feat: add Cursor Bridge provider

[Mushro0mCloud]

Summary

• add Cursor Bridge as an OpenAI-compatible provider
• register all 35 model IDs currently exposed by the live Cursor catalog, including Claude Fable 5, Claude Opus/Sonnet, GPT/Codex, Gemini, Grok, Kimi, GLM, Composer 2.5, and Cursor Default
• include model-specific reasoning effort controls backed by Cursor variants
• include a monochrome provider logo compatible with light and dark themes

Why

Cursor Bridge lets users access models available to their Cursor account through an OpenAI-compatible API. Adding it to models.dev makes the provider and its model catalog selectable in OpenCode without requiring a custom provider definition.

Cursor Bridge is unofficial community software and is not affiliated with or endorsed by Cursor.

Provider details

• API: https://cursorbridge.dev/v1
• Authentication: CURSOR_API_KEY
• Documentation: https://cursorbridge.dev/docs#model-metadata
• SDK: @ai-sdk/openai-compatible

Model metadata

• 33 upstream models use canonical base_model metadata
• default and Cursor-proprietary composer-2.5 are defined directly
• 20 models declare verified low, medium, high, xhigh, and/or max effort variants
• models with fixed reasoning behavior use an empty reasoning-options list
• fixed token costs are intentionally omitted because availability and billing depend on the authenticated Cursor account and plan

Provider evidence

Authenticated GET /v1/models/{model} responses expose the exact context_window, max_output_tokens, reasoning, reasoning_options, and applicable release_date values represented in this PR:

curl -H "Authorization: Bearer $CURSOR_API_KEY" https://cursorbridge.dev/v1/models/claude-fable-5

Each TOML with configurable reasoning links directly to its corresponding endpoint. The public documentation defines the response fields and explains that effort values come from Cursor usable-model variants. Unsupported effort values are rejected rather than silently substituted.

For Cursor-owned virtual models, release_date is documented as the first date that ID was observed in the live Cursor Bridge catalog.

Validation

• verified the authored catalog exactly matches the 35 model IDs observed across live Cursor account catalogs
• programmatically compared all 20 TOML effort arrays with their live authenticated endpoint responses
• verified inline limits and dates against the live default and composer-2.5 endpoint responses
• validated all TOML files parse and every base_model target exists
• validated required resolved model fields and reasoning declarations
• validated SVG structure and currentColor requirements
• git diff --check passes

Full repository validation runs in GitHub Actions because Bun is not installed in the local environment.

[github-actions]

I have enough to finalize. All 35 base_model targets exist, every referenced base model carries reasoning = true (so reasoning_options = [] is schema-valid), the inline default/composer-2.5 files have all required fields, family = "auto" is valid, and the logo/provider TOML are compliant. The substantive issues are the reasoning-surface claim and citation mapping. Comparable OpenAI-compatible aggregators (Requesty, Poe) proxying these same models declare effort controls and document the request format, which contradicts the blanket [] claim.

Action items

• [medium] [possible mistake] providers/cursor-bridge/models/claude-opus-4-6.toml:2 (applies to all 35 model files) - Check: reasoning_options must describe the inference provider's actual accepted request surface, not merely the upstream model. Why: Every reasoning model is declared reasoning_options = [] ("no verified control"), but Cursor Bridge is an OpenAI-compatible endpoint where reasoning_effort / reasoning.effort (or a thinking budget) is the standard reasoning surface, and comparable OpenAI-compatible aggregators proxying these same models (Requesty: effort + budget_tokens; Poe: effort) declare controls and document the request format in provider.toml. The PR body asserts "no provider-wide reasoning control is guaranteed" without evidence of inspecting Cursor Bridge's docs/API or testing the field; if a control is forwarded, opencode users lose the ability to steer reasoning. Action: Verify via Cursor Bridge's docs/API whether reasoning_effort / reasoning.effort / a thinking budget is accepted and effective for these models. If so, replace [] with a verified effort (and/or budget_tokens) option and document the request syntax in provider.toml as Requesty/Poe do. If no effective control exists, cite the docs or test result that establishes that.
• [low] [possible mistake] providers/cursor-bridge/models/composer-2.5.toml:2 and providers/cursor-bridge/models/default.toml:4 - Check: Each citation must state what it supports, and material provider-specific facts need first-party evidence. Why: composer-2.5.toml cites https://cursorbridge.dev/v1/models for context = 272_000, output = 32_768, release_date = "2026-07-09", and capabilities, but a /v1/models endpoint lists model IDs, not limits or dates. default.toml cites the docs URL without mapping it to its specific limits/dates. The PR body lists API/docs URLs but does not map them to the 35-model catalog claim or the inline limits, so these factual claims cannot be reviewed from the provided citations. Action: Add first-party citations that support the specific context/output limits, release/last-updated dates, and capability flags for default and composer-2.5, and state what each cited URL proves (including confirmation that the 35 model IDs match the live Cursor catalog).

[github-actions]

Action items

• [high] [possible mistake] providers/cursor-bridge/models/claude-opus-4-6.toml:2 (and the 19 other files changed in commit 3) - Check: reasoning_options effort values must reflect Cursor Bridge's actual inference API surface with provider evidence. Why: The PR body states reasoning models declare reasoning_options = [] because "no provider-wide reasoning control is guaranteed across Cursor plans and model variants," but commit 3 replaces [] with per-model effort arrays. No citation in the PR body or TOML maps to any claimed value, and the audit-reasoning-options skill requires explicit evidence for xhigh and max. The per-model variation is unexplained and suspicious: claude-opus-4-6 = ["high","max"] while claude-opus-4-7/4-8 = the full 5-level set; gpt-5.1 = 3 values but gpt-5.2 = 4; glm-5.2 = ["high","max"] here vs ["high","xhigh"] on OpenRouter. Tools consuming models.dev would send unsupported effort values if these are wrong. Action: Provide Cursor Bridge API documentation or reproducible request evidence proving each claimed effort value is accepted for that specific model ID, or revert all 20 files to reasoning_options = [] to match the PR body's stated rationale.
• [low] [possible mistake] providers/cursor-bridge/models/composer-2.5.toml:14 and providers/cursor-bridge/models/default.toml:18 - Check: Citations must state what they support for material factual claims. Why: composer-2.5 cites https://cursorbridge.dev/v1/models and default cites https://cursorbridge.dev/docs, but neither maps to the specific context/output limits (composer-2.5 context=272_000/output=32_768, default context=200_000/output=64_000) or release dates. An OpenAI-compatible /v1/models endpoint typically lists model IDs only and does not expose token limits, so the cited source may not support these values. Action: Provide a source (model docs page or API response field) that specifically supports each model's limit values and release date, and state what each link proves.

[Mushro0mCloud]

Addressed both action items in 8922bf6b1 and updated the PR description.

Reasoning-option evidence

Cursor Bridge now documents its model metadata contract at https://cursorbridge.dev/docs#model-metadata. Authenticated GET /v1/models/{model} responses return the exact accepted reasoning_options for that model ID, derived from Cursor usable-model variants. Unsupported effort values are rejected by the inference routing layer.

Reproduction example:

curl -H "Authorization: Bearer $CURSOR_API_KEY"   https://cursorbridge.dev/v1/models/claude-fable-5

That response currently reports:

{"reasoning_options":["low","medium","high","xhigh","max"]}

All 20 configurable model TOMLs now cite their model-specific endpoint. I also programmatically compared every TOML effort array against its live endpoint response; all 20 match. Models with fixed reasoning behavior retain reasoning_options = [].

Inline limits and dates

The docs define the metadata fields and clarify that release_date for Cursor-owned virtual models means the first date the model ID was observed in the live Cursor Bridge catalog. The cited authenticated endpoints now report the exact authored values:

• /v1/models/default: context 200,000; output 64,000; first observed 2026-07-13
• /v1/models/composer-2.5: context 272,000; output 32,768; first observed 2026-07-09

The comments in both TOMLs now state exactly what each citation supports. The stale PR wording about having no provider-wide reasoning control has also been removed.

[github-actions]

The repository checkout is the trusted base revision, so the new files don't exist yet — my analysis is based on the diff content. I have completed my review of all changed files, the schema, base model metadata, conventions, and documentation.

Action items

• [low] [possible mistake] providers/cursor-bridge/models/default.toml:6-13, providers/cursor-bridge/models/composer-2.5.toml:5-12 - Check: Cited evidence must cover the capability and modality fields declared for the two directly-defined models, not only limits and dates. Why: The leading comment citations explicitly state that the per-model API URLs "report the 200k context, 64k output limit, and 2026-07-13 first-observed release date" (default) and "report the 272k context, 32,768 output limit, and 2026-07-09 first-observed release date" (composer-2.5). They do not claim to cover attachment, reasoning, tool_call, structured_output, temperature, or modalities — all material factual claims that cannot be reviewed without a source. Action: Update the citation comments to state that the per-model API responses also report these capability/modality values, or add separate citations (e.g., provider docs or API response excerpts) that support each capability value.

• [low] [violation] providers/cursor-bridge/ - Check: AGENTS.md recommends adding a sync module for new providers whose API is context-rich enough to authoritatively track model data. Why: The PR body and per-model citations describe a provider API (cursorbridge.dev/v1/models/<model>) that returns model IDs, per-model reasoning effort variants, token limits, and release dates — rich enough to authoritatively track model availability and reasoning_options over time while preserving hand-authored capability and cost fields, matching the partial-sync pattern used by the xAI and Google sync modules in sync.md. Action: Consider adding a sync module (packages/core/src/sync/providers/cursor-bridge.ts) that fetches the model catalog, updates model IDs / reasoning_options / limits, and deletes models no longer served, while preserving hand-authored capabilities and omitting cost (which the PR notes is account/plan-dependent). This is a recommendation, not a merge blocker.