Embacle — LLM Runners

Standalone Rust library that wraps AI CLI tools and SDKs as pluggable LLM providers.

Instead of integrating with LLM APIs directly (which require API keys, SDKs, and managing auth), Embacle delegates to CLI tools that users already have installed and authenticated — getting model upgrades, auth management, and protocol handling for free. For GitHub Copilot, an optional SDK mode maintains a persistent JSON-RPC connection for native tool calling.

Supported Runners

CLI Runners (subprocess-based)

SDK Runners (persistent connection)

Quick Start

Add to your Cargo.toml:

[dependencies]
embacle = "0.3"

Use a CLI runner:

use std::path::PathBuf;
use embacle::{ClaudeCodeRunner, RunnerConfig};
use embacle::types::{ChatMessage, ChatRequest, LlmProvider};

#[tokio::main]
async fn main() -> Result<(), embacle::types::RunnerError> {
    let config = RunnerConfig::new(PathBuf::from("claude"));
    let runner = ClaudeCodeRunner::new(config);

    let request = ChatRequest::new(vec![
        ChatMessage::user("What is the capital of France?"),
    ]);

    let response = runner.complete(&request).await?;
    println!("{}", response.content);
    Ok(())
}

Copilot SDK (feature flag)

Enable the copilot-sdk feature for persistent JSON-RPC instead of per-request subprocesses:

[dependencies]
embacle = { version = "0.3", features = ["copilot-sdk"] }

use embacle::{CopilotSdkRunner, CopilotSdkConfig};
use embacle::types::{ChatMessage, ChatRequest, LlmProvider};

#[tokio::main]
async fn main() -> Result<(), embacle::types::RunnerError> {
    // Reads COPILOT_SDK_MODEL, COPILOT_GITHUB_TOKEN, etc. from env
    let runner = CopilotSdkRunner::from_env();

    let request = ChatRequest::new(vec![
        ChatMessage::user("Explain Rust ownership"),
    ]);

    let response = runner.complete(&request).await?;
    println!("{}", response.content);
    Ok(())
}

The SDK runner starts copilot --headless once and reuses the connection across requests. Configuration via environment variables:

MCP Server (embacle-mcp)

A standalone MCP server binary that exposes embacle runners via the Model Context Protocol. Connect any MCP-compatible client (Claude Desktop, editors, custom agents) to use all embacle providers.

Usage

# Stdio transport (default — for editor/client integration)
embacle-mcp --provider copilot

# HTTP transport (for network-accessible deployments)
embacle-mcp --transport http --host 0.0.0.0 --port 3000 --provider claude_code

MCP Tools

Client Configuration

Add to your MCP client config (e.g. Claude Desktop claude_desktop_config.json):

{
  "mcpServers": {
    "embacle": {
      "command": "embacle-mcp",
      "args": ["--provider", "copilot"]
    }
  }
}

REST API Server (embacle-server)

An OpenAI-compatible HTTP server that proxies requests to embacle runners. Any client that speaks the OpenAI chat completions API can use it without modification.

Usage

# Start with default provider (copilot) on localhost:3000
embacle-server

# Specify provider and port
embacle-server --provider claude_code --port 8080 --host 0.0.0.0

Endpoints

Model Routing

The model field determines which provider handles the request. Use a provider:model prefix to target a specific runner, or pass a bare model name to use the server's default provider.

# Explicit provider
curl http://localhost:3000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "claude:opus", "messages": [{"role": "user", "content": "hello"}]}'

# Default provider
curl http://localhost:3000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-4o", "messages": [{"role": "user", "content": "hello"}]}'

Multiplex

Pass an array of models to fan out the same prompt to multiple providers concurrently. Each provider runs in its own task; failures in one don't affect others.

curl http://localhost:3000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": ["copilot:gpt-4o", "claude:opus"], "messages": [{"role": "user", "content": "hello"}]}'

The response uses object: "chat.completion.multiplex" with per-provider results and timing.

Streaming is not supported for multiplex requests.

SSE Streaming

Set "stream": true for Server-Sent Events output in OpenAI streaming format (data: {json}\n\n with data: [DONE] terminator).

Authentication

Optional. Set EMBACLE_API_KEY to require bearer token auth on all endpoints. When unset, all requests are allowed through (localhost development mode). The env var is read per-request, so key rotation doesn't require a restart.

EMBACLE_API_KEY=my-secret embacle-server
curl http://localhost:3000/v1/models -H "Authorization: Bearer my-secret"

Architecture

Your Application
    └── embacle (this library)
            │
            ├── CLI Runners (subprocess per request)
            │   ├── ClaudeCodeRunner    → spawns `claude -p "prompt" --output-format json`
            │   ├── CopilotRunner       → spawns `copilot -p "prompt"`
            │   ├── CursorAgentRunner   → spawns `cursor-agent -p "prompt" --output-format json`
            │   ├── OpenCodeRunner      → spawns `opencode run "prompt" --format json`
            │   ├── GeminiCliRunner     → spawns `gemini -p "prompt" -o json -y`
            │   ├── CodexCliRunner      → spawns `codex exec "prompt" --json --full-auto`
            │   ├── GooseCliRunner      → spawns `goose run --quiet --no-session`
            │   ├── ClineCliRunner      → spawns `cline task --json --act --yolo`
            │   └── ContinueCliRunner   → spawns `cn -p --format json`
            │
            ├── SDK Runners (persistent connection, behind feature flag)
            │   └── CopilotSdkRunner    → JSON-RPC to `copilot --headless`
            │
            ├── Provider Decorators (composable wrappers)
            │   ├── FallbackProvider    → ordered chain, first success wins
            │   ├── MetricsProvider     → latency, token, and error tracking
            │   └── QualityGateProvider → response validation with retry
            │
            ├── Agent Loop
            │   └── AgentExecutor       → multi-turn tool calling with configurable max turns
            │
            ├── Structured Output
            │   └── request_structured_output()  → schema-validated JSON extraction with retry
            │
            ├── MCP Tool Bridge
            │   └── McpToolBridge       → MCP tool definitions ↔ text-based tool loop
            │
            ├── MCP Server (separate binary crate)
            │   └── embacle-mcp         → JSON-RPC 2.0 over stdio or HTTP/SSE
            │
            ├── REST API Server (separate binary crate)
            │   └── embacle-server      → OpenAI-compatible HTTP, SSE streaming, multiplex
            │
            └── Tool Simulation (text-based tool calling for CLI runners)
                └── execute_with_text_tools()  → catalog injection, XML parsing, tool loop

All runners implement the same LlmProvider trait:

• complete() — single-shot completion
• complete_stream() — streaming completion
• health_check() — verify the runner is available and authenticated

The CopilotSdkRunner additionally provides:

• execute_with_tools() — native tool calling via the SDK's session and tool handler infrastructure

Text-Based Tool Calling (CLI runners)

CLI runners don't have native tool calling, so Embacle provides a text-based simulation layer. It injects a tool catalog into the system prompt and parses <tool_call> XML blocks from the LLM response, looping until the model stops calling tools.

use embacle::tool_simulation::{
    FunctionDeclaration, FunctionCall, FunctionResponse,
    execute_with_text_tools,
};
use embacle::types::{ChatMessage, LlmProvider};
use std::sync::Arc;
use serde_json::json;

let declarations = vec![
    FunctionDeclaration {
        name: "get_weather".into(),
        description: "Get current weather for a city".into(),
        parameters: Some(json!({"type": "object", "properties": {"city": {"type": "string"}}})),
    },
];

let handler = Arc::new(|name: &str, args: &serde_json::Value| -> FunctionResponse {
    FunctionResponse {
        name: name.to_owned(),
        response: json!({"temperature": 22, "conditions": "sunny"}),
    }
});

let mut messages = vec![ChatMessage::user("What's the weather in Paris?")];
let result = execute_with_text_tools(
    &runner,          // any LlmProvider (CopilotRunner, ClaudeCodeRunner, etc.)
    &mut messages,
    &declarations,
    handler,
    5,                // max iterations
).await?;

println!("{}", result.content);           // final response with tools stripped
println!("Tool calls: {}", result.tool_calls_count);

The pure functions are also available individually for custom loop implementations:

Agent Loop

AgentExecutor runs a multi-turn tool-calling loop: inject a tool catalog, send the prompt, parse <tool_call> blocks from the response, execute tools, feed results back, repeat until the model stops calling tools or max_turns is reached.

use embacle::agent::{AgentExecutor, AgentConfig};
use embacle::tool_simulation::FunctionDeclaration;
use embacle::types::{ChatMessage, ChatRequest};
use std::sync::Arc;
use serde_json::json;

let declarations = vec![FunctionDeclaration {
    name: "lookup".into(),
    description: "Look up a value".into(),
    parameters: Some(json!({"type": "object", "properties": {"key": {"type": "string"}}})),
}];

let handler = Arc::new(|name: &str, _args: &serde_json::Value| {
    embacle::tool_simulation::FunctionResponse {
        name: name.to_owned(),
        response: json!({"value": 42}),
    }
});

let agent = AgentExecutor::new(Arc::new(runner), declarations, handler)
    .with_max_turns(5);

let request = ChatRequest::new(vec![ChatMessage::user("Look up answer")]);
let result = agent.execute(&request).await?;
println!("Turns: {}, Tool calls: {}", result.total_turns, result.tool_calls);

Fallback Chains

FallbackProvider wraps multiple providers and tries them in order. The first successful response wins; if all fail, the last error is returned.

use embacle::fallback::FallbackProvider;

let provider = FallbackProvider::new(vec![
    Box::new(primary_runner),
    Box::new(backup_runner),
])?;

// Uses primary_runner; falls back to backup_runner on error
let response = provider.complete(&request).await?;

Health checks pass if any provider is healthy. Capabilities are the union of all inner providers.

Metrics

MetricsProvider wraps any provider to track latency, token usage, call counts, and errors.

use embacle::metrics::MetricsProvider;

let provider = MetricsProvider::new(Box::new(runner));
let response = provider.complete(&request).await?;

let report = provider.report();
println!("Calls: {}, Avg latency: {}ms", report.call_count, report.avg_latency_ms);

Quality Gate

QualityGateProvider validates responses against a policy (minimum length, refusal detection) and retries with feedback if validation fails.

use embacle::quality_gate::{QualityGateProvider, QualityPolicy};

let policy = QualityPolicy {
    max_retries: 2,
    min_content_length: 10,
    ..QualityPolicy::default()
};
let provider = QualityGateProvider::with_policy(Box::new(runner), policy);
let response = provider.complete(&request).await?;

Structured Output

Forces any provider to return schema-valid JSON by injecting schema instructions and validating the response, retrying with error feedback on schema violations.

use embacle::structured_output::{request_structured_output, StructuredOutputRequest};
use serde_json::json;

let schema = json!({
    "type": "object",
    "properties": {
        "city": {"type": "string"},
        "temperature": {"type": "number"}
    },
    "required": ["city", "temperature"]
});

let result = request_structured_output(
    &runner,
    &StructuredOutputRequest { request, schema, max_retries: 2 },
).await?;

let data: serde_json::Value = serde_json::from_str(&result.content)?;

MCP Tool Bridge

Bridges MCP tool definitions to embacle's text-based tool loop, so CLI runners can use tools from any MCP-compatible tool server.

use embacle::mcp_tool_bridge::{McpToolBridge, McpToolDefinition};

let mcp_tools = vec![McpToolDefinition {
    name: "search".into(),
    description: Some("Search the web".into()),
    input_schema: serde_json::json!({"type": "object", "properties": {"query": {"type": "string"}}}),
}];

let declarations = McpToolBridge::to_declarations(&mcp_tools);

Features

• Zero API keys — uses CLI tools' own auth (OAuth, API keys managed by the tool)
• Auto-discovery — finds installed CLI binaries via which
• Auth readiness — non-blocking checks, graceful degradation
• Capability detection — probes CLI version and supported features
• Container isolation — optional container-based execution for production
• Subprocess safety — timeout, output limits, environment sandboxing
• Agent loop — multi-turn tool calling with configurable max turns and turn callbacks
• Fallback chains — ordered provider failover with automatic retry
• Metrics — latency, token, and error tracking as a provider decorator
• Quality gate — response validation with retry on refusal or insufficient content
• Structured output — schema-validated JSON extraction from any provider
• MCP tool bridge — connect MCP tool servers to CLI runners via text-based tool loop
• Feature flags — SDK integrations are opt-in to keep the default dependency footprint minimal

Modules

License

Licensed under the Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0).