Connect JetBrains AI Assistant to Tokens

JetBrains AI Assistant is the AI plugin for IntelliJ IDEA, PyCharm, WebStorm and the other JetBrains IDEs. Besides JetBrains' own subscription, it can use your own API key with a third-party provider. Tokens fits the OpenAI-compatible provider type, which sends OpenAI Chat Completions requests to https://tokens.bd/v1. The models then appear in the AI Chat model selector.

This guide was checked against JetBrains' AI Assistant 2026.2 documentation (the providers page is dated 15 July 2026 and the overview 5 August 2026; read October 2026). It was checked against the documentation, not run end to end with a live key.

What you need#

• A JetBrains IDE with AI Assistant installed. JetBrains lists CLion, DataGrip, DataSpell, GoLand, IntelliJ IDEA, PhpStorm, PyCharm, Rider, RubyMine, RustRover and WebStorm. Use version 2026.1.2 or later (see the path bug below).
• A Tokens key. API keys covers spend caps and allow-lists.
• A model id from the model catalog. This guide uses deepseek/deepseek-v4.1-flash.

If your organization uses JetBrains IDE Services or JetBrains Central, JetBrains says an administrator can restrict providers or block your own API keys. If the settings below are missing, ask your admin.

Add Tokens as a provider#

1. Open Settings and go to Tools | AI Assistant | Providers & API keys.
2. In Third-party AI providers, set Provider to OpenAI-compatible.
3. In URL, enter https://tokens.bd/v1.
4. In API Key, paste your Tokens key.
5. Set Tool calling on or off (see below).
6. Click Test Connection, then Apply.
7. Open AI Chat and click the model selector. Your Tokens models are listed under the provider's section.

Which URL to enter#

JetBrains' pages say only "Specify the URL of the provider's API endpoint". They give no example that includes or leaves out /v1, and they do not say whether the field takes a path.

Enter https://tokens.bd/v1, which ends in /v1 and has no /chat/completions. The reasons:

• Two bug reports in JetBrains' tracker show the IDE building the model-list request itself. In LLM-22721 a custom path was ignored and the IDE asked for /api/v1/models (closed as a duplicate). In LLM-22911 a base URL ending in /v4/ was rewritten to /v1. JetBrains lists LLM-22911 as fixed in 2026.1.2.
• Tokens serves the model list at https://tokens.bd/v1/models, so a URL that ends in /v1 gives the right path whether the IDE keeps your path or replaces it with /v1.

If Test Connection fails and the Tokens request log shows a 404 unsupported_endpoint with a doubled path such as /v1/v1/models, the IDE is adding /v1 itself. Try https://tokens.bd (the same host without /v1) instead. This fallback is an inference from the bug reports, not something JetBrains documents.

JetBrains' pages do not say how the model list is built, but the bug reports above show the IDE requesting /models. Tokens serves it filtered by key, plan and wallet, so a key with an allow-list shows only the allowed models.

Tool calling#

JetBrains describes the Tool calling setting as whether the model supports calling tools configured through MCP, and it appears only for OpenAI-compatible providers. Turn it on for a model that supports tool calling (check Choosing a model) if you use MCP servers in AI Chat. Leave it off for a model that does not. JetBrains' page does not document tool use outside MCP.

Use Tokens for inline completion#

AI Completion (inline completion and next edit suggestions) has its own provider setting.

1. In the same settings page, find the AI Completion section and set Provider to OpenAI Compatible.
2. Enter your Tokens key as API key and https://tokens.bd/v1 as Base URL.
3. Enter the Model, Model context, Max output tokens and Prompt schema, then click Test Connection and Apply.

JetBrains says the model must be served by the endpoint in Base URL, and that inline completion needs a model with Fill-in-the-Middle (FIM) support. Next edit suggestions need edit-prediction support. Most chat models in the catalog are not FIM models, so check the catalog before pointing completion at Tokens. Completion fires on short pauses while you type, which adds up on a metered key and against the per-minute rate limit. Use a separate key with a spend cap.

Models Assignment#

For local models and OpenAI-compatible endpoints, JetBrains asks you to assign models to feature groups yourself under Models Assignment:

The page also shows a Context window field, which defaults to 64 000 tokens for local models. Set it to the context window listed for your model in the catalog so long chats are not cut short.

Check that it works#

Test the key and model outside the IDE first:

export TOKENS_API_KEY=tok_live_your_key
curl -s https://tokens.bd/v1/models -H "Authorization: Bearer $TOKENS_API_KEY"
curl -s https://tokens.bd/v1/chat/completions \
  -H "Authorization: Bearer $TOKENS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "deepseek/deepseek-v4.1-flash", "max_tokens": 16, "messages": [{"role": "user", "content": "Reply with OK"}]}'

Then click Test Connection in the IDE, pick the Tokens model in AI Chat and send a message. The request appears in your dashboard usage analytics.

Limits and what does not work#

• JetBrains' provider pages do not mention agents. Junie, Claude Agent and the other agents are configured separately.
• JetBrains states that AI Assistant does not support invoking tools from configured MCP servers when using local models. The page makes no such statement for OpenAI-compatible endpoints, but it also does not confirm that every tool-calling feature works there.
• If a feature has no model assigned that supports it, JetBrains can fall back to a JetBrains AI subscription, if you have one.
• The model-selector list is only as complete as GET /v1/models for your key.

Troubleshooting#

Test Connection fails. Check the URL first (see above), then the key. Run the curl commands above to separate a Tokens problem from an IDE problem. If you are on a version older than 2026.1.2, update, because older builds could replace the path.

401 invalid_api_key or missing_api_key. The key is wrong, was left empty, or was rotated. Rotation stops the old secret immediately. Paste the new key and apply again.

The model list is empty. An empty data array usually means no active plan and no wallet balance, or an allow-list that excludes every model. Subscribe or top up in billing.

404 model_not_found. The model id must be the full Tokens id, including the provider prefix and the slash, such as deepseek/deepseek-v4.1-flash.

Errors after turning Tool calling on. A 400 invalid_request can mean the upstream rejected the tool definitions. Turn the setting off or choose a model that supports tools.

402 insufficient_credits or 429 rate_limited. Top up in billing, or wait for Retry-After. Inline completion is the usual cause of rate limits.

Every code is listed in Errors, and the request format is in Chat Completions.

Sources: JetBrains AI Assistant help, Providers and API keys, Use custom models, Agents, tracker issues LLM-22911 and LLM-22721, checked October 2026.