Provider and local model setup

Choose an access method

Provider type selects subscription access, an official API, or Custom Gateway. Available subscription choices depend on the active agent framework. The captured Codex setup shows Codex subscription, xAI OAuth, official APIs, and Custom Gateway; do not assume another framework presents the same choices.

Choice	What you need	Check before continuing
Codex subscription	A compatible Codex sign-in	Inspect the Codex authentication choice; importing existing access copies authentication into Open-Science
Official API	Access to that provider and the requested model	Confirm provider, region when applicable, and API credential
Custom Gateway	A compatible endpoint, exact model ID, and an API key when required	Confirm the API format and supported model features with the gateway operator

Choose Import existing Codex sign-in to copy a working local sign-in into Open-Science. The import can include a compatible non-secret loopback route; other global configuration, Skills and sessions stay separate. Under Advanced settings → Transport, keep Auto (recommended) unless your connection requires a different transport.

Choose a provider region or a free catalog model

For SenseNova, select China or Global in the provider form before choosing a model. Use the API key for that region, review the resulting model list, select Save, and wait for connection validation before the change is committed. Switching regions can change both the endpoint and available models; a key or model name from the other region may not work.

For gateways such as OpenRouter or OpenCode Zen, select a free model only when that exact entry is offered for the active framework. Use the account and credential required by the service. A free catalog entry does not remove usage limits or establish support for every tool or image input. Do not append :free to an arbitrary model ID. Send a small request and check the returned model and result before using the connection for research.

Connect an existing Codex subscription

1. Open Settings → Model → Add provider.
2. Set Provider type to Codex subscription.
3. In Codex authentication, choose Import existing Codex sign-in. This requires a usable sign-in on this computer. It copies authentication into the application profile; it does not import your other Codex sessions or Skills.
4. Select Save. Wait while the provider row shows Testing…; a saved row alone is not the success check.
5. Confirm Connection verified and Authentication imported into Open-Science in the provider row. The released interface may display the product name without a hyphen.
6. In Main model, select an available subscription model. For example, select an available gpt-5.6-sol entry if your account offers it. Check the model name and provider together, especially when multiple providers offer similarly named models.
7. Open a project and send a bounded request. A connection test verifies authentication, while an actual response verifies the request path. Confirm the response and any tool-permission request appear in that session.

Provider-row control	Use it when	Success check
Check Codex login	The saved connection may have expired.	The pending check settles into the displayed verified or failed state.
Re-import Codex login	You have refreshed the external sign-in and want to update the application copy.	Authentication is imported and checked again.
Edit	You need to review authentication or transport settings.	Select Save and wait for successful validation before the edit is committed.
Delete	An unused provider should be removed.	Availability depends on whether the provider is still required; an active dependency can prevent deletion.

If import reports that a file-backed Codex login is missing, sign in through the supported Codex flow and retry Re-import Codex login. A login held only in an external credential store is not necessarily an importable file.

Do not interpret Testing… as failure, or Connection verified as proof that every listed model and tool can run. If import fails, complete the supported Codex sign-in flow and retry; do not paste authentication JSON into a prompt or documentation.

The agent runtime runs the work; the model provider supplies the model. Installing Codex does not automatically connect a provider. In first-time setup, this page follows Agent runtime. After setup, open Settings → Model to manage provider access.

Update or remove an API credential

After changing a key at the service, find its provider in Settings → Model, select Edit, enter the replacement in API key, and select Save. Leaving this field blank keeps the existing key; it does not clear it. The connection is tested before the edit is committed. If authentication fails, check the endpoint, the account the key belongs to, and its validity before retrying.

After Connection verified, complete a small request with that provider. Remove an unused provider with Delete, checking its name in the confirmation. Removing the application configuration does not revoke the key at the service.

Custom Gateway: every visible field

Start by selecting Custom Gateway. Changing the provider type can preserve the display name from the previous selection, so review the name instead of assuming it was reset.

Field or control	Input and behavior
Provider type	Selects the provider family and changes the visible form
Name / Provider name	Optional display name, such as Lab gateway; not a model identifier
Base URL	Required gateway base address. Remote model endpoints require HTTPS; HTTP is allowed for localhost and loopback addresses. Use your operator’s actual address, not the nonworking https://gateway.example placeholder.
API format	Select Chat Completions, Messages, or Responses; the displayed route helps identify the corresponding protocol
API key	Required for remote gateways; optional for a local loopback gateway that needs no authentication. Enter a real credential if your local server requires one
Eye / Show API key	Toggles visibility of the current key input; keep it hidden before capturing or sharing the screen
Model	Required exact model identifier accepted by the endpoint; the screenshot's demo-model is only a placeholder
Context window	Optional model context limit; blank requests the provider default
Context presets	32K, 64K, 128K, 200K, 256K, 1M; selecting 128K fills 128000
Advanced settings	Expands or collapses capability and token-limit fields
More information (i)	Opens contextual help beside the associated label
Back	Returns to Agent runtime; the wizard owns the form draft so it can survive navigating back
Test & continue	Validates required fields, then tests the provider before committing valid settings; advances after a successful applicable validation

The three API formats shown in the menu are:

• Chat Completions — /v1/chat/completions.
• Messages — /v1/messages.
• Responses — /v1/responses.

These are protocol choices, not instructions to append every listed route to the Base URL. A gateway can support one format without supporting the others.

Advanced fields and conditional controls

An older remote HTTP configuration remains editable but cannot send requests. Obtain an HTTPS endpoint from the service operator, save it and test again. A local loopback model server can keep its HTTP address; a remote LAN server still needs HTTPS.

Advanced fields and conditional controls

Field or control	How to set it
Image input	Enable only if both gateway and selected model accept image content
Thinking mode	Enable only if the gateway/model accept thinking or effort controls
Supported effort levels	Appears with thinking enabled; select the levels actually supported, rather than inferring them from a model name
Reasoning request format	Appears for Chat Completions with thinking enabled; select how the gateway expects effort parameters
Maximum input tokens	Optional separate input limit; blank uses the provider default. Presets: 32K, 64K, 128K, 200K, 256K, 1M
Maximum output tokens	Optional separate output limit. Presets: 4K, 8K, 16K, 32K, 64K, 128K

Enable Thinking mode to configure supported effort levels. For Chat Completions, also select the reasoning request format supported by your endpoint. These declarations must match the provider's API capabilities.

Test the gateway configuration

1. Select Custom Gateway and expand Advanced settings.
2. Enter the Base URL and exact model ID supplied by the gateway operator, plus an API key when required. Missing required fields produce inline errors and keep you on this page.
3. Enter a recognizable display name. For a real connection, enter the actual endpoint and model supplied by your provider; demonstration placeholders cannot pass a connection test.
4. Choose a context preset and confirm the numeric value.
5. Enable Thinking mode only when supported, then inspect the newly visible effort fields. Changing API format may change the available fields.
6. If a key is required, enter it privately and keep it hidden. Select Test & continue when you are ready for a provider request.
7. Wait for the result. Testing connection… indicates pending validation; repeated clicks are disabled. Subscription flows instead use Sign in & continue, Waiting for sign-in…, and Cancel sign-in where applicable.

Connect a local model endpoint

Example Connect a local Qwen model through Ollama

A local model server runs separately from Open-Science. Choose Custom Gateway for a compatible endpoint, and use an Agent that supports its API format. The example below uses Ollama with OpenCode. Installing a Python Notebook interpreter does not install a model server.

Start the server and download the model

Install Ollama, then start a local-only test server in a terminal:

OLLAMA_HOST=127.0.0.1:11435 OLLAMA_CONTEXT_LENGTH=32768 ollama serve

Keep that terminal open. In another terminal, download the model to that server:

OLLAMA_HOST=127.0.0.1:11435 ollama pull qwen3:0.6b

Wait for the download to finish. If the server is running but the requested model is absent, Open-Science can report Test failed: the configured model was not found. Finish the download, confirm the exact model ID, and select Test connection again.

Enter the provider settings

Open Settings → Model → Add provider and enter:

Field	This local connection example
Provider type	Custom Gateway
Name	Local Qwen demo
Base URL	http://127.0.0.1:11435
API format	Chat Completions (/v1/chat/completions)
API key	Leave blank for this unauthenticated loopback endpoint; use the actual credential for an authenticated gateway
Model	qwen3:0.6b
Context window	32768, matching the running server
Advanced settings → Maximum output tokens	4096
Image input / Thinking mode	Off for this connection check

The form appends /v1 to the gateway root. Open-Science accepts a blank API key for loopback addresses such as localhost, 127.0.0.1 and [::1]; the older screenshot may show a placeholder. A remote or LAN gateway still requires HTTPS and an API key. Use the API format your local server supports.

Set an output budget that leaves room for input and conversation history. OpenCode reserves an output budget when this field is blank; a large reserve can cause repeated compaction in a small context window. The declared context window also needs to match the model server’s allocation. Changing the form alone does not change Ollama’s runtime configuration.

Select a compatible Agent and check a reply

In Settings → Agent, install OpenCode → App-managed download if it is missing, then select its card and confirm Switch. Return to Model and select the local model. Start a new conversation with a short connection-only request before using it for research. Check that the request actually finishes; a saved provider or a successful connection test alone does not establish reliable scientific reasoning, tool use or image support.

The connection check completed with Local model connected. using the configured local endpoint and OpenCode. It verifies a text request, not a biomedical analysis.

Keep the server running while using the model. For an Agent on another host, localhost refers to that host. A browser reaching an endpoint does not prove the Agent can reach it.

Check an actual tool call

Example Check a local model’s Notebook tool call

After confirming a connection, use a small task with a known result to test the tool path. Ask the agent to execute this through the Python Notebook instead of returning mental arithmetic:

print(8664 + 18515)
print((8664 + 18515) == 27179)

These are the first GSE60450 sample's zero-count and detected-gene counts. Inspect the proposed code in the permission panel, approve it, then open Notebook and verify 27179 / True.

Local qwen2.5:7b completed this call through the Codex framework and a local Chat Completions endpoint. Its initial proposal referenced an unavailable helper module; the check succeeded after declining that proposal and specifying the dependency-free code above. This verifies a bounded tool operation, not reliable planning of a complete RNA-seq analysis or equivalent behavior under another Agent framework.

If setup does not advance

Symptom	Next check
Required-field messages	Complete the named fields; a display name alone is insufficient
Secure key storage unavailable	Unlock or authorize the operating-system credential vault; keys cannot be saved until it is available
Connection/authentication failure	Check the credential, endpoint, format, and access to the specific model
Provider changed during testing	Review the current provider and test again; a superseded result must not complete setup
Sign-in canceled	Start again when ready; cancellation is not a successful connection
Installed runtime but no usable provider	Finish model connection; runtime installation and provider authorization are separate

HTTP error lookup

For 400, 401, 403, 404, 429 or 5xx responses, use the HTTP troubleshooting table. Keep the responding service and its detailed message with the status code.

Source: ProviderForm.tsx, ProviderStep.tsx.

Save a provider change in v0.31.0 and later

Provider edits are tested before they are committed. Select Save, wait for the connection result, and confirm success before closing the form. A failed test does not replace a working saved configuration. If a previously saved connection is rejected during a request, its availability is updated; check the credential and endpoint, then test again. Conversation models, Classification models and Local parsing models have different purposes; see model settings.

StepFun and region selection

Choose StepFun in the provider catalog, confirm China or Global, then select a model and supply credentials for that region. v0.32.0 adds Step-5 Preview, with multimodal and 1M-context catalog metadata. Actual model access, quota and input support still depend on the provider account and selected Agent's compatibility.

Save and check the connection before choosing it in a conversation. Existing provider configurations retain their previous endpoint; updating the app does not switch their region or Main model.

Updated provider catalogs

The v0.33.0 catalog adds Xiaomi MiMo v2.6 models and xAI Grok 4.7, and refreshes OpenCode Zen and Go. New MiMo configurations default to mimo-v2.6-pro; existing v2.5 choices remain available. Grok 4.7 becomes xAI’s new default while previous model IDs remain listed.

Select the exact model offered for your active Agent, check its input and reasoning options, and run a small request. A catalog default does not change the model in every existing session or guarantee your account has access. Codex can remain Main while you configure a classification service separately.

New catalog choices

The v0.33.1 provider catalogs add GPT-6 and Claude Opus 5.5. Open the provider's model list and choose an entry supported by the active framework and your account. Catalog presence does not grant access or switch the saved Main model. Validate the provider and send a small request before using a changed model for research.

MiniMax M3.1 Flash is the released multimodal model with a 1M-token context, available in both MiniMax regions. Choose the region that issued your key, select the model and test the saved configuration. v1.0.2 fixes its built-in connection through Codex. Claude Sonnet 5.5 remains in the provider catalog. Selecting a catalog entry does not automatically replace the saved Main model.

v0.34.1 adds gpt-6.1-sol to the Codex model catalog. Select it explicitly in the model picker, then send a small request to confirm account access. Updating the app does not automatically replace the saved Main model.

v0.35.0 expands the OpenCode Zen catalog and adds Jev classification support. Check the role-specific picker: a model available for a classification task is not automatically the Main model. Keep your configured Main provider unless you deliberately change it; classification setup is covered in Models.

Add Requesty

1. Open Settings → Model → Add provider.
2. In Provider type, choose Requesty under Official API. This built-in entry supplies its endpoint; a Custom Gateway URL is not required.
3. Check Name and enter your Requesty key in API key. Get an API key opens Requesty's key-management page. Keep the key hidden.
4. Review Supported models, then select Test connection or Save and wait for validation. A missing key leaves these controls unavailable.
5. After the provider is saved successfully, select an offered model in the intended model picker and send a small request. Check the actual response before starting research.

Requesty uses https://router.requesty.ai; the active Agent determines which compatible protocol and models are offered. Access and quota depend on the Requesty account. Adding this provider does not automatically replace the saved Main model. v0.35.1 also refreshes the OpenRouter catalog; use the current picker rather than copying a model ID from an old screenshot.

Recover a model connection

Use the recovery action attached to an expired or invalid credential, update the correct provider entry, then test it before returning to the same session. The Composer is blocked only while recovery requires your action. Preserve the draft and check whether the previous turn is still running before resending. For Codex subscription models, context usage reflects the selected model's reported capacity; it can differ from an older displayed estimate.

Select Kimi K3.1 Pro

v1.0.3 adds Kimi K3.1 Pro models to the provider catalog. Open the intended provider's model picker and select an available entry, then test the connection. Model availability still depends on the active Agent framework, endpoint and account; a catalog entry alone does not establish access.