Skip to main content

Skills

AgentFlow skills are bundled assistant instructions for coding agents such as Codex, Claude, and GitHub Copilot. They are copied from:

agentflow-api/agentflow_cli/cli/templates/skills

Use these skills when you want an assistant to understand AgentFlow package boundaries, graph patterns, CLI behavior, API routes, TypeScript client conventions, testing, and production guidance while editing an AgentFlow project.

agentflow skills --agent codex
agentflow skills --agent claude
agentflow skills --agent github

What gets installed

The same base skill bundle is copied into the assistant-specific location.

AssistantInstalled files
Codex.agents/skills/agentflow/
Claude.claude/skills/agentflow/
GitHub Copilot.github/instructions/agentflow.instructions.md and .github/skills/agentflow/

For Codex and Claude, the installed folder is copied from:

agentflow-api/agentflow_cli/cli/templates/skills/agent-skills

For GitHub Copilot, AgentFlow also copies:

agentflow-api/agentflow_cli/cli/templates/skills/copilot/agentflow.instructions.md

That file points Copilot at the installed skill bundle under .github/skills/agentflow/.

What the skill contains

The base bundle contains:

agent-skills/
+-- SKILL.md
+-- references/
+-- architecture.md
+-- agents-and-tools.md
+-- state-graph.md
+-- state-and-messages.md
+-- checkpointing-and-threads.md
+-- dependency-injection.md
+-- media-and-files.md
+-- memory-and-store.md
+-- streaming.md
+-- production-runtime.md
+-- api-client.md
+-- remote-tools.md
+-- callbacks-and-command.md
+-- prebuilt-agents-and-tools.md
+-- testing-and-evaluation.md
+-- publishers-and-runtime-protocols.md
+-- context-id-background.md
+-- providers-and-adapters.md
+-- security-and-validators.md
+-- cli-commands.md
+-- api-configuration.md
+-- auth-and-authorization.md
+-- api-settings-and-middleware.md
+-- rest-api-and-errors.md
+-- id-and-thread-name-generators.md
+-- client-auth-and-errors.md
+-- client-messages-invoke-stream.md
+-- client-threads-memory-files.md

SKILL.md is the entry point. It tells the assistant when to use the AgentFlow skill, which packages exist, where the public docs live, and which reference file to read before changing a subsystem.

The reference files cover:

AreaWhat the assistant learns
ArchitecturePackage layout across agentflow, agentflow-api, agentflow-client, docs, and playground
Agents and graphsAgent, ToolNode, StateGraph, prebuilt agents, state, messages, tools, and handoffs
Runtime behaviorCheckpointing, dependency injection, memory, media, streaming, publishers, and protocols
API and CLIagentflow init, api, play, build, skills, agentflow.json, auth, settings, middleware, routes, and errors
TypeScript clientAuth, invoke, stream, messages, threads, memory, files, and client-side tool execution
Quality and safetyTesting, evaluation, provider adapters, validators, and prompt-injection safeguards

Install for one assistant

Run from the project root:

agentflow skills --agent codex

Supported values are codex, claude, github, or menu numbers 1, 2, 3.

If you omit --agent in an interactive terminal, AgentFlow prompts you to choose:

Which agent?
- 1. Codex
- 2. Claude
- 3. GitHub

In non-interactive environments, pass --agent or --all.

Install for every assistant

agentflow skills --all

If an installation already exists, --all skips that assistant unless you also pass --force.

Install into another project

agentflow skills --agent claude --path ./my-agent

The command refuses to install directly into the filesystem root or your home directory. Point --path at a project folder.

Update an existing install

agentflow skills --agent github --force

Use --force to replace an existing installed AgentFlow skill after updating the CLI.

List supported assistants

agentflow skills --list

Options

OptionDefaultDescription
--agent, -ainteractive promptTarget assistant: codex, claude, github, or menu number 1, 2, 3
--path, -p.Project directory where skills should be installed
--force, -ffalseOverwrite an existing install
--allfalseInstall skills for every supported assistant
--list, -lfalseList supported assistants and exit
--verbose, -vfalseEnable verbose logging
--quiet, -qfalseSuppress output except errors