# Set up Interceptor

Install a complete released package and use a local Docker daemon matching its
architecture. Released packages need no local Go toolchain, including for wrapper
builds on macOS.

## Install the release package

Choose your documented release from [official releases](https://releases.intrusive.ai)
or [Interceptor releases](https://github.com/intrusiveai/interceptor_sandbox/releases).
Download the OS/architecture archive and its matching checksum file. From the
download directory, replace `YOUR_INTERCEPTOR_VERSION` with the selected release
number without the leading `v`. This macOS ARM64 example installs into a separate
version folder:

```sh
INTERCEPTOR_VERSION=YOUR_INTERCEPTOR_VERSION
PLATFORM=darwin_arm64
shasum -a 256 --check "interceptor_sandbox-v${INTERCEPTOR_VERSION}-${PLATFORM}-SHA256SUMS"
mkdir -p "$HOME/.local/lib/interceptor/$INTERCEPTOR_VERSION"
tar -xzf "interceptor_sandbox-v${INTERCEPTOR_VERSION}-${PLATFORM}.tar.gz" \
  -C "$HOME/.local/lib/interceptor/$INTERCEPTOR_VERSION"
export PATH="$HOME/.local/lib/interceptor/$INTERCEPTOR_VERSION/interceptor_sandbox:$PATH"
interceptor version
```

Set the version to your selection. For Linux, use `linux_amd64` or `linux_arm64`
and `sha256sum --check`. Keep packages in separate directories when upgrading.

The macOS archive contains `interceptor`, `interceptor-linux`, `release.json`, and
license material. Keep them together; the Linux companion runs in the wrapper.
Copying only the CLI or mixing companion versions breaks package verification.
Linux packages use their native executable for the wrapper. Docker Desktop on
macOS and Docker Engine on Linux must run matching native Linux images.

## Prepare private host storage

Choose a local private data directory if you need an explicit location:

```sh
umask 077
mkdir -p ./interceptor-data
export INTERCEPTOR_DATA_DIR="$PWD/interceptor-data"
```

Use the same setting in terminals that operate the session. Otherwise Interceptor
uses its OS user-configuration directory. The directory contains evidence and host
bindings; do not place it in a broadly shared or synchronized folder.

## Configure model access

Use [Model and credential providers](model-and-credential-providers.md) for the
complete generation and embedding provider matrix, configuration examples, cloud
identities, and secret stores. Supported stores include AWS Secrets Manager,
Azure Key Vault, Google Cloud Secret Manager, HashiCorp Vault, and 1Password.
Real provider credentials are resolved on the host and are never made available
to the target container.

Keep real provider credentials on the host. The agent's blueprint uses the local
model endpoint and declared synthetic SDK credentials. Choose the generation
provider, base URL, authentication method, and allowed model explicitly.

For OpenAI Chat with a host variable already populated securely, the relevant
`interceptor run` options are:

```sh
--model-provider openai-chat \
--provider-base-url https://api.openai.com/v1 \
--provider-auth api-key-env \
--model-api-key-env TARGET_MODEL_KEY \
--allow-model YOUR_TARGET_MODEL
```

These are options to the full run command, not a command by themselves. Substitute
an available model. The target's blueprint model name must match the allowed model.
The variable's value does not go in command arguments or blueprint assets.

If the target uses remote OpenAI embeddings, configure them independently:

```sh
--embeddings-provider openai-compatible \
--embeddings-base-url https://api.openai.com/v1 \
--embeddings-auth api-key-env \
--embeddings-api-key-env TARGET_EMBEDDING_KEY \
--allow-embedding-model text-embedding-3-large
```

Generation settings and credentials are not inherited by embeddings. The host may
explicitly use the same credential for both, but each connector is selected separately.
Native Bedrock Titan uses `bedrock-titan`, AWS authentication, and its own regional
model selection. It is not interchangeable with an OpenAI embeddings endpoint.

See [model authentication](model-and-credential-providers.md#select-model-authentication)
and [credential providers](model-and-credential-providers.md#configure-credential-providers)
for the other supported configurations, including separate embedding credentials.

## Verify the installation through a target

Check `interceptor version`, Docker availability, and the complete package layout.
Then [prepare an agent](bring-your-agent.md), validate/seal its blueprint, build and
verify its wrapper, start a session, and invoke a legitimate operation. A CLI version
check or available key alone does not verify model access or the target's dependencies.

`interceptor run` owns the local API at `http://127.0.0.1:8080` and remains in the
foreground. One active instance is supported on a trusted local host. No separate
gateway or service manager is needed for the walkthrough. Keep its terminal open
until Operator finishes and you have collected evidence.
