add this configuration before importing customers or sending requests, then restart the gateway:

config.yaml, additions to your existing configuration

general_settings:
  use_spend_logs_ledger: true

litellm_settings:
  callbacks: [customer_budget_webhook]

callback_settings:
  customer_budget_webhook:
    endpoint: https://billing.example.com/litellm/customer-budget
    headers:
      X-Source: litellm

Merge these entries with your existing configuration. If you already have callbacks, add customer_budget_webhook to that list. For values supplied through your secret store, use CUSTOMER_BUDGET_WEBHOOK_URL and CUSTOMER_BUDGET_WEBHOOK_HEADERS; see configuration precedence.

After a positive spend update commits, LiteLLM checks the affected organization, team, and member. Spend at or above soft_budget produces soft_budget_crossed; spend at or above max_budget produces budget_crossed. One update can produce both events at several levels. Free requests and requests rejected before inference do not produce a customer budget spend event.

Example event for Alice's $40 soft threshold

{
  "event": "soft_budget_crossed",
  "scope": "customer_member",
  "entity_id": "acme-alice",
  "customer_org_id": "org-acme",
  "customer_org_alias": "Acme",
  "customer_team_id": null,
  "customer_team_alias": null,
  "customer_member_id": "acme-alice",
  "customer_member_alias": "Alice",
  "spend": 41.25,
  "budget": 40,
  "period": "2026-12-13T12:00:00+00:00",
  "timestamp": "2026-11-13T12:00:00Z"
}

budget is the triggering threshold. period is the recorded next reset timestamp, or "lifetime" if no reset is recorded. Route the event with customer_org_id to update the customer's balance in your product. A member budget covers that member across teams, so member events contain the organization and member IDs; the team fields are null.

Return a successful 2xx response within 10 seconds. Delivery is best effort: a cache claim suppresses the same event, scope, entity, and period until the budget reset, including after failure. Lifetime budgets use a cache claim with a 100-year TTL. Cross-worker deduplication needs shared Redis; a local fallback can produce duplicates. The HTTP client retries a connection error once, but there is no durable retry queue or schedule for timeouts or unsuccessful HTTP responses. Reconcile through the usage APIs to recover missed events.

Verify delivery in step 4, after the customer and SDK examples are ready.

2. Import a customer

Create the organization first, then its team and members. Use IDs from your application, or save the IDs LiteLLM returns when you omit them.

Create the organization and default member budget

Acme has a $1,000 budget. Its default member template gives each member a $50 budget, a $40 soft threshold, and rate limits.

POST /management/v1/customer_organizations

curl --request POST "$LITELLM_BASE_URL/management/v1/customer_organizations" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "customer_org_id": "org-acme",
    "alias": "Acme",
    "budget": {
      "max_budget": 1000,
      "budget_duration": "30d"
    },
    "default_budgets": [
      {
        "name": "standard",
        "max_budget": 50,
        "soft_budget": 40,
        "budget_duration": "30d",
        "rpm_limit": 60,
        "tpm_limit": 100000
      }
    ],
    "default_budget_template": "standard"
  }'

Expect 201 Created. A soft budget is a notification threshold and does not block requests. Webhook delivery requires the customer budget callback.

Create the team

POST /management/v1/customer_teams

curl --request POST "$LITELLM_BASE_URL/management/v1/customer_teams" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "customer_team_id": "acme-support",
    "parent_org_id": "org-acme",
    "alias": "Support",
    "budget": {
      "max_budget": 250,
      "budget_duration": "30d"
    }
  }'

Expect 201 with parent_org_id: "org-acme".

Create Alice and Bob

Create each member with the Support team's ID in parent_team_ids. Omitting budget_id selects Acme's default member template.

Create Alice

curl --request POST "$LITELLM_BASE_URL/management/v1/customer_members" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "customer_member_id": "acme-alice",
    "parent_team_ids": ["acme-support"],
    "alias": "Alice"
  }'

Create Bob

curl --request POST "$LITELLM_BASE_URL/management/v1/customer_members" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "customer_member_id": "acme-bob",
    "parent_team_ids": ["acme-support"],
    "alias": "Bob"
  }'

Expect 201 for each member. Read effective_budget in the response to verify the inherited $50 cap. Use budget_id to select another template or budget_override to customize a member. The current API rejects the old budget_template_id field.

Add an email (preview)

The email field requires #51617. On a build containing that change, add Alice's email:

PATCH /management/v1/customer_members/acme-alice

curl --request PATCH "$LITELLM_BASE_URL/management/v1/customer_members/acme-alice" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "email": "alice@acme.example"
  }'

Repeat an import

A repeated create returns 409 for an existing ID. Retrieve that ID, confirm its parent, then patch the fields you need to change. This gives your backend an upsert flow with the existing APIs.

Retrieve the existing member

curl "$LITELLM_BASE_URL/management/v1/customer_members/acme-alice" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY"

After confirming parent_team_ids contains acme-support, update Alice:

Update the existing member

curl --request PATCH "$LITELLM_BASE_URL/management/v1/customer_members/acme-alice" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "alias": "Alice"
  }'

Use the same create, retrieve, and update flow for organizations and teams. Native upserts and automatic organization creation remain under review in #51607.

3. Send a request for Alice

Use your existing application key and pass Alice's identity in request headers:

x-litellm-customer-org-alias: Acme
x-litellm-customer-member-id: acme-alice

LiteLLM resolves acme-alice to acme-support, then org-acme. The same identity headers work with Responses, Chat Completions, and Anthropic Messages.

OpenAI Responses SDK

Run pip install openai in your application project. The raw response exposes x-litellm-call-id, which you use to find this request in spend logs.

import os

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["LITELLM_API_KEY"],
    base_url=f"{os.environ['LITELLM_BASE_URL'].rstrip('/')}/v1",
)
raw_response = client.responses.with_raw_response.create(
    model=os.environ["LITELLM_MODEL"],
    input="Write a short welcome message.",
    max_output_tokens=128,
    extra_headers={
        "x-litellm-customer-org-alias": "Acme",
        "x-litellm-customer-member-id": "acme-alice",
    },
)
response = raw_response.parse()
print(response.output_text)
print("LiteLLM call ID:", raw_response.headers.get("x-litellm-call-id"))

Expect a successful Responses result. LiteLLM checks the organization, team, and member budgets, model access, and rate limits, then attributes usage to all three records.

LangChain Deep Agents

Run pip install deepagents langchain-openai. Use a model alias that supports tool calling. Configure the model for the authenticated customer's request:

import os

from deepagents import create_deep_agent
from langchain_openai import ChatOpenAI

model = ChatOpenAI(
    model=os.environ["LITELLM_MODEL"],
    base_url=f"{os.environ['LITELLM_BASE_URL'].rstrip('/')}/v1",
    api_key=os.environ["LITELLM_API_KEY"],
    use_responses_api=True,
    default_headers={
        "x-litellm-customer-org-alias": "Acme",
        "x-litellm-customer-member-id": "acme-alice",
    },
)
agent = create_deep_agent(model=model)
result = agent.invoke({"messages": [{"role": "user", "content": "Say hello to Alice."}]})
print(result["messages"][-1].text)

Select the customer's headers for each authenticated request. Do not change a shared model instance's headers while it serves other customers.

An existing member ID can resolve its parents without the organization header. With an existing organization alias and a new member ID, LiteLLM registers the member under org-acme:default and assigns the organization's default member template. An alias alone attributes usage to the organization.

With #51617, replace the member-ID header with x-litellm-customer-member-email: alice@acme.example; keep the organization alias header. An email string in the member-ID header is an ID, not an email lookup.

4. Check budgets and webhook delivery

Create a separate customer with a zero budget to check rejection without changing Acme's settings:

Create a zero-cap test customer

curl --request POST "$LITELLM_BASE_URL/management/v1/customer_organizations" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "customer_org_id": "org-budget-check",
    "alias": "Budget check",
    "budget": {
      "max_budget": 0
    }
  }'

Use the client from step 3 with the same application key and model, and this customer's alias:

from openai import APIStatusError

try:
    response = client.responses.create(
        model=os.environ["LITELLM_MODEL"],
        input="Budget check",
        max_output_tokens=128,
        extra_headers={"x-litellm-customer-org-alias": "Budget check"},
    )
    print("Request succeeded:", response.id)
except APIStatusError as error:
    print(error.status_code)

Expect 422 and a budget error naming customer_org=org-budget-check, before a provider request. A deployment can change this status to 429; see Handle a budget refusal in your client.

Use a positive-cost model to test the webhook. Set the test cap to $1 and the soft threshold to $0, then rerun the same SDK call. The request succeeds and incurs the model's normal cost. Zero-cost requests do not produce a spend event.

Allow the test request and trigger a soft-budget alert

curl --request PATCH "$LITELLM_BASE_URL/management/v1/customer_organizations/org-budget-check" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "budget": {
      "max_budget": 1,
      "soft_budget": 0
    }
  }'

After accounting commits the request's cost, check your webhook receiver for soft_budget_crossed, with scope: "customer_org" and entity_id: "org-budget-check". The event's budget is 0 and its spend is greater than 0. The earlier rejected request does not generate this event.

Expected webhook excerpt, using the local demo cost

{
  "event": "soft_budget_crossed",
  "scope": "customer_org",
  "entity_id": "org-budget-check",
  "customer_org_id": "org-budget-check",
  "spend": 0.14,
  "budget": 0
}

If the event does not arrive, confirm that the callback is enabled, the gateway can reach your endpoint, and accounting has recorded positive spend. Check gateway logs for delivery errors. Use a fresh test organization ID and alias when repeating this test: duplicate events remain suppressed for the budget period, including after a failed delivery.

Restore the test customer's zero cap and remove its soft threshold:

Restore the test cap after the check

curl --request PATCH "$LITELLM_BASE_URL/management/v1/customer_organizations/org-budget-check" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "budget": {
      "max_budget": 0,
      "soft_budget": null
    }
  }'

The first exhausted scope rejects a request: organization, then team, then member. Fresh customer spend counters across workers are on main through #51568. Use a build containing that change; older builds can admit requests against cached spend until refresh. Allow for requests already in flight and background accounting when you show balances.

Handle a budget refusal in your client

A budget refusal returns 422 by default. A deployment that sets budget_exceeded_status_code: 429 under litellm_settings returns 429, so identify the refusal from the parsed error body and not from the status or the message text.

config.yaml, optional: return 429 for budget refusals

litellm_settings:
  budget_exceeded_status_code: 429

OpenAI-compatible routes such as /v1/chat/completions set error.type to budget_exceeded:

Chat Completions, default 422

{
  "error": {
    "message": "ExceededBudget: customer_org=org-budget-check over budget. Spend=0.0, Budget=0.0. Spend resumes when the budget window resets.",
    "type": "budget_exceeded",
    "param": null,
    "code": "422"
  }
}

/v1/messages keeps an Anthropic error type and adds error.litellm_error_type:

Anthropic Messages, default 422

{
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "message": "ExceededBudget: customer_org=org-budget-check over budget. Spend=0.0, Budget=0.0. Spend resumes when the budget window resets.",
    "litellm_error_type": "budget_exceeded"
  }
}

Route	Status	error.type	error.litellm_error_type
/v1/chat/completions	422, default	budget_exceeded	Absent
/v1/messages	422, default	invalid_request_error	budget_exceeded
/v1/chat/completions	429, configured	budget_exceeded	Absent
/v1/messages	429, configured	rate_limit_error	budget_exceeded

An ordinary rate limit, such as a scope's rpm_limit, is also a 429 but has a different body: error.type is throttling_error on OpenAI-compatible routes, and on /v1/messages it is rate_limit_error with no litellm_error_type.

The OpenAI SDK exposes the inner error object as error.body:

Detect a budget refusal with the OpenAI SDK

import os

from openai import APIStatusError, OpenAI

client = OpenAI(
    api_key=os.environ["LITELLM_API_KEY"],
    base_url=f"{os.environ['LITELLM_BASE_URL'].rstrip('/')}/v1",
)


def is_budget_refusal(error: APIStatusError) -> bool:
    body = error.body
    return isinstance(body, dict) and body.get("type") == "budget_exceeded"


try:
    client.chat.completions.create(
        model=os.environ["LITELLM_MODEL"],
        messages=[{"role": "user", "content": "Budget check"}],
        max_tokens=16,
        extra_headers={"x-litellm-customer-org-alias": "Budget check"},
    )
except APIStatusError as error:
    if is_budget_refusal(error):
        print("Budget exhausted, do not retry:", error.status_code)
    else:
        raise

The Anthropic SDK exposes the whole response body, so read the nested error object. Run pip install anthropic first.

Detect a budget refusal with the Anthropic SDK

import os

from anthropic import Anthropic, APIStatusError

client = Anthropic(
    api_key=os.environ["LITELLM_API_KEY"],
    base_url=os.environ["LITELLM_BASE_URL"].rstrip("/"),
)


def is_budget_refusal(error: APIStatusError) -> bool:
    body = error.body
    detail = body.get("error") if isinstance(body, dict) else None
    return isinstance(detail, dict) and detail.get("litellm_error_type") == "budget_exceeded"


try:
    client.messages.create(
        model=os.environ["LITELLM_MODEL"],
        messages=[{"role": "user", "content": "Budget check"}],
        max_tokens=16,
        extra_headers={"x-litellm-customer-org-alias": "Budget check"},
    )
except APIStatusError as error:
    if is_budget_refusal(error):
        print("Budget exhausted, do not retry:", error.status_code)
    else:
        raise

Retrying straight away does not help: the request is refused before it reaches a provider, and the response carries no Retry-After header. The same request can succeed after the exhausted budget or window resets, after an approved budget increase request, or after an administrator raises the limit. A budget without budget_duration or a window has no automatic reset, whatever the message text says, so read budget_reset_at and budget_limits[].reset_at from the management API before you tell a customer when to try again.

Both SDKs retry 429 twice by default, with backoff, before your code sees the exception. With the default 422 the refusal reaches your except block after one request. With a configured 429 the check above still identifies the refusal, but only after the SDK has sent the request three times. To avoid those extra requests, keep the default 422, or pass max_retries=0 to the client, or use client.with_options(max_retries=0) for one call, and retry in your own code only when the error is not a budget refusal.

5. Fetch usage for your product

The organization aggregated endpoint returns complete spend, token, and request totals for the date range, without pagination. It accepts up to 100 customer organization IDs.

Complete Acme usage

curl --get "$LITELLM_BASE_URL/management/v1/customer_organizations/daily/activity/aggregated" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --data-urlencode "customer_org_ids=org-acme" \
  --data-urlencode "start_date=$START_DATE" \
  --data-urlencode "end_date=$END_DATE"

Read the response's top-level metadata:

Usage response excerpt from the local example

{
  "metadata": {
    "total_spend": 0.14,
    "total_tokens": 120,
    "total_api_requests": 1
  }
}

Dates are inclusive UTC days, with at most a 400-day difference between start and end. total_spend is cost in USD. You can use it in your product's pricing or chargeback calculation; this API does not create invoices or collect payment. The key breakdown shows the top key, while totals include every key.

Team and member usage

Support team usage, first page

curl --get "$LITELLM_BASE_URL/management/v1/customer_teams/daily/activity" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --data-urlencode "customer_team_ids=acme-support" \
  --data-urlencode "start_date=$START_DATE" \
  --data-urlencode "end_date=$END_DATE" \
  --data-urlencode "page=1" \
  --data-urlencode "page_size=100"

Alice's usage, first page

curl --get "$LITELLM_BASE_URL/management/v1/customer_members/daily/activity" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --data-urlencode "customer_member_ids=acme-alice" \
  --data-urlencode "start_date=$START_DATE" \
  --data-urlencode "end_date=$END_DATE" \
  --data-urlencode "page=1" \
  --data-urlencode "page_size=100"

These endpoints paginate stored rows. Follow metadata.has_more and increment page until it is false. Sum each page's totals once; the same date can occur on several pages. See the pagination contract. Do not add organization, team, and member totals together: they include the same requests.

Accounting runs in the background. If usage has not appeared after a successful request, check accounting completion, the UTC dates, and the identity headers. Zero-cost models can report requests and tokens with $0 spend.

6. Fetch budgets and caps

Retrieve Acme's running budget spend and organization cap:

Organization budget

curl "$LITELLM_BASE_URL/management/v1/customer_organizations/org-acme" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY"

Retrieve the Support team's running spend and cap:

Team budget

curl "$LITELLM_BASE_URL/management/v1/customer_teams/acme-support" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY"

Retrieve Alice's inherited budget and active overrides:

Member budget

curl "$LITELLM_BASE_URL/management/v1/customer_members/acme-alice" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY"

Read budget.max_budget for an organization and effective_budget.max_budget for a team or member. The same objects contain rpm_limit, tpm_limit, model limits, and allowed models. A team or member's effective budget combines its template with active overrides; an expired override no longer applies.

spend is the current budget counter, budget_usage lists the current counter and reset_at of each budget window, and budget_reset_at records the next reset. Usage reads report historical spend by UTC day for the dates you select. A reset sets the counters back without erasing that history. Budgets reset on the separately configured budget reset timezone, so a window boundary need not line up with a UTC day. To resolve a member's parents in your product, read its parent_org_id and parent_team_ids. The first team is its default team.

Open Customer Management → Customers, select Acme, and choose Select Time Range to include your requests. The usage page shows spend, requests, tokens, and the member caps. Choose Teams in Break down by to inspect Support.

Actual customer records and accounting data in the local Customer Management. Alice's controlled-provider request records $0.14 and 120 tokens.

7. Debug a request in spend logs

Use the x-litellm-call-id printed by the Responses SDK. Make that value available as LITELLM_CALL_ID in the environment where you run curl.

Find a request by its LiteLLM call ID

curl --get "$LITELLM_BASE_URL/spend/logs/v2" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --data-urlencode "request_id=$LITELLM_CALL_ID" \
  --data-urlencode "start_date=$START_DATE 00:00:00" \
  --data-urlencode "end_date=$END_DATE 23:59:59"

This API requires a date range even with a request ID. Customer identity fields require #51596. Check the saved metadata to verify Alice → Support → Acme:

Saved customer identity excerpt

{
  "metadata": {
    "user_api_key_customer_member_id": "acme-alice",
    "user_api_key_customer_member_alias": "Alice",
    "user_api_key_customer_team_id": "acme-support",
    "user_api_key_customer_team_alias": "Support",
    "user_api_key_customer_org_id": "org-acme",
    "user_api_key_customer_org_alias": "Acme"
  }
}

Compare spend and total_tokens with usage. A metadata.status of "failure" identifies a failed request; an absent or null status follows the default success behavior.

Open /ui/logs/?log_id=<x-litellm-call-id>. In the request details, check Customer for the resolved hierarchy and Metrics for cost and tokens:

Actual saved request in the local gateway, using a controlled provider and the accounting ledger. The Customer section requires #51596.

The log API has no customer organization, team, or member filter. user_id, team_id, and end_user refer to separate LiteLLM entities. Find an individual request by its call ID and use the customer usage APIs for customer totals.

Shared team budget or per-member allowance

A team budget is one counter for the whole team: every member's spend draws down the same amount. A member template is a set of limits that each assigned member inherits with its own counter, so three members on a $10 template can spend $10 each. Use a team budget to cap a group's total and a template to give each person an allowance. You can combine both.

The examples below continue from step 2 and use org-acme. They need jq to read an ID from a response. The effective_budget values shown were returned by a build of main; the spend figures in the prose are illustrations.

One shared team budget

Create a Research team with a $100 budget and add three members. They inherit Acme's standard template, so each also has a $50 member cap.

Create a team with a $100 shared budget

curl --request POST "$LITELLM_BASE_URL/management/v1/customer_teams" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "customer_team_id": "acme-research",
    "parent_org_id": "org-acme",
    "alias": "Research",
    "budget": {
      "max_budget": 100,
      "budget_duration": "30d"
    }
  }'

Add three members to the team

for member in acme-carol acme-dan acme-erin; do
  curl --request POST "$LITELLM_BASE_URL/management/v1/customer_members" \
    --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
    --header "Content-Type: application/json" \
    --data "{
      \"customer_member_id\": \"$member\",
      \"parent_team_ids\": [\"acme-research\"]
    }"
done

The team's spend is the sum of its members' requests. If Carol spends $45, Dan $40, and Erin $15, the team counter reaches $100 and LiteLLM refuses the next request from any of the three, even though each is under the $50 member cap.

A separate allowance for each member

Add a template with a $10 monthly window and a $100 annual window. The response is the organization, so read the new template's ID from its default_budgets array by name.

Add a template with monthly and annual windows

export TEMPLATE_ID="$(curl --silent --request POST \
  "$LITELLM_BASE_URL/management/v1/customer_organizations/org-acme/default_budgets" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "name": "monthly-allowance",
    "budget_limits": [
      {"budget_duration": "1mo", "max_budget": 10},
      {"budget_duration": "12mo", "max_budget": 100}
    ]
  }' | jq -r '.default_budgets[] | select(.name == "monthly-allowance") | .budget_id')"
echo "$TEMPLATE_ID"

Create a Sales team without a team budget and assign the template to three members through budget_id.

Create a team and assign the template to three members

curl --request POST "$LITELLM_BASE_URL/management/v1/customer_teams" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "customer_team_id": "acme-sales",
    "parent_org_id": "org-acme",
    "alias": "Sales"
  }'

for member in acme-frank acme-grace acme-heidi; do
  curl --request POST "$LITELLM_BASE_URL/management/v1/customer_members" \
    --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
    --header "Content-Type: application/json" \
    --data "{
      \"customer_member_id\": \"$member\",
      \"parent_team_ids\": [\"acme-sales\"],
      \"budget_id\": \"$TEMPLATE_ID\"
    }"
done

Each response has the same inherited limits:

effective_budget excerpt for each member

{
  "budget_limits": [
    {"budget_duration": "1mo", "max_budget": 10, "reset_at": "2026-12-01T00:00:00Z"},
    {"budget_duration": "12mo", "max_budget": 100, "reset_at": "2027-01-01T00:00:00Z"}
  ]
}

All three inherit the same limits from the template, but each has a separate spend counter. If Frank spends $10 this month, LiteLLM refuses Frank's next request until the monthly window resets, while Grace and Heidi can still spend $10 each. Sales has no team budget, so the three can spend $30 in a month between them.

Change the allowance for every assigned member

Patch the template to raise the monthly window to $20. A budget_limits patch replaces the whole array, so send the annual window again to keep it.

Raise the monthly allowance to $20

curl --request PATCH \
  "$LITELLM_BASE_URL/management/v1/customer_organizations/org-acme/default_budgets/$TEMPLATE_ID" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "budget_limits": [
      {"budget_duration": "1mo", "max_budget": 20},
      {"budget_duration": "12mo", "max_budget": 100}
    ]
  }'

Retrieve two of the members:

Read the effective limits for Frank and Grace

for member in acme-frank acme-grace; do
  curl --silent "$LITELLM_BASE_URL/management/v1/customer_members/$member" \
    --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
    | jq '{customer_member_id, spend, limits: .effective_budget.budget_limits}'
done

Both now report the $20 monthly window, and each keeps its own spend:

Output after the template edit

{
  "customer_member_id": "acme-frank",
  "spend": 0,
  "limits": [
    {"budget_duration": "1mo", "max_budget": 20, "reset_at": "2026-12-01T00:00:00Z"},
    {"budget_duration": "12mo", "max_budget": 100, "reset_at": "2027-01-01T00:00:00Z"}
  ]
}
{
  "customer_member_id": "acme-grace",
  "spend": 0,
  "limits": [
    {"budget_duration": "1mo", "max_budget": 20, "reset_at": "2026-12-01T00:00:00Z"},
    {"budget_duration": "12mo", "max_budget": 100, "reset_at": "2027-01-01T00:00:00Z"}
  ]
}

A template edit reaches every member still assigned to it. Changing the organization's default_budget_template_id is different: it selects the template for members created afterward and does not reassign existing members. Carol, Dan, and Erin stay on standard if monthly-allowance becomes the default.

Override one member

Give Frank a $50 monthly window until the end of the year. Use a future timestamp with a timezone.

Override Frank's monthly window

curl --request PATCH "$LITELLM_BASE_URL/management/v1/customer_members/acme-frank" \
  --header "Authorization: Bearer $LITELLM_ADMIN_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "budget_override": {
      "budget_limits": [
        {"budget_duration": "1mo", "max_budget": 50}
      ]
    },
    "budget_override_expires_at": "2027-01-01T00:00:00Z"
  }'

Frank's effective_budget now has a $50 monthly window and the template's $100 annual window. An override window replaces the template window with the same budget_duration, and the other template windows still apply. Other non-null override fields replace the matching template fields. Two cases differ: an override that sets its own max_budget drops every template window and keeps only the windows in the override, and a non-empty allowed_models override replaces the inherited list.

After budget_override_expires_at passes, Frank's effective_budget returns to the template's $20 and $100 windows. The response still includes the stored budget_override, so read effective_budget for the limits that apply now.

Monthly and annual windows

Each budget_limits entry accepts "1mo" for a monthly window or "12mo" for an annual window, at most one of each. Other values, including "1y", return 422 with Input should be '1mo' or '12mo'. This rule covers window entries only. The scalar budget_duration field beside max_budget has its own format, such as "30d".

How the limits combine

A request must pass every limit that applies to it: the organization's budget, the budget of the team the request resolves to, and the member's effective budget, including each of its windows. Each limit is compared with its own counter, so the smallest number does not decide the outcome alone. Erin can be refused at $15 of personal spend because Research has reached $100, and Frank can be refused at his $20 monthly window while Acme and Sales have room. LiteLLM checks the organization first, then the team, then the member, and the error names the first exhausted scope.

Fail-open handling (pending)

Fail-open identity handling remains under review in #51604. Main rejects identities it cannot resolve or register, and identities that conflict. A validated configuration that preserves asserted customer IDs in logs is required before this guide can provide a setup.

Troubleshooting

Result	Check
409 during import	Retrieve the existing ID, verify its parent, then patch mutable fields.
422 during member creation	Use parent_team_ids and budget_id. Supply a template or override if the organization has no default template.
401 on inference	Check the organization alias and member ID. A new member needs an existing organization alias and default template for registration.
403 on inference	Confirm that the organization alias matches the member's organization.
Budget 422, or 429 when configured	Confirm it is a budget refusal, then read organization, team, and member spend and caps to find the first exhausted scope.
429 without a budget error type	Check the scope's RPM, TPM, and model limits.
Missing customer IDs in logs	Check whether your build includes #51596.