Adding Models

Adding Models

This skill guides you through adding a new LLM model to Letta Code.

Quick Reference

Key files:

• src/agent/remote-model-catalog.ts - Runtime catalog loading and projection
• src/agent/model-catalog.ts - Model lookup and compatibility aliases
• .github/workflows/ci.yml - CI test matrix (optional)
• src/tools/manager.ts - Toolset detection logic (rarely needed)

Workflow

Step 1: Find Valid Model Handles

First identify the agent source. These inputs are deliberately different:

In Cloud mode, never use base/hosted rows from GET /v1/models to filter, supplement, delay, or provide a fallback for the hosted catalog. This once made GPT-4o appear in a selector even though the Cloud catalog deliberately omitted it. GET /v1/models remains necessary for organization-specific BYOK rows.

Query the Cloud hosted catalog to see hosted preset IDs, handles, and capabilities:

curl -s https://api.letta.com/v1/models/catalog | jq '.models[] | [.id, .handle]'

To inspect organization BYOK rows from a Cloud backend, query its model inventory and filter by provider_category:

curl -s https://api.letta.com/v1/models/ \
  | jq '.[] | select(.provider_category == "byok") | [.handle, .provider_type]'

Do not use this response as a second hosted catalog.

Common provider prefixes:

• anthropic/ - Claude models
• openai/ - GPT models
• google_ai/ - Gemini models
• google_vertex/ - Vertex AI
• openrouter/ - Various providers

Step 2: Update the Owning Catalog

Letta Code does not bundle a model catalog:

• Cloud hosted rows and presets come from the server's GET /v1/models/catalog response.
• Cloud GET /v1/models contributes only organization BYOK rows to selectors.
• Local model inventory comes from pi-ai and the active provider runtimes.

Add the model at the source that owns it. A hosted preset belongs in the server catalog. A local provider model belongs in pi-ai or that provider's discovery runtime.

Only change this repository when the model needs Letta Code-specific compatibility behavior, such as preserving an established CLI alias or recognizing a new provider for toolset selection. Keep that logic narrow and derive the handle and metadata from the runtime catalog rather than copying model definitions here.

Step 3: Test the Model

Test with headless mode:

bun run src/index.ts --new --model <model-id> -p "hi, what model are you?"

Example:

bun run src/index.ts --new --model gemini-3-flash -p "hi, what model are you?"

Step 4: Add to CI Test Matrix (Optional)

To include the model in automated testing, add it to .github/workflows/ci.yml:

# Find the headless job matrix around line 122
model: [gpt-5-minimal, gpt-4.1, sonnet-4.5, gemini-pro, your-new-model, glm-4.6, haiku]

Toolset Detection

Models are automatically assigned toolsets based on provider:

• openai/* → codex toolset
• google_ai/* or google_vertex/* → gemini toolset
• Others → default toolset

This is handled by isGeminiModel() and isOpenAIModel() in src/tools/manager.ts. You typically don't need to modify this unless adding a new provider.

Common Issues

"Handle not found" error: The model handle is incorrect. Run the validation script to see valid handles.

Model works but wrong toolset: Check src/tools/manager.ts to ensure the provider prefix is recognized.