Custom proxies

About the Ingrazzio proxy

Ingrazzio is JetBrains' internal proxy protocol that Junie uses to communicate with LLM providers. The production Ingrazzio endpoint is https://ingrazzio-cloud-prod.labs.jb.gg.

When Junie connects through an Ingrazzio proxy, it uses the base URL to access several sub-endpoints:

• LLM chat — the base path handles streaming chat completion requests.

• Web search — the /search path provides web search capabilities.

• URL extraction — the /extract path fetches and extracts content from URLs.

• Authentication — the /auth/test and /auth/reset paths validate the token and license state.

When using the default JetBrains provider, Junie authenticates automatically via JetBrains Account (JBA). When using a custom proxy, JBA authentication is bypassed — you must supply any required credentials (such as Authorization: Bearer <token>) through the proxy's headers field. All the sub-endpoints above, including authentication, are routed through the selected proxy with those headers.

Configure a proxy

Add a proxies array to your config.json. Each entry describes a named proxy endpoint:

Select a proxy as the active provider

To make Junie use a configured proxy, set the provider field to the proxy name:

You can also override the provider at runtime with the --provider CLI flag:

When a proxy is selected as the provider, Junie does not use JetBrains AI authentication. All authentication must be handled through the headers you configure on the proxy entry.

Legacy: the INGRAZZIO_URL environment variable

Setting the INGRAZZIO_URL environment variable creates an implicit Ingrazzio-kind proxy named ingrazzio-env pointing at that URL — no config.json entry is required:

The synthetic proxy behaves like a proxy declared in config.json:

• All requests — LLM chat, web search, URL extraction, and authentication (/auth/test, /auth/reset) — are routed through the given URL, and it appears in the model selector next to configured proxies.

• It automatically becomes the default provider unless another provider is selected via the CLI (--provider) or config.json.

• If config.json declares any Ingrazzio-kind proxy, that configuration takes precedence and INGRAZZIO_URL is ignored entirely.

• If an ingrazzio token is supplied via the environment (for example, JUNIE_API_KEY or EJ_AUTH_INGRAZZIO_TOKEN), it is sent as the Authorization header of the proxy's requests; otherwise Junie falls back to the JetBrains Account token.

Restricting the model list

available-models lists the JetBrains model ids a proxy is allowed to serve, for example anthropic-claude-4-6-sonnet. It works for every proxy kind.

The list is a filter, not a catalogue: Junie shows the intersection of it with the models that would otherwise be visible for that proxy. An OpenAI proxy can only ever serve OpenAI models, so listing an Anthropic id there adds nothing. Ids that this version of Junie does not recognize are ignored, and a list that matches nothing at all leaves the proxy out of the model selector entirely — check the log if a configured proxy does not appear.

Leave the field out to accept whatever the proxy kind supports.

Supported proxy kinds

The kind field determines which protocol Junie uses to communicate with the proxy.

Bedrock proxies

A Bedrock proxy routes Anthropic models through an AWS Bedrock-compatible gateway. Requests are sent to <api-url>/model/<wire-name>/invoke, where <wire-name> is the provider-side model id with Bedrock's default cross-region inference prefix.

Bedrock is the one kind where available-models is effectively required: omitting it lets Junie offer every Anthropic model it knows, and your gateway is unlikely to serve all of them.

Because Bedrock model ids differ from Junie's, a Bedrock proxy also accepts model-overrides, which maps a JetBrains model id to the name your gateway knows it by. It is the only kind that does — other kinds always send the JetBrains model name.

If the proxy's headers include an Authorization header, it is used as-is; otherwise Junie falls back to the JetBrains Account token.

Quick setup example

1. Create a configuration file (for example, my-config.json):

2. Run Junie with the custom config:

You can still override the model at runtime with the --model flag:

Merging proxies across configuration files

When multiple configuration files define proxies, they are merged by name:

• If two files define a proxy with the same name, the higher-priority file's fields override the lower-priority file's fields on a per-field basis.

• Headers from both files are combined (deduplicated).

• available-models is replaced, not combined — otherwise a lower-priority file could widen a list the higher-priority one deliberately narrowed.

• model-overrides entries from both files are combined per model id, with the higher-priority file winning (Bedrock only).

• Proxies with different names are all included in the final list.

For the configuration precedence order, see Configuration files.

Limitations

• The Ingrazzio, Bedrock, OpenAI, Google, and Anthropic proxy kinds are currently supported. The JetBrainsAI and OpenRouter kinds are reserved for future use and fail when selected as the active provider.

• Proxy configuration is only available through config.json. There are no dedicated CLI flags for defining proxy entries.

• When using a proxy provider, JetBrains AI authentication is bypassed entirely. You must supply any required credentials via the headers field.