feat: add Avian as LLM provider

[avianion]

Summary

Adds Avian as a new LLM provider. Avian is an OpenAI-compatible inference API offering competitive pricing on frontier models.

Models:

Model	Context Window	Input (per 1M)	Output (per 1M)
deepseek/deepseek-v3.2	164K	$0.26	$0.38
moonshotai/kimi-k2.5	131K	$0.45	$2.20
z-ai/glm-5	131K	$0.30	$2.55
minimax/minimax-m2.5	1M	$0.30	$1.10

Features supported:

• Chat completions
• Streaming responses
• Function/tool calling
• Temperature control

Changes

• components/icons.tsx — Add AvianIcon SVG component
• providers/types.ts — Add 'avian' to ProviderId union type
• providers/models.ts — Add Avian provider definition with model pricing and capabilities
• providers/avian/index.ts — Provider implementation using OpenAI SDK (follows DeepSeek pattern)
• providers/avian/utils.ts — Streaming utility using shared createOpenAICompatibleStream
• providers/registry.ts — Register avianProvider in the provider registry
• providers/utils.ts — Add Avian to provider metadata map

Implementation

The provider follows the same pattern as other OpenAI-compatible providers (DeepSeek, Cerebras). It uses the openai npm package with baseURL: 'https://api.avian.io/v1' and supports all standard features including tool calling loops, forced tool usage, and streaming with cost tracking.

Authentication is via user-provided API key (AVIAN_API_KEY as Bearer token).

Test plan

• Verify Avian models appear in model selector dropdown
• Test chat completion with deepseek/deepseek-v3.2 model
• Test streaming response
• Test function calling with a tool-enabled workflow
• Verify cost tracking displays correct pricing

cc @waleedlatif1 @emir-karabeg

[vercel]

@avianion is attempting to deploy a commit to the Sim Team on Vercel.

A member of the Team first needs to authorize it.

[greptile-apps]

Greptile Summary

Added Avian as a new LLM provider with OpenAI-compatible API integration. The implementation follows the established DeepSeek provider pattern, reusing proven architecture for OpenAI SDK-based providers. The integration includes proper tool calling support, streaming responses, cost tracking, and provider metadata registration. All 4 models are correctly configured with pricing and context window specifications.

Confidence Score: 5/5

• This PR is safe to merge with no concerns
• The implementation is a well-executed pattern replication of the existing DeepSeek provider with appropriate naming changes. All imports follow established conventions, model definitions are accurate, and the provider is properly registered across all necessary files. No logic changes or risks introduced.
• No files require special attention

Important Files Changed

Filename	Overview
apps/sim/providers/avian/index.ts	New provider implementation following established DeepSeek pattern with OpenAI SDK, supports tool calling, streaming, and cost tracking
apps/sim/providers/models.ts	Added Avian provider definition with 4 models, pricing, and configuration
apps/sim/providers/registry.ts	Registered avianProvider in the provider registry map
apps/sim/providers/types.ts	Added 'avian' to ProviderId union type
apps/sim/providers/utils.ts	Added Avian to provider metadata map using buildProviderMetadata

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[User selects Avian provider] --> B[Provider Registry]
    B --> C[avianProvider from /providers/avian/index.ts]
    C --> D[Initialize OpenAI SDK]
    D --> E[baseURL: api.avian.io/v1]
    E --> F{Request Type?}
    F -->|Streaming + No Tools| G[Create stream response]
    F -->|With Tools| H[Tool execution loop]
    F -->|Standard| I[Standard completion]
    H --> J[Execute tools with prepareToolsWithUsageControl]
    J --> K[Track forced tool usage]
    K --> L[Iterate up to MAX_TOOL_ITERATIONS]
    L --> M{Stream final?}
    M -->|Yes| G
    M -->|No| N[Return response with timing]
    G --> O[Calculate costs using model pricing]
    I --> O
    N --> O
    O --> P[Return to user with tokens & cost]

Loading

_{Last reviewed commit: f77f80c}

[greptile-apps]

_{7 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[avianion]

Friendly follow-up — this PR is still active and ready for review. Would appreciate a look when you get a chance! cc @waleedlatif1 @emir-karabeg

[avianion]

Friendly follow-up — this PR is still active and ready for review. All feedback has been addressed. Would appreciate a look when you get a chance! cc @waleedlatif1

[avianion]

Hey @waleedlatif1 @emir-karabeg — friendly follow-up on this PR. Avian is an OpenAI-compatible inference provider that's already live and powering apps like ISEKAI ZERO. This is a lightweight integration (standard OpenAI-compatible endpoint) and we're happy to address any feedback or make adjustments. Would love to get this merged if you have a moment to review. Thanks!

[cursor]

PR Summary

Medium Risk
Large new executor (~580 lines) on the inference path with tool loops and streaming, but it is additive and mirrors existing OpenAI-compatible providers rather than changing auth or shared security logic.

Overview
Adds Avian as a first-class LLM provider so workflows can call Avian’s OpenAI-compatible API (https://api.avian.io/v1) with user-supplied API keys.

The new avianProvider implements chat completions with streaming (no-tools path and post-tool final response), function/tool calling with usage control and forced-tool sequencing, token/cost tracking, and trace timing—aligned with other OpenAI-compatible providers (e.g. DeepSeek). A thin createReadableStreamFromAvianStream wraps the shared createOpenAICompatibleStream helper.

Catalog & UI wiring: ProviderId gains 'avian'; PROVIDER_DEFINITIONS lists four models (DeepSeek v3.2, Kimi K2.5, GLM-5, MiniMax M2.5) with pricing, context windows, and toolUsageControl; AvianIcon is added; the provider is registered in registry.ts and client metadata in providers/utils.ts.

^{Reviewed by Cursor Bugbot for commit f885d73. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
docs	Skipped		Jun 5, 2026 10:44am

[cursor]

~560-line provider duplicates DeepSeek implementation verbatim

Low Severity

The entire avian/index.ts (~560 lines) is a near-verbatim copy of deepseek/index.ts, differing only in the provider name string and base URL. The avian/utils.ts wrapper is also a trivial one-liner delegating to createOpenAICompatibleStream, identical in structure to deepseek/utils.ts, groq/utils.ts, etc. Given that a shared createOpenAICompatibleStream utility and checkForForcedToolUsageOpenAI helper already exist in providers/utils.ts, this full duplication increases maintenance burden — any future bug fix to the tool-calling loop, streaming, or cost tracking logic needs to be replicated across every copy-pasted provider.

Additional Locations (1)

• apps/sim/providers/avian/utils.ts#L1-L14

 

[avianion]

Re: Bugbot feedback on ~560-line duplication with DeepSeek

Investigated the duplication concern. The Avian provider is indeed a near-verbatim copy of DeepSeek (only the provider name and base URL differ), but this is the established pattern across the entire codebase — every OpenAI-compatible provider (DeepSeek, Groq, xAI, Mistral, Cerebras, OpenRouter) maintains its own full copy of the same ~550-line implementation:

Provider	Lines
DeepSeek	561
Groq	519
xAI	611
Mistral	574
Cerebras	561
OpenRouter	621
Avian	561

There is no shared createOpenAICompatibleProvider factory in this codebase. Creating one would be a worthwhile refactor, but it would be a cross-cutting change touching 7+ providers and is well outside the scope of this PR. Introducing it just for Avian would create an inconsistency where one provider uses an abstraction and six others don't.

The Avian implementation correctly follows the existing pattern, uses the shared utilities that do exist (createOpenAICompatibleStream, prepareToolsWithUsageControl, trackForcedToolUsage, calculateCost), and matches DeepSeek's proven implementation exactly. No refactoring applied.

[avianion]

Addressed the review feedback:

1. SVG icon: Replaced with a self-contained SVG using only inline path data — no external references, no embedded base64 images.

2. Provider implementation: Updated to match the current pattern on main — added sumToolCosts import, typed toolResults as Record<string, unknown>[], added toolCost field to cost tracking, and fixed the result.success && result.output guard. Merged upstream main to resolve conflicts.

3. Lint/formatting: Ran biome check and fixed formatting issues (callback argument placement in streaming functions). All files now pass biome check cleanly.

[avianion]

Addressed feedback: confirmed no dead code paths in the Avian provider — the implementation only handles text model completions with no unreachable branches. Reviewed all provider logic; the current code correctly returns text responses directly without any image-handling dead code.

[cursor]

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

There are 3 total unresolved issues (including 1 from previous review).

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

^{Reviewed by Cursor Bugbot for commit 7d94f16. Configure here.}

[avianion]

Addressed feedback: fixed abort signal propagation and trace enrichment in tool-calling loop