Chidori vs. Vercel AI SDK Gap Analysis

This note compares Chidori's current functionality against the Vercel AI SDK documentation reviewed on 2026-06-17.

The originally requested URL, https://ai-sdk.dev/v7/docs, was not reachable during the review. The live documentation available at https://ai-sdk.dev/docs identified v6 as the latest docs surface. The comparison below is therefore against the current public AI SDK docs that were available at review time.

Summary

Chidori and the Vercel AI SDK overlap on LLM calls, streaming, tool calling, agent loops, MCP tools, and multi-turn context. They are not direct substitutes.

AI SDK is broader as an application-facing TypeScript toolkit: many providers, framework UI hooks, structured output helpers, embeddings, reranking, media generation, provider middleware, and a large frontend/server integration surface.

Chidori is deeper as a durable agent runtime: host-call capture, deterministic replay, durable pause/resume, human input, multiplayer signals, branch stores, recorded HTTP/filesystem effects, runtime policy, and replayable checkpoints are runtime semantics rather than library conventions.

Current Overlap

Chidori Advantages

These are capabilities where Chidori has a stronger or more opinionated runtime contract than AI SDK.

Durable Execution By Default

Every side effect flows through the host boundary and is recorded. Replay uses the call log to return recorded host-call results with zero LLM calls. This covers prompts, tools, HTTP, input pauses, signals, memory, templates, and checkpoints.

AI SDK offers deterministic unit testing through mock providers, but it does not provide a built-in durable execution journal that can replay real production runs byte-for-byte.

Crash And Human Pause Recovery

Chidori sessions can pause on chidori.input(...), policy approval, or chidori.signal(...), persist state, and resume later through the session API. This is a runtime-level primitive for long-running human-gated work.

AI SDK supports application-managed persistence and resumable chat streams, but the application owns durable orchestration semantics.

Recorded External Effects

Chidori captures base effects such as fetch, node:http, VFS-backed workspace operations, crypto, timers, and host tools under policy and replay. Requests made inside dependencies inherit the same capture and approval model.

AI SDK focuses on model/tool abstractions. It does not make arbitrary HTTP or filesystem effects replayable by default.

Branching And Replayable Exploration

chidori.branch(...) forks a run into variant sub-runs from an anchored state, persists branch stores, and can resume or rerun branches independently. Replaying the parent returns the recorded branch outcomes without re-running the variants.

AI SDK supports workflow patterns and subagents, but does not expose an equivalent durable branch store or replayable fan-out primitive.

Runtime Policy And Capability Confinement

Chidori has explicit policy profiles, deterministic Date/Math.random policies, import policy, captured networking policy, and workspace mutation gates. These policies are part of the runtime.

AI SDK assumes the host JavaScript runtime and application enforce these constraints.

Chidori Gaps Against AI SDK

These are the main gaps to close if Chidori wants closer functional parity with the Vercel AI SDK framework.

1. Provider Breadth

Current Chidori support is intentionally narrow:

• Native Anthropic provider.
• Native OpenAI / OpenAI-compatible provider.
• LiteLLM-compatible routing as an escape hatch.
• MCP tools for tool integration, not model-provider expansion.

AI SDK has a much broader provider ecosystem, including official packages for OpenAI, Anthropic, Google, Google Vertex, Azure, Amazon Bedrock, Mistral, Groq, Cohere, DeepSeek, Cerebras, xAI, Together.ai, Fireworks, DeepInfra, Perplexity, ElevenLabs, Deepgram, AssemblyAI, and many community providers.

Gap:

• No first-class provider package interface.
• No published provider specification equivalent.
• No plugin-like provider ecosystem.
• Limited native model capability matrix.

2. Structured Output

Chidori exposes format: "json" and structured tool-call responses, but the repo explicitly documents no JSON-mode or structured-output plumbing beyond tool calls.

AI SDK supports structured output through Output.object(...), Output.array(...), schemas, validation, and streaming partial structured outputs.

Gap:

• No first-class schema-validated generateObject/Output.object equivalent.
• No Zod/Valibot/JSON Schema validation pipeline for model outputs.
• No partial structured output stream equivalent to AI SDK's streamed object generation.
• No typed structured-output authoring surface in the TypeScript SDK.

3. Embeddings And Reranking

Chidori does not currently expose embeddings or reranking as model primitives. The local roadmap states that embeddings and additional modalities are not yet implemented.

AI SDK documents embeddings and reranking as first-class core capabilities.

Gap:

• No embed / embedMany equivalent.
• No rerank API.
• No embedding model provider abstraction.
• Memory/vector storage exists at the host API level, but model-side embedding generation is not a first-class API.

4. Media Modalities

Chidori currently focuses on text, tool calls, captured effects, and durable agent execution.

AI SDK documents image generation, transcription, speech, and video generation surfaces where providers support them.

Gap:

• No first-class image generation API.
• No audio transcription API.
• No speech synthesis API.
• No video generation API.
• No media artifact streaming or provider capability abstraction.

5. Frontend UI Toolkit

Chidori exposes CLI streaming, session APIs, SSE events, and TypeScript/Python HTTP clients.

AI SDK UI provides framework-specific hooks such as useChat, useCompletion, and useObject for React, Vue, Svelte, Angular, and SolidJS usage.

Gap:

• No React/Vue/Svelte/Angular/Solid client hooks.
• No drop-in chat state manager.
• No UI message abstraction equivalent to AI SDK UI's UIMessage stream model.
• No generative UI helper layer.
• No client-side structured object hook equivalent to useObject.

6. Stream Protocol Compatibility

Chidori streams host calls and prompt lifecycle events:

• call
• prompt_start
• prompt_delta
• prompt_end
• paused
• done

AI SDK has broader stream helpers and UI-oriented stream protocols for chat, custom data, metadata, tool use, resumable streams, and UI consumption.

Gap:

• No compatibility layer for AI SDK UI stream protocols.
• No documented bridge from Chidori SSE events to AI SDK UI hooks.
• No standardized typed client event model beyond the current TypeScript SDK parser.

7. Provider Options And Middleware

Chidori supports basic prompt options such as model, temperature, max tokens, tools, stream labels, and cache posture.

AI SDK has a richer provider options and language-model middleware surface, including provider-specific settings, model wrapping, guardrails, RAG-style middleware, and lifecycle hooks.

Gap:

• No language model middleware API.
• Limited provider-specific option pass-through.
• No reusable model wrapper stack.
• No built-in guardrail/middleware pattern matching AI SDK's integration model.

8. Memory Ecosystem

Chidori has a durable chidori.memory.* host API and replay-aware memory operations.

AI SDK documents provider-defined memory tools and integrations with memory providers such as Letta, Mem0, Supermemory, and Hindsight.

Gap:

• No packaged memory-provider integrations.
• No provider-defined memory tool mapping.
• No semantic memory provider interface.
• No standard memory injection or retrieval middleware.

9. Testing Surface

Chidori's strongest testing mechanism is replaying real recorded runs. That is valuable for integration and regression testing.

AI SDK has dedicated mock language/embedding models, mock value generators, and stream simulation helpers designed for unit tests.

Gap:

• No TypeScript SDK test helper package comparable to ai/test.
• No mock provider API exposed to agent authors beyond environment-level static provider behavior.
• No stream simulation utilities for frontend/client tests.
• The TypeScript SDK now has a client test suite (sdk/typescript/test/, Node's built-in runner), but there is still no published mock-provider or stream-simulation test-helper package comparable to ai/test.

10. Package And Ecosystem Fit

Chidori agents run inside the embedded pure-Rust JavaScript engine. This gives Chidori control over determinism and captured effects, but it also means normal Node/npm compatibility is intentionally limited.

AI SDK runs in standard JavaScript application environments and can compose with the broader npm ecosystem.

Gap:

• No general npm package compatibility inside agents.
• Limited node: allowlist.
• No standard bundler/runtime integration story for arbitrary AI SDK ecosystem packages inside a Chidori agent.
• No adapter that lets Chidori agents call AI SDK model providers directly inside the durable runtime.

Positioning Implication

Chidori should not try to compete with AI SDK by matching every provider, frontend framework hook, media endpoint, and middleware feature immediately. The strongest current position is complementary:

• AI SDK is the application/model integration toolkit.
• Chidori is the durable execution substrate for expensive, long-running, replay-sensitive, human-gated, or branch-heavy agents.

The highest-leverage parity work would be the parts that preserve Chidori's runtime differentiation while reducing adoption friction:

1. Structured output with schema validation.
2. A provider abstraction or adapter story compatible with the AI SDK ecosystem.
3. Frontend stream adapters for AI SDK UI-style chat clients.
4. Embeddings/reranking as durable host calls.
5. A TypeScript test helper package that combines AI SDK-style mocks with Chidori replay fixtures.

References

Local references:

• README.md - Chidori positioning and durable host-call model.
• llm.txt - TypeScript agent and host API reference.
• docs/core-concepts.md - Host functions, context, prompt caching, and conversational agents.
• docs/replay.md - Replay model and SDK checkpoint flow.
• docs/context-management.md - Prompt caching, context composition, and compaction.
• docs/branching-execution.md - Branching design and branch stores.
• docs/signals.md - Multiplayer signal model.

External references reviewed:

• https://ai-sdk.dev/docs/introduction
• https://ai-sdk.dev/docs/ai-sdk-core/overview
• https://ai-sdk.dev/docs/ai-sdk-core/generating-structured-data
• https://ai-sdk.dev/docs/ai-sdk-core/tool-calling
• https://ai-sdk.dev/docs/ai-sdk-core/model-context-protocol
• https://ai-sdk.dev/docs/ai-sdk-core/testing
• https://ai-sdk.dev/docs/ai-sdk-core/telemetry
• https://ai-sdk.dev/docs/ai-sdk-ui/overview
• https://ai-sdk.dev/docs/agents/overview
• https://ai-sdk.dev/docs/agents/memory
• https://ai-sdk.dev/docs/agents/subagents