🏗️ Production-Tested

Same runtime behind Copilot CLI, battle-tested at scale

🌐 Multi-Language

TypeScript, Python, Go, .NET, and Java SDKs available

🔧 Extensible

Custom tools, agents, MCP servers, hooks, and skills

⚡ Auto-managed or external CLI

💾 Session persistence

👥 Multi-client support (v3)

📡 40+ streaming event types

📈 Horizontal scaling

🔍 OpenTelemetry tracing

🔧 Custom Tools

Define functions Copilot can invoke

🤖 Custom Agents

Specialized AI personas with scoped tools

🔌 MCP Servers

External tool integration via standard protocol

📚 Skills

Reusable prompt modules from directories

🪝 Hooks

Intercept sessions for security and auditing

📡 Streaming

40+ real-time event types

🔑 BYOK

Bring your own model provider API keys

📊 Observability

Built-in OpenTelemetry instrumentation

🔍 Automated Code Review

Problem: Manual reviews miss security issues and take days

→ Deploy a security auditor agent that reviews every PR in seconds using OWASP-focused prompts and read-only tools

🚀 CI/CD Pipeline Intelligence

Problem: Pipelines fail silently; debugging takes hours

→ Run agents in your pipeline that analyze failures, suggest fixes, and auto-generate test coverage

🗄️ Database Query Assistants

Problem: Non-technical teams can't self-serve data queries

→ Connect MCP servers to your database — users ask in natural language, agent writes and explains SQL

📝 Documentation Generation

Problem: Docs are always outdated, nobody wants to write them

→ Agents analyze your codebase and produce up-to-date, comprehensive documentation on every release

💡 AI-Powered SaaS Products

Problem: Building AI features requires custom orchestration and model plumbing

→ Embed Copilot's agent runtime in your product with per-user OAuth, session isolation, and streaming — ship AI features in days, not months

🤝 Multi-Agent Workflows

Problem: Complex tasks need multiple specialized AI roles working together

→ Use Microsoft Agent Framework to orchestrate Copilot alongside Azure OpenAI and Anthropic in sequential or parallel pipelines

⚙️ DevOps Automation

Problem: Incident response and infrastructure changes require deep expertise

→ Agents that diagnose production issues, manage infrastructure, and automate deployments with guardrailed shell access

🏗️ Internal Developer Platforms

Problem: Internal tools lack intelligence and require constant maintenance

→ Embed Copilot in your developer portal, CLI tools, and dashboards — let third-party teams build on your AI-powered platform

📖 File System — Reading

• read_file — Read file contents
• view — View file or directory
• glob — Find files by name pattern
• grep — Search contents (ripgrep)
• list_dir — List directory contents

✏️ File System — Writing

• edit — Targeted edits to existing files
• create — Create new files
• write_file — Write / overwrite files

⚡ Shell, Web & Agent

• bash / shell — Execute commands
• web_fetch — Fetch content from URLs
• task — Delegate to sub-agents
• ask_user — Ask the user a question
• store_memory — Persist facts

1. Session Config

Whitelist or blacklist at creation

session = await client.create_session(
  available_tools=[
    "grep", "glob", "view"
  ],
  # OR
  excluded_tools=[
    "bash", "shell"
  ],
)

2. onPreToolUse Hook

Dynamic per-call decisions

ALLOWED = [
  "read_file", "glob",
  "grep", "view"
]

async def on_pre_tool_use(inp, inv):
  if inp["toolName"] not in ALLOWED:
    return {
      "permissionDecision": "deny"
    }
  return {
    "permissionDecision": "allow"
  }

3. Override Built-ins

Replace with your own (v0.1.30+)

@define_tool(
  description="Custom grep",
  overrides_built_in_tool=True,
)
async def grep(params):
  # Your custom implementation
  return custom_search(
    params.query
  )

🔧 Define

Name, description, and typed parameters via JSON Schema

⚡ Handle

Your async function runs when the agent calls the tool

🔗 Register

Pass tools array to createSession()

🔍 Research Agent

Tools: grep, glob, view

Analyze code, answer questions. Read-only.

✏️ Editor Agent

Tools: view, edit, bash

Make minimal, surgical code changes.

🛡️ Security Auditor

Tools: grep, glob, view

Focus on OWASP Top 10 vulnerabilities.

Local / Stdio

Runs as a subprocess, communicates via stdin/stdout

Local tools, file access, scripts

Remote / HTTP

Accessed via HTTP endpoint, supports SSE

Shared services, cloud tools

onSessionStart

Add context, audit logging

onUserPromptSubmitted

Modify or filter user input

onPreToolUse

Approve, deny, or modify tool calls

onPostToolUse

Transform results, redact secrets

onErrorOccurred

Custom error handling, retry logic

onSessionEnd

Cleanup, finalize audit trail

OAuth Flow

Code Example

const client = new CopilotClient({
  githubToken: userAccessToken,
  useLoggedInUser: false,
});

const session = await client.createSession({
  sessionId: `user-${userId}`,
  model: "gpt-4.1",
  onPermissionRequest: approveAll,
});

OpenAI

Direct API access

Azure AI Foundry

Enterprise deployments

Anthropic

Claude models

Ollama

Local models

Foundry Local

On-device AI

Other

vLLM, LiteLLM, etc.

Mode 1: stdio (Auto-Managed)

SDK spawns CLI as a child process

• Zero config — just new CopilotClient()
• CLI dies with your app
• Best for: local dev, desktop apps

Mode 2: TCP (External Server)

CLI runs as a headless server on a port

• copilot --headless --port 4321
• Independent lifecycle, multi-client
• Best for: APIs, backends, scaling

Supported Platforms

• Desktop apps (Electron, Tauri)
• Docker containers
• Distributable CLI tools

Multi-Platform Binaries

• copilot-darwin-arm64
• copilot-linux-x64
• copilot-win-x64.exe

💻 Local CLI

Hobbyist

SDK spawns CLI as child process. Simplest path for personal projects and development.

📦 Bundled CLI

Desktop App

Ship the CLI binary with your app. Standalone apps with zero external dependencies.

🖥️ Backend Services

Enterprise

CLI as headless server over TCP. Web backends, APIs, and microservices.

📈 Scaling

Platform

Horizontal scaling with session isolation. Serve thousands of users reliably.

Sequential Workflow

Concurrent Workflow

✨ New Features

• System prompt customization — surgically edit 10 prompt sections (identity, tone, safety, …) with replace/append/prepend/transform actions
• Blob attachments — send images and binary data inline without writing to disk
• skipPermission on tools — bypass confirmation for low-risk read-only tools
• OpenTelemetry — full distributed tracing across all 4 SDKs
• New RPC methods — skills, MCP, extensions, shell, and plugin management

⚠️ Breaking Changes

• Python API overhaul — keyword args replace TypedDicts
• autoRestart removed — deprecated across all SDKs
• Go Start() context — cancelling context no longer kills CLI process; use Stop()
• Go LogOptions.Ephemeral — changed from bool to *bool
• Python internal modules — copilot.jsonrpc → copilot._jsonrpc

Actions per section

Example

await client.createSession({
  systemMessage: {
    mode: "customize",
    sections: {
      identity: {
        action: (current) =>
          current.replace("Copilot",
            "Acme Assistant"),
      },
      tone: {
        action: "replace",
        content: "Be concise.",
      },
    },
  },
});