---
title: "Set up Pioneer AI"
description: "Choose where Pioneer AI answers come from, set budgets and conversation retention, tune your own endpoint, and choose which MSP clients get it."
url: "https://support.outpostplatform.com/pioneer-ai/set-up/"
product: "depot"
type: "how-to"
status: "draft"
ai_summary: "Settings > AI for ai:manage: Pioneer AI chat, Keep conversations for (days), Automatic approval, Outpost Cloud AI tier and monthly budget, the Home Lab free allowance, your own endpoint (URL, API style, key, models, context window, reasoning effort, cache key), prices and budgets, per-person budgets on Settings > Team, Verify errors, and the MSP client policy and override."
availability:
  plans: ["home-lab", "essentials", "enterprise", "msp"]
  permission: "ai:manage"
  preview: false
source: "https://gitlab.com/outpostplatform/docs/-/edit/main/src/content/docs/pioneer-ai/set-up.md"
license: "CC BY 4.0"
---

# Set up Pioneer AI

This page shows how to choose where [Pioneer AI](/pioneer-ai/) gets its
answers, how to limit what it can cost in a month, how long conversations are
kept, and, on the MSP plan, which clients get Pioneer AI. When you finish, the
people on your team who hold `ai:use` can work with Pioneer AI from the source
you chose.

Pioneer AI chat is on by default with Outpost Cloud AI as the source, so on
most Outposts it already works before anyone visits this page.

## Before you begin

- You need the `ai:manage` permission, which the owner holds. Anyone with
  `ai:use` can open **Settings** > **AI** and read the settings, but cannot
  change them. See [Roles and permissions](/depot/roles-and-permissions/).
- To use your own endpoint, have its base URL, its API style (OpenAI-compatible
  or Anthropic), an API key if it needs one, the exact model name, and the
  model's context window.
- On an Outpost hosted by Outpost, your own endpoint must be reachable on the
  public internet. Addresses on private networks, such as `10.0.0.0/8`,
  `192.168.0.0/16`, loopback, and link-local, are refused. A self-hosted Depot
  or Remote Outpost can use an endpoint on your LAN.

## Turn Pioneer AI chat on or off

1. Go to **Settings** > **AI**.
2. On the **Pioneer AI** card, turn **Pioneer AI chat** on or off.
3. Select **Save**.

When chat is off, the chat says **Pioneer AI is turned off.** for everyone.

The **Proactive AI features** switch beside it lets Outpost use AI without
anyone asking, such as the **Ask AI** step in a workflow. It is off by default,
and AI steps take their fallback until it is on. See
[AI steps in workflows](/service-desk/workflow-ai-steps/).

## Choose how long conversations are kept

1. Go to **Settings** > **AI**.
2. In **Keep conversations for (days)**, enter a number from `0` to `365`. The
   default is `30`.
3. Select **Save**.

The period counts from a conversation's last message. `0` deletes a
conversation after a day without activity. A shorter period applies to
existing conversations too, so conversations older than the new period are
deleted within the hour. See
[Pioneer AI data and privacy](/pioneer-ai/data-and-privacy/#how-long-they-are-kept).

## Choose automatic approval

**Automatic approval** on the same card lets a separate AI safety review
approve some actions for technicians who opt in. It is off by default. See
[Turn on automatic approval](/pioneer-ai/automatic-approval/).

## Choose whether read-only shell commands ask {#read-only-shell-commands}

Pioneer AI judges every command it sends to a shell before it runs. See
[Shell commands](/pioneer-ai/approve-actions/#shell-commands).
**Run read-only shell commands without asking**, on the **Pioneer AI** card,
decides what happens to commands that only read, such as `hostname` or
`ipconfig`:

- **On**, the default: they run at once, and the reply says so with the
  reason.
- **Off**: they ask for approval like a change.

Commands that change the endpoint always follow the approval rules, whatever
this setting says. Select **Save** after changing it.

## Choose web search

Pioneer AI can search the web for current versions, patches, and advisories.
It is on by default. See [Pioneer AI web search](/pioneer-ai/web-search/) for
what a search may contain and what it costs.

1. Go to **Settings** > **AI**.
2. Turn **Web search** on or off.
3. Select **Save**.

With your own endpoint, Pioneer AI searches only with your own key. Enter a
Brave Search API key in **Your own search key (optional)**, under the endpoint
fields, and select **Save**. The key is encrypted and never shown again. When
no search is available from your source, the **Web search** setting says so.

## Choose Outpost documentation

Pioneer AI can search and read this documentation site to answer questions
about Outpost itself. It is on by default and free. See
[Pioneer AI and the Outpost documentation](/pioneer-ai/outpost-docs/).

1. Go to **Settings** > **AI**.
2. Turn **Outpost documentation** on or off.
3. Select **Save**.

## Choose what sub-agents may do

Pioneer AI can hand parts of a large task to
[sub-agents](/pioneer-ai/use-pioneer-ai/#sub-agents). Under **Sub-agents** on
**Settings** > **AI**, select what they may do, then select **Save**:

| Area | What it allows | Default |
| --- | --- | --- |
| **Read Outpost data** | Looking things up, as Pioneer AI does | On |
| **Research: web search, documentation and memories** | Web search, your documentation and knowledge base, and memories | On |
| **Draft scripts** | Writing script text in the report, without running it | On |
| **Endpoints**, **Run and save scripts**, **Tickets**, **Software deployment**, **Policies** | Changes in that area, listed under **Changes** | Off |

Changes a sub-agent makes still need approval, the same as Pioneer AI's own.
A sub-agent can never take a high-impact action, whatever you allow.

## Set a spending limit per reply

**Spending limit per reply (USD)** on **Settings** > **AI** caps what one reply
may cost, including its sub-agents and web searches. The default is `0.50`,
and the range is $0.01 to $100. When a reply reaches it, Pioneer AI says what
it found and what is left, and waits for the technician to continue.

The limit uses the estimated cost of each call. On your own endpoint the
estimate needs both prices under **Cost and budget**, so without them there is
no limit.

## Give Pioneer AI instructions

**Instructions for Pioneer AI** and **Safety review policy** on the same card
tell Pioneer AI how your team works and add rules to the automatic approval
review. See [Give Pioneer AI instructions](/pioneer-ai/instructions/).

## Use Outpost Cloud AI

Outpost Cloud AI is not available on [Remote Outpost](/remote-outpost/), and
the option is not shown there. The card says **This Depot runs offline, so
answers come from an endpoint you run.**

1. Go to **Settings** > **AI**.
2. Under **Source**, select **Outpost Cloud AI**.
3. In **Default tier**, choose **Fast**, **Standard**, or **Premium**. Each
   option names the model that answers for it. Under the list, the console
   shows that model, its input and output rates per 1M tokens, and the model it
   falls back to if it is unavailable.
4. Optional: in **Monthly budget (USD, optional)**, enter the most you want
   Outpost Cloud AI to cost in a month, such as `50.00`. Leave it blank for no
   limit.
5. Select **Save**.

The card also shows **Estimated so far this month**. When the month's estimate
reaches the budget, Pioneer AI stops for the rest of the month and the chat
shows **This month's Pioneer AI budget has been reached. Raise it in Settings
to keep using Pioneer AI.** Raising or clearing the budget lets it start again
at once.

Outpost Cloud AI is billed by use, at the rates the console shows for each
tier. Usage appears as **Pioneer AI** lines on your Outpost invoice. See
[Invoices, taxes, and payments](/account/invoices/).

### On Home Lab

Home Lab is not billed for Outpost Cloud AI. Instead, it includes a free
monthly allowance. When the month's allowance is used, the chat shows **Your
plan's free Pioneer AI allowance for this month is used up. It resets on**
followed by the date. The allowance resets on the first day of each month
(UTC). To keep working before then, use your own endpoint.

### During a free trial

Outpost Cloud AI has a spending limit of its own during a trial. When the trial
reaches it, the chat shows **This trial has reached its Pioneer AI limit.
Activate your subscription to keep using it.** See
[Free trials](/account/free-trials/).

## Use your own endpoint

1. Go to **Settings** > **AI**.
2. Under **Source**, select **Your own endpoint**.
3. In **Endpoint URL**, enter the base URL including the port, such as
   `http://ollama.example.com:11434/v1`. OpenAI-compatible URLs usually end in
   `/v1`. Anthropic URLs do not, for example `https://api.anthropic.com`.
4. In **API style**, choose **OpenAI-compatible** or **Anthropic**.
5. In **API key**, enter the key. Leave it blank if the endpoint needs none.
6. In **Model**, enter the model name exactly as the endpoint expects it.
7. Optional: in **Fast model (optional)** and **Premium model (optional)**,
   enter the models for those tiers. A tier with no model of its own uses
   **Model**.
8. Optional: in **Vision model (optional)**, enter a model that can read
   images. Pioneer AI uses it to read images attached in a conversation and to
   see the screen when it
   [controls a remote session](/remote-access/pioneer-control/). Leave it
   blank to turn both off.
9. In **Default tier**, choose the tier Pioneer AI uses.
10. Under **Advanced**, set the options in the next section.
11. Select **Save**.

The key is encrypted when it is saved and is never shown again. Once a key is
saved, the field shows **Saved**. To keep it, leave the field blank when you
save other changes.

If you change the host, port, or scheme of **Endpoint URL**, or change **API
style**, enter the key again in the same save. Otherwise the save is refused
with **Changing the endpoint address or API style would send the stored API key
somewhere new. Enter the API key again as part of this change.**

Usage through your own endpoint is never billed by Outpost.

### Advanced options

| Option | What to enter |
| --- | --- |
| **Context window (tokens)** | The model's real context size, such as `131072`, and at least 16,384. Pioneer AI summarizes older messages before a conversation outgrows it. Left blank, it assumes 32,768 tokens. |
| **Reasoning effort** | **Model default**, **Low**, **Medium**, or **High**, for models that take a reasoning effort. Takes effect only with **Send the reasoning effort**. |
| **Send a prompt cache key** | Sends a key for each conversation, so an OpenAI-compatible endpoint can reuse the conversation's cached prefix. Leave it off if the endpoint rejects unknown fields. |
| **Send the reasoning effort** | Sends **Reasoning effort** with each request. Only for models that take one. Leave it off if the endpoint rejects unknown fields. |

The Anthropic API style ignores both send options.

## Set budgets for your own endpoint

Outpost does not bill your own endpoint, but it can estimate each call from
the prices your provider charges you, and stop Pioneer AI when a budget is
reached.

1. Go to **Settings** > **AI**.
2. Under **Cost and budget**, enter **Input price (USD per 1M tokens)** and
   **Output price (USD per 1M tokens)**. Without both prices, every call is
   estimated at nothing and no budget can be set.
3. Optional: in **Monthly budget (USD, optional)**, enter the organization
   budget, such as `50.00`.
4. Select **Save**.

The card then shows **Estimated so far this month** and when the budget
resets. Budgets run by calendar month in UTC. When a budget is reached,
Pioneer AI stops for the people it covers until the month resets.

To give one person a budget of their own:

1. Go to **Settings** > **Team**.
2. On the **Pioneer AI budgets** card, find the person and select **Set
   budget**, or **Change** if they have one.
3. Under **How it relates to the organization budget**, choose one:

   | Option | Effect |
   | --- | --- |
   | **Personal cap within the organization budget** | Their spend counts toward the organization budget, and they stop at whichever runs out first. |
   | **Separate budget, not counted** | Their spend does not count toward the organization budget, and the organization budget does not stop them. |
   | **Separate budget, still counted** | Their spend counts toward the organization budget, so it can stop others, but the organization budget does not stop them. |

4. In **Monthly budget (USD)**, enter the amount, such as `25.00`.
5. Select **Save**.

The card lists each person's **Spent this month**, their **Budget**, and how
many of their requests a budget refused. **Use the organization budget**
removes a personal budget. A person with a budget sees what is left under
the message box.

## Verify the source

**Verify** sends one short question to the saved source and shows the answer.
It checks what is saved, not what is on the screen, so save first.

1. On the **Verify** card, select **Verify**.
2. Read the result:
   - **The endpoint answered.**, with the reply, the model, and the time it
     took, means the source works.
   - Anything else is a message that says what failed.

The card keeps the last result, such as **Last verified** with **Working** or
**Failed**.

| Message starts with | What to check |
| --- | --- |
| **The host name could not be resolved.** | The URL for typos, and that the Depot server can resolve the name. |
| **The connection was refused.** | The port, and that the AI service is running and listening on it. |
| **The TLS handshake failed.** | `http` against `https`, and that the certificate is valid for the host. |
| **The endpoint did not answer in time.** | The address, firewalls, and that the model is loaded. |
| **The endpoint rejected the API key** | The key and its permissions. |
| **The endpoint returned 404.** | The base URL path, the API style, and the model name. |
| **The endpoint is rate limiting requests** | Your plan with the provider, or try again shortly. |
| **The endpoint returned a server error** | The AI service's logs. |
| **The endpoint answered, but not in the expected format.** | That **API style** matches the service. |
| **This address is on a private network** | Use a publicly reachable endpoint. Shown only on an Outpost hosted by Outpost. |

## Choose which clients get Pioneer AI

On the MSP plan, the settings on **Settings** > **AI** in your practice view
apply to every client, and the **Pioneer AI** card has a **Clients** list.

1. Go to **Settings** > **AI**.
2. In **Clients**, choose one:

   | Option | Effect |
   | --- | --- |
   | **Off for all clients** | Pioneer AI is off for every client and for your practice. Per-client overrides are hidden. |
   | **Off unless enabled per client** | Off for each client until you turn it on for that client. |
   | **On unless disabled per client** | On for each client until you turn it off for that client. This is the default. |

3. Select **Save**.

To override the default for one client:

1. Go to **Clients** and open the client.
2. On the **AI** card, in **Pioneer AI**, choose **Use MSP default**, **Force
   off**, or **Force on**.
3. Select **Save**.

**In effect now** shows whether Pioneer AI is on for that client. The **AI**
card is not shown when the policy is **Off for all clients**. Changing an
override needs `ai:manage`. Anyone who can read the client sees the card read
only.

## Verify

1. Open Pioneer AI with **Open Pioneer AI** in the header.
2. Send a short question, such as `Which of my endpoints are offline?`

A reply streams in under **Pioneer AI**. On the MSP plan, select a client whose
setting is off and open Pioneer AI again. It shows **Pioneer AI is turned off
for this client.**

Every change to these settings, to a personal budget, and to a client override
is recorded in the [audit log](/depot/audit-log/). The API key is never
recorded.

## Next steps

- [Use Pioneer AI](/pioneer-ai/use-pioneer-ai/)
- [Turn on automatic approval](/pioneer-ai/automatic-approval/)
- [Review Pioneer AI usage and conversations](/pioneer-ai/usage/)