crig

crig is a Crystal library for building modular, builder-first LLM applications.

It ports the public behavior of Rig's rig-core crate into Crystal. The codebase includes agents, extractors, provider clients, embeddings, vector stores, streaming, tool servers, CLI/Discord integrations, and 100+ ported examples.

Current upstream parity source:

• Repository: https://github.com/0xPlaygrounds/rig.git

• Submodule path: vendor/rig

• Rust crate under port: vendor/rig/crates/rig-core

• Pinned upstream commit: f77a5819ec2a71e98583480a68a341f816a75c8a

• What is crig?

• Features

• Get Started

  • Simple agent
  • Structured extraction
  • Streaming

• Providers

• Examples

• Architecture

• Development

• Upstream Relationship

What is crig?

crig gives Crystal applications a unified, builder-first interface for:

• completion, chat, and streaming across 25+ providers
• agent workflows with tools, dynamic context, memory, and multi-turn prompting
• structured extraction
• embeddings, vector stores, and RAG pipelines
• tool servers (local and MCP)
• transcription, audio generation, and image generation

Every workflow starts from a provider client and branches through ergonomic builders:

• client.agent(model) — build an agent with preamble, tools, and context
• client.extractor(Type, model) — structured data extraction
• client.embeddings(model) — embeddings for search and RAG
• model.completion_request(prompt) — raw completion building
• model.stream(prompt) — streaming responses via channels

Features

• 25+ providers under a unified client/model abstraction
• Agent workflows: static context, dynamic (RAG) context, tools, tool servers, memory, multi-turn prompting with hooks
• Streaming via Crystal channels instead of async futures
• Structured extraction using the same submit-tool strategy as upstream Rig
• Embeddings builders, in-memory vector stores, vector search with distance metrics
• Tool definitions, embedding-backed tools, MCP tools, concurrent tool-server execution
• Transcription, audio generation, and image generation
• 100+ ported examples covering agents, extractors, streaming, RAG, and every provider

Get Started

Add crig to your shard.yml:

dependencies:
  crig:
    github: dsisnero/crig

shards install

Set your provider API key (here OpenAI):

export OPENAI_API_KEY="sk-..."

Simple agent

require "crig"

client = Crig::Providers::OpenAI::Client.from_env

agent = client
  .agent(Crig::Providers::OpenAI::GPT_5_2)
  .preamble("You are a helpful assistant.")
  .build

puts agent.prompt("What can you tell me about Crystal?").send

Structured extraction

require "crig"

struct Sentiment
  include JSON::Serializable
  getter sentiment : String
  getter confidence : Float64
end

client = Crig::Providers::OpenAI::Client.from_env

extractor = client
  .extractor(Sentiment, Crig::Providers::OpenAI::GPT_4O_MINI)
  .build

result = extractor.extract("I absolutely love this library!")
# => ExtractionResponse(@output=Sentiment(@sentiment="positive", @confidence=0.95))

Streaming

require "crig"

client = Crig::Providers::Anthropic::Client.from_env

model = client.completion_model(Crig::Providers::Anthropic::CLAUDE_SONNET_4_6)

model.stream("Write a haiku about programming.").each_item do |item|
  print item.text.try(&.text) if item.kind.text?
end
puts

Providers

25 providers under src/crig/providers/, all accessed through the same builder API:

Integration modules also exist for:

• CLI chatbots (Crig::Integrations::CliChatbot)
• Discord bots (Crig::Integrations::DiscordBot)
• MCP / RMCP tool servers (Crig::Tool::Server)
• Conversation memory (Crig::Memory)
• Telemetry / OpenTelemetry spans

Examples

100+ ported examples live under examples/. A few highlights:

Architecture

crig follows rig-core's decomposition with Crystal-native execution:

• src/crig/client/ — generic client infrastructure, provider builder traits, dyn-client factory
• src/crig/completion/ — message types, completion request/response, chat traits
• src/crig/agent/ — agent builder, prompt requests, streaming multi-turn loop, hooks
• src/crig/embeddings/ — embedding model traits, builders, distance metrics, vector stores
• src/crig/providers/ — 25 provider implementations with request/response conversion
• src/crig/tool/ — tool definitions, tool servers, MCP integration
• src/crig/pipeline/ — composable pipeline ops (map, then, chain, parallel, try)
• src/crig/loaders/ — file, PDF, and EPUB document loaders
• src/crig/evals.cr — LLM-as-judge evaluation metrics
• src/crig/memory.cr — conversation memory traits and in-memory backend

The builder API is the primary surface — every workflow starts from a client, chains builder calls, and terminates with .build, .send, .stream, or .extract. Crystal fibers and Channel replace Rust's async/futures for concurrency.

Development

make install        # install dependencies
make format         # crystal tool format --check
make lint           # ameba
make test           # crystal spec

Parity tracking lives under plans/inventory/. Bootstrap and validate:

./scripts/ensure_parity_plan.sh . vendor/rig/crates/rig-core rust auto 0
./scripts/check_port_inventory.sh . plans/inventory/rust_port_inventory.tsv vendor/rig/crates/rig-core rust
./scripts/check_source_parity.sh . plans/inventory/rust_source_parity.tsv vendor/rig/crates/rig-core rust
./scripts/check_test_parity.sh . plans/inventory/rust_test_parity.tsv vendor/rig/crates/rig-core rust

Current parity: 2,500 source items ported, 249 intentional divergences, 0 missing. Source parity tracks 2,155 API items; test parity tracks 507 upstream test equivalents.

Upstream Relationship

Upstream Rust behavior remains the source of truth for the tracked Rig surface, but the repository should be read as a Crystal library in its own right.

That means:

• docs should describe the APIs and capabilities that exist in src/ today
• examples should look like first-class Crystal examples, not temporary parity notes
• the only intentional architectural adaptation is concurrency mechanics: Rust uses async/futures, while crig uses Crystal fibers and channels

If you are contributing, follow AGENTS.md and update the parity inventories alongside implementation work.