Code graph model providers
The providers that serve each default code graph model, and how the allowed providers and zero data retention settings narrow them.
The code graph sends pieces of your code to AI models when it judges links between nodes, names domains, and writes summaries. If your security policy limits who may process source code, you need to know which companies serve those models before you set the allowed providers. This page lists them for each default model.
The model calls go through OpenRouter, which routes each call to a provider that hosts the model. The lists below are the providers OpenRouter showed on 12 November 2026. They change as providers join and leave, so read OpenRouter's page for the model again before you rely on a list. Each model's providers are in OpenRouter's public model data at https://openrouter.ai/api/v1/models/<model>/endpoints.
Default models
The defaults come from the codegraph/models.toml and codegraph/embeddings.toml settings.
| Task | Default model | Used when |
|---|---|---|
| Linker | GLM 5.3 Flash | It judges the links the build suggests. |
| Stronger model | GLM 5.3 | The linker could not settle a pair across repositories. |
| Build time and summaries | The linker model | It names domains, writes briefs, drafts decisions, and writes summaries a caller asks for. |
| Embeddings | voyage-4-large | A build embeds cards for meaning search, when embeddings are on. |
Providers for each model
Read on 12 November 2026 from OpenRouter's endpoint list for each model.
GLM 5.3 Flash
Standard calls can reach 37 providers: AtlasCloud, BaseTen, Cohere, CoreWeave, Crusoe, Darkbloom, DeepInfra, DekaLLM, DigitalOcean, Fireworks, Friendli, GMICloud, Inceptron, InferenceNet, KT, Makora, Mistral, Modal, Morph, Near AI, Nebius, NextBit, Novita, OpenInference, Parasail, Particle AI, PrimeIntellect, Reka, Relace, Sail Research, SiliconFlow, StreamLake, Tenstorrent, Together, Venice, Wafer, and Z.AI.
Batch calls (z-ai/glm-5.3-flash:batch) can reach two: DeepInfra and Sail Research.
GLM 5.3
Standard calls can reach 36 providers: AkashML, Alibaba, Amazon Bedrock, AtlasCloud, Baidu, BaseTen, Cloudflare, CoreWeave, Crusoe, Decart, DeepInfra, DigitalOcean, Fireworks, Friendli, GMICloud, Inceptron, InferenceNet, Makora, Mistral, Modal, Morph, Nebius, Novita, Parasail, PrimeIntellect, Reka, Relace, River, Sail Research, SambaNova, SiliconFlow, Tenstorrent, Together, Venice, Wafer, and Z.AI.
Batch calls (z-ai/glm-5.3:batch) can reach two: DeepInfra and Morph.
voyage-4-large
Embeddings do not go through OpenRouter. The Oxagen provider sends them to Voyage AI on Oxagen's key, and Voyage AI is the only provider. With the custom provider, your own endpoint serves the embeddings and the allowed providers setting does not apply to it. With embeddings off, nothing is sent to an embedding provider.
Narrow the list
allowed_providers and zero_data_retention in codegraph/models.toml cut these lists down. You change both in AI models, which opens a steering pull request on the file.
allowed_providersnames the OpenRouter providers that may serve calls, using the names above. A call goes only to a provider on both the model's list and yours. Left unset, every provider on the list may serve it. It does not apply to embeddings.zero_data_retentionlimits calls to providers that keep no copy of the request. When both settings are on, a call must satisfy both, so the providers left are the ones on your list that also keep no data.
Zero data retention leaves out a few providers. On 12 November 2026, OpenRouter's zero data retention list covered 31 of the 37 GLM 5.3 Flash providers and 31 of the 36 GLM 5.3 providers.
Before you set them:
- Batch needs a provider you allow. Batch calls run on the two providers listed for the model. If your allowed providers leave both out, the batch calls fail and so do the linker and enrich rounds that send them. Allow one of the two batch providers, or set the call mode to standard.
- Zero data retention means standard calls. OpenRouter offers no batch endpoint that keeps no data, so a workspace with zero data retention on makes standard calls at the full price.
Check each model against your list, so that every default model keeps at least one provider.
Code graph from the CLI
Build a repository's code graph on your machine, serve it to your agents over MCP, and upload CI reports to the Oxagen code graph.
Configuration
Environment configuration for running and deploying Oxagen: transactional email (SMTP), local-development behavior, and the storage driver.