ag2ai/ag2-skills
22 skills708 installs176 starsGitHub
Install
npx skills add https://github.com/ag2ai/ag2-skillsSkills in this repo
1Ag2 Ag Uiag2-ag-ui is a Claude Code skill that exposes an AG2 beta Agent over the AG-UI protocol so a web frontend can stream responses, render tool calls, sync shared state and track sub-task steps. It wraps the agent with AGUIStream and mounts it in FastAPI. A developer uses it when they want a web UI, such as CopilotKit or a custom React app, in front of an AG2 agent rather than a CLI or script.36installs2Ag2 Hitlag2-hitl is a Claude Code skill that adds human-in-the-loop control to an AG2 beta Agent. It pauses the agent mid-run to collect typed human input via context.input() with a hitl_hook, and gates specific tool calls behind approval_required() middleware. A developer uses it when the agent should ask for confirmation, request missing information, or require human approval before running sensitive, irreversible or expensive tool calls such as sending emails, deleting records or payments.36installs3Ag2 Add Custom Toolag2-add-custom-tool is a Claude Code skill that adds a custom Python tool to an AG2 beta Agent using the @tool decorator. It covers sync and async tools, Pydantic parameter typing, typed Input and ToolResult returns for text, data, images and binary, final=True early exit, and dependency injection. A developer uses it to give an AG2 agent a new capability backed by Python code such as API calls, database queries or computations.35installs4Ag2 Knowledge And MemoryThis skill teaches an AG2 beta agent to persist state across runs, control what the LLM sees each turn, and cap history to fit a context window. It documents KnowledgeStore backends, KnowledgeConfig wiring, aggregation strategies, and assembly policies like SlidingWindowPolicy and TokenBudgetPolicy. A developer uses it when they want an agent to remember between conversations or manage long event histories.35installs5Ag2 QuickstartThis skill builds a minimal AG2 beta Agent end to end: pick a model provider, set a prompt, call agent.ask(), then chain follow-up turns with reply.ask() to preserve context. A developer uses it when starting a new AG2 beta project or when unsure which provider config to use. It covers OpenAIConfig, AnthropicConfig, GeminiConfig, OllamaConfig, env-var fallback for API keys, and OpenAI-compatible endpoints.35installs6Ag2 Subagent DelegationThis skill covers single-agent recursion and parallel fan-out inside one AG2 beta Agent. A developer opts in with tasks=TaskConfig(...) to get auto-injected run_subtask and run_subtasks(parallel=True) tools for self-delegation, or uses Agent.as_tool() to call one named specialist agent from inside another. It documents context flow, recursion safety, and persistent_stream for sub-task history, and points to ag2-network-quickstart for full multi-agent collaboration.35installs7Ag2 Multimodal InputThis skill covers sending images, audio, video, or documents into an AG2 beta Agent alongside text using ImageInput, AudioInput, VideoInput, and DocumentInput. A developer uses it when they want an agent to describe a photo, transcribe audio, summarise a PDF, or analyse a video. It documents the per-provider support matrix, the four ways to source data, and provider-specific features like Gemini YouTube URLs and Anthropic attachment caching.34installs8Ag2 OverviewThis skill is a capability map of AG2 beta (autogen.beta), an async, protocol-driven agent framework, and an index of which sibling skill to reach for. A developer loads it first when a request mentions building with AG2 beta but the specific feature is not yet clear. It covers the three setup prerequisites (provider extra, API key, install check) and a table mapping user intents to skills like ag2-quickstart, ag2-add-custom-tool, and the network skills.34installs9Ag2 Use Builtin ToolsThis skill shows how to wire AG2 beta's shipped tools into an Agent rather than writing custom Python. It covers provider-native server-side tools such as web search, web fetch, code execution, MCP server integration, and image generation, plus locally executed toolkits like FilesystemToolkit, DuckDuckGo, Exa, and Tavily search. A developer uses it to add ready-made capabilities to an agent and to check per-tool provider support before wiring.34installs10Ag2 Observers And AlertsThis skill monitors an AG2 beta agent's event stream to log events, detect repeated tool calls, track token spend, and build trigger-driven observers. A developer uses it for observability, runtime safety guards, alerts, or batch/time-based reactive logic. It covers stateless @observer functions, stateful BaseObserver classes, built-ins like TokenMonitor and LoopDetector, Watch primitives, ObserverAlert severities, and halting on FATAL conditions.33installs11Ag2 Shell ToolThis skill gives an AG2 beta Agent the ability to run shell commands. It covers SandboxShellTool, a client-side subprocess via LocalEnvironment that works with any provider, and the provider-native ShellTool for OpenAI Responses server-side execution. A developer uses it when the agent needs to execute commands, build or test code, or operate on a workspace, always paired with sandboxing filters.33installs12Ag2 Structured OutputThis skill shows how to get a typed, validated Python value back from an AG2 beta Agent instead of free text. A developer passes response_schema= (a Pydantic model, dataclass, primitive, union, ResponseSchema, or @response_schema validator) and reads the parsed result via await reply.content(). It is used for classification, extraction, scoring, and normalisation where downstream code parses the reply, and it covers validation retries and PromptedSchema for providers without native structured output.33installs13Ag2 TelemetryThis skill adds OpenTelemetry tracing to an AG2 beta Agent using TelemetryMiddleware. It emits spans for the full turn, each LLM call, each tool execution, and each human-input request following OpenTelemetry GenAI semantic conventions. A developer uses it to get per-turn latency breakdowns, attribute token usage, and push traces into an existing observability backend such as Jaeger, Grafana Tempo, Datadog, Honeycomb, or Langfuse.33installs14Ag2 TestingThis skill shows how to write pytest tests for AG2 beta agents and tools without making real LLM API calls. A developer passes TestConfig(...) as the agent's config to mock LLM responses and injects ToolCallEvent objects to simulate tool execution, then asserts success and error paths. It documents overriding Depends/Inject dependencies, capturing stream events, and the per-ask() cursor behaviour of the mocked response list.33installs15Ag2 Network GovernanceThis skill governs an AG2 multi-agent network from the hub side, covering identity (Passport, Resume), per-agent Rule blocks for access and limits, channel-level expectations, and an append-only audit log. A developer uses it when a network needs rate limits, access policy, SLAs, compliance trails, live metrics, or capability-driven peer ranking. It also documents live HubListener observability and task observation that builds each agent's track record.30installs16Ag2 Network QuickstartThis skill builds a multi-agent AG2 beta network, the standard multi-agent pattern in autogen.beta, activated whenever two or more Agents need to interact. It covers Hub.open, LocalLink, HubClient.register, Passport, Resume, the channel lifecycle, and the two 2-party channel adapters (consulting for strict question-and-reply and conversation for free-form). It is the entry point that routes to the discussion, workflow, governance, and tools-and-views skills.30installs17Ag2 Network Tools And ViewsThis skill shapes what an AG2 network agent perceives and which actions its LLM can take. It covers the five auto-injected NetworkPlugin tools (delegate, peers, channels, tasks, context) plus the adapter-owned say, replacing the default handler for gateways and headless workers, the ViewPolicy protocol, peer discovery via skill markdown, and the Envelope wire format with its EV_* event taxonomy.30installs18Ag2 Network WorkflowThis skill builds an AG2 network workflow channel, an orchestrated N-party adapter driven by a declarative TransitionGraph. A developer uses it for conditional handoffs, multi-step pipelines, context-aware routing, feedback loops, or migrating from the classic GroupChat and Agent.handoffs pattern. It covers TransitionGraph factories, built-in targets and conditions, the typed Handoff return for dynamic routing, and side-by-side migration from classic GroupChat.30installs19Ag2 Evaluationag2-evaluation is a Claude Code skill that evaluates, tests and tracks an AG2 beta agent offline. It builds a Suite of tasks, runs the agent with run_agent, and grades answers with prebuilt scorers or a custom scorer including an LLM judge. A developer uses it to score correctness, tool use, cost or quality, build a CI regression gate, or diff runs to catch regressions. For head-to-head or leaderboard comparison it points to ag2-eval-comparison.28installs20Ag2 Eval Comparisonag2-eval-comparison is a Claude Code skill that compares AG2 beta agents, models or prompts to decide which is better. It ranks named variants on a leaderboard with run_variants, runs head-to-head LLM judging with a dual-order position swap and Wilson confidence intervals, and collects blinded human preference votes. A developer uses it to A/B test prompts or models, run a leaderboard, or pick a winner. For grading a single agent it points to ag2-evaluation.27installs21Ag2 MiddlewareThis skill covers intercepting the AG2 beta agent loop with BaseMiddleware, which exposes four hooks: on_turn, on_llm_call, on_tool_execution, and on_human_input. It is for cross-cutting behaviour like retry, logging, history trimming, request mutation, tool auditing, and rate limiting. It documents built-ins such as LoggingMiddleware, RetryMiddleware, HistoryLimiter, and TokenLimiter, plus how to write your own.27installs22Ag2 Network DiscussionThis skill opens an AG2 network discussion channel, an N-party round-robin where each participant speaks in fixed order, cycling until an explicit close or TTL. A developer uses it for a brainstorm with a fixed cast, a panel discussion, or round-robin reviewers. It documents the ordering knob, the expected_next_speaker rotation, and the can_send probe pattern that lets agents skip wasted LLM calls when it is not their turn.27installs