Migrate templates to Model Gateway — remove separate model API keys, bump SDKs

[jay-sahnan]

Why

Model Gateway launches April 2nd, routing LLM requests through Browserbase so users only need their BROWSERBASE_API_KEY. No more juggling separate OpenAI/Anthropic/Google keys per template.

What

• TypeScript (5 files): Collapsed model: { modelName, apiKey } objects to model: "string"
• Python (~27 files): Removed model_api_key parameter from Stagehand/AsyncStagehand/StagehandConfig constructors, including variable declarations and validation blocks
• SDK bumps: @browserbasehq/stagehand → ^3.1.0, stagehand (Python) → >=3.19.0, browserbase (Python) → >=1.7.0
• .env.example files (~16): Removed model API key entries
• READMEs (~40): Removed model key setup instructions, error messages, and pitfall entries. Added Model Gateway section to root README.
• Error messages/comments (~30 files): Removed "Verify OPENAI_API_KEY" / "Verify GOOGLE_API_KEY" hints from catch blocks

Notes

• CUA templates excluded: gemini-cua, company-address-finder, business-lookup, microsoft-cua — CUA models are not supported by Model Gateway at launch
• Other exclusions: cerebras-docs-checker (unsupported provider), go/hackernews (Go SDK not in scope)
• cartesia-form-filling: Removed Stagehand model_api_key but kept GEMINI_API_KEY for its separate genai.Client (non-Stagehand usage)
• Pre-existing ruff lint errors in amazon-global-price-comparison, exa-browserbase, sec-filing-research were fixed as part of this PR since the linter runs on all staged .py file

---

Note

Medium Risk
Broad, cross-template changes plus SDK version bumps could break example runtimes if any template still relies on removed model_api_key/project ID wiring or older Stagehand/Browserbase APIs.

Overview
Updates templates to use Browserbase Model Gateway so most examples only require BROWSERBASE_API_KEY, removing MODEL_API_KEY/provider-key setup, validation, and troubleshooting text across Python/TypeScript/Go templates.

Refactors Stagehand client initialization accordingly (dropping model_api_key and often browserbase_project_id arguments, and collapsing TS model config to a routed model string), and bumps dependencies (stagehand>=3.19.0, browserbase>=1.7.0, and TS packages to newer/latest versions). Root README.md adds a Model Gateway section and notes that CUA models remain excluded and still require provider keys.

^{Reviewed by Cursor Bugbot for commit a90a164. Bugbot is set up for automated code reviews on this repo. Configure here.}

[Kylejeong2]

we no longer need project_id right? or is python not the same?

[jay-sahnan]

Yep

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 548cc58. Configure here.}

[cursor]

Prompt newlines lost in string reformatting

Medium Severity

The extract instruction was refactored from a triple-quoted f-string (with newlines between each numbered item) to concatenated string literals, which silently strips all newlines. The old prompt had each extraction field on its own line with a blank line before the "Only extract…" closing note; the new version joins everything into one continuous line. This makes the numbered list significantly harder for the LLM to parse and may reduce extraction quality.

 

^{Reviewed by Cursor Bugbot for commit 548cc58. Configure here.}

[cursor]

Extra blank line added to agent system prompt

Low Severity

A blank line was inserted into AGENT_SYSTEM_PROMPT between the "For cover letter…" and "For location/relocation…" bullet points during reformatting. The original prompt had these as consecutive items in a single bullet list. This changes the actual prompt content sent to the LLM, likely unintentionally since this PR's goal was removing API keys, not modifying prompts.

 

^{Reviewed by Cursor Bugbot for commit 548cc58. Configure here.}

[Kylejeong2]

pre-approving