@nathancatania/pi-glean0.1.1
Pi provider for the Glean LLM Gateway.
Untested on omp
This is a pi package. No one has reported whether it works on omp yet. If you try it, add your result below.
Reports 0
No reports yet.
Readme
pi-glean
A Pi provider for the Glean LLM Gateway.
The provider discovers the Glean model catalog at startup and routes each model to the API that Glean exposes for that model.
[!NOTE] This is a community repository. It is not endorsed, sponsored, managed, or supported by Glean. You use the contents at your own risk.
Quick Start
Requirements
- Node.js 22.19.0 or later
Install
Install the package from GitHub:
pi install git:github.com/nathancatania/pi-glean
OR
Install the package from npm:
pi install npm:@nathancatania/pi-glean
Log in
Run /login, then select Glean. Pi asks for two values.
| Prompt | Value |
|---|---|
| Glean tenant id | The label before -be.glean.com. For https://acme-be.glean.com, enter acme. |
| Glean LLM Gateway API key | The LLM Gateway API key for your tenant. Note: This is different to the Client API key. |
For headless runs, set GLEAN_LLM_API_KEY and GLEAN_TENANT_ID as environment variables instead. A stored credential takes priority over the variables.
Pi verifies the pair against the Glean model catalog https://{{GLEAN_TENANT_ID}}-be.glean.com/rest/api/v1/models. If successful, Pi stores the API key and the tenant id in ~/.pi/agent/auth.json.
Environment variables
Use the environment variables for CI and headless runs:
| Variable | Value |
|---|---|
GLEAN_LLM_API_KEY |
The Glean LLM Gateway API key. |
GLEAN_TENANT_ID |
The Glean tenant id. |
A stored credential wins over the environment variables. The provider uses GLEAN_LLM_API_KEY and GLEAN_TENANT_ID when no credential exists.
Headless and CI runs
Use the environment variables for CI and headless runs:
| Variable | Value |
|---|---|
GLEAN_LLM_API_KEY |
The Glean LLM Gateway API key. |
GLEAN_TENANT_ID |
The Glean tenant id. |
You must set both variables in the environment. Pi will use them when no stored credential exists.
export GLEAN_LLM_API_KEY="<api key>"
export GLEAN_TENANT_ID="acme"
pi install npm:@nathancatania/pi-glean
pi --print --model glean/<model id> "Summarize the failing test"
--print sends one prompt, writes the answer to stdout, and exits. Read the model ids of your tenant first:
pi --list-models glean
Model routing
Glean publishes a supportedEndpoints list for each model. Pi picks the first match in this order.
| Glean endpoint | Pi API | Request URL |
|---|---|---|
v1/responses |
OpenAI Responses | https://{tenant}-be.glean.com/rest/api/v1/openai/v1/responses |
v1/chat/completions |
OpenAI Chat Completions | https://{tenant}-be.glean.com/rest/api/v1/openai/v1/chat/completions |
v1/messages |
Anthropic Messages | https://{tenant}-be.glean.com/rest/api/v1/anthropic/v1/messages |
The Responses API wins when a model lists both OpenAI endpoints. Glean supports Chat Completions for streaming only, and Pi always streams.
Every route sends Authorization: Bearer <api_key>. The Anthropic route also sends anthropic-version: 2023-06-01. Pi removes the x-api-key header on that route, because the Anthropic SDK adds it by default.
The provider marks a Claude model for adaptive thinking when Glean lists the xhigh or max thinking level.
Authentication headers
Every route sends Authorization: Bearer <glean_api_key>.
The Anthropic route also sends anthropic-version: 2023-06-01. The provider removes the x-api-key header, because the Anthropic SDK adds it by default.
Catalog mapping
Pi converts each Glean catalog entry into a Pi model.
| Glean field | Pi field |
|---|---|
displayName |
Model name |
First modelAliases entry, then model |
Model id |
supportsReasoning |
Reasoning flag |
supportedThinkingLevels |
Thinking level map |
pricingPerInputToken, pricingPerOutputToken, pricingPerCachedInputTokenRead |
Cost per million tokens |
maxContextTokens, maxInputTokens |
Context window |
maxOutputTokens |
Output limit |
imageExtensions |
Image input |
An unlisted thinking level becomes null, so Pi clamps that level.
Pi assumes 128000 tokens when the entry carries no context size, and 8192 tokens when it carries no output limit.
Supported models
The provider keeps models that use v1/responses, v1/chat/completions, or v1/messages.
The provider currently drops models with other endpoints. This includes Gemini streamGenerateContent, Bedrock ConverseStream, realtime, audio, image, and embedding models.
Development
Test a local checkout for one run:
pi -e ./pi-glean
Run the tests:
npm test
Run the type check:
npm install
npm run check
The tests use the Node.js test runner and the Node.js type stripping. They do not need a Glean tenant.
License
MIT. See LICENSE.