Claude Code

Anthropic’s official CLI for Claude — an agentic coding and knowledge-work tool that runs in the terminal, uses tools (file I/O, shell, web), and supports skills, subagents, and MCP integrations. One of the most active platforms in the agentic AI ecosystem as of 2026.

Key facts

• Type: Agentic AI CLI tool
• Maker: Anthropic
• Key properties: Skills system, subagents, slash commands, MCP protocol support, context/memory management, bash execution, web fetch
• Models: Claude 4.5/4.6 family (Opus, Sonnet, Haiku)
• Role in ecosystem: Frequently used as a foundation layer for higher-level orchestration tools like Paperclip and OpenClaw ^[src-001]

What it does

Claude Code runs Claude as an autonomous CLI agent with full tool access — reading/writing files, running shell commands, searching the web, calling MCP servers. It’s the building block most community tools assume when they talk about “the agent” — Paperclip ships with Claude Code as its default adapter, and Nate Herk’s examples consistently reach for it as the default runtime ^[src-001].

Two notable usage patterns from ^[src-001]:

1. Claude Code as adapter inside Paperclip. Each agent in a Paperclip company runs via the Claude Code CLI under the hood. Paperclip manages the heartbeats, task queue, and org chart; Claude Code does the actual reasoning and tool use.

2. Claude Code as a meta-layer ABOVE Paperclip. Nate Herk runs a separate Claude Code project whose only job is to help him configure, debug, and scale Paperclip itself. The Claude Code project has ingested the Paperclip repo, API docs, heartbeat protocol, and secret management patterns — and acts as Nate’s “partner in crime” for managing the orchestrator. This is the pattern: use Claude Code to orchestrate the orchestrator.

Claude Code as an agentic runtime

From ^[src-003], Claude Code fits the definition of an Agentic AI system: it runs a ReAct Loop (Reason + Act) (reason → call a tool → observe → iterate) with full tool access (file I/O, shell, web, MCP servers). It’s listed alongside Devin and GitHub Copilot Workspace as a leading agentic tool for software development.

Claude Code is also the reference runtime for the LLM Knowledge Bases (Karpathy pattern) pattern ^[src-002] — the ability to read entire project directories, maintain context across files, and edit markdown in place is what makes Karpathy-style wikis practical. Other LLMs with similar file access (GPT-4o, Gemini) would work in principle but the tooling ergonomics favour Claude Code in 2026.

Claude Code + MCP

Claude Code natively supports the Model Context Protocol (MCP) for tool integration — every external system Robin’s assistant talks to (WordPress, Gmail, Calendar) goes through an MCP server. This is why adding new capabilities is cheap: if an MCP server exists, it plugs straight in.

Related

• See also: Paperclip, OpenClaw, Obsidian, Agent Orchestration, Heartbeats (Agent Wake-ups), Agentic AI, ReAct Loop (Reason + Act), Model Context Protocol (MCP), LLM Knowledge Bases (Karpathy pattern)
• Compared against: _(no comparison pages yet)_

Updates from bulk ingest

From src-004 (cluster 1)

• Hit $2.5B annualised run rate by Feb 2026 — more than doubled since Jan 2026
• 29M daily installs within VS Code by Feb 2026
• Built on Bun + TypeScript + React; is both an MCP client and MCP server
• Source code partially leaked April 2026 via an npm source map file, exposing ~85 slash commands and internal feature flags
• Architecture: CLI parser -> query engine -> LLM API -> tool execution loop -> terminal renderer
• Internal feature flags reference voice mode, Chyros, daemon mode, and coordinator mode — some gated to anthropic user type
• Ultra Plan research preview runs on cloud Opus 4.6 with multi-agent exploration and ~30 minute cap
• Memory 2.0 with AutoDream launched March 2026 — background sub-agent consolidates memory files
• Auto Mode permission tier released March 2026 as safer alternative to –dangerously-skip-permissions
• Scheduled Tasks (desktop) and Loop skill (all surfaces) provide cron-style execution
• Channels feature (iMessage, Telegram, Discord) lets external chat apps push events into running sessions
• Remote Control lets mobile/web steer a running local session via QR code — Pro and Max plans only
• Native computer use released in research preview March 2026 — Mac OS desktop app only at launch
• Key commands: /clear, /compact, /context, /cost, /plan, /ultra-plan, /memory, /dream, /resume, /summary, /review, /security-review, /loop, /remote-control, /mcp, /plugins, /schedule
• Skills system with progressive context loading: level 1 frontmatter, level 2 skill.md, level 3 references
• Official Skill Creator runs evals, benchmarks, and trigger tuning on user skills
• Codex plugin from OpenAI brings GPT-5.4 into Claude Code via /codex commands (free with ChatGPT subscription)
• claude.md is operating context, not documentation — Anthropic explicitly treats it as an onboarding document for the agent
• Managed Agents (April 2026) let non-engineers deploy hosted agents with vault credentials, MCP OAuth, and 8 cents/hour billing
• Can be run with local Ollama models or OpenRouter free models by overriding ANTHROPIC env vars in settings.local.json
• Google Workspace CLI integration lets Claude Code control Drive/Gmail/Calendar/Docs/Sheets/Slides via a single CLI instead of MCP servers
• Creator: Boris Cherny

From src-005 (cluster 2)

• Connects to n8n via the community n8n MCP server (1100+ nodes, 2700+ templates) and the seven-skill n8n Skills repo for expression syntax, validation, and node configuration
• Can fix failing n8n workflows automatically via an error-workflow pattern that tunnels through ngrok to a local Claude Code instance
• Cannot fix credential, auth, or rate-limit errors in n8n workflows – escalates to the user via ClickUp/Slack notifications
• Runs inside VS Code via the Claude Code extension; ‘bypass permissions’ mode requires flipping ‘allow dangerously skip permissions’ in settings
• Uses claude.md files as project-level system prompts, plan mode for brainstorming, and slash commands like /clear to reset context
• Included in Anthropic plans from Pro ($17/mo) upward; Pro hits limits quickly on real workflow builds, $100/$200 tiers recommended for serious use
• Opus 4.5 is the recommended model for complex n8n builds; Sonnet 4.5 for writing tasks
• Can call MCP servers, read/write files, and execute Python/JavaScript – enabling the WAT framework pattern of markdown workflows + Python tools
• Builds full web apps end-to-end: optimises an n8n workflow, generates a Tailwind front end, pushes to GitHub, and auto-deploys via Vercel
• Runs security reviews on demand – can flag exposed credentials, hard-coded keys, and unsafe patterns on command
• Plugin ‘Ralph Wiggum for Claude Code’ adds continuous loops with max-iteration and done-signal guardrails for long-running tasks

Updates from src-009 (Claude Design cluster)

• Claude Code is now positioned as the deployment and deep-logic layer of a two-tool Anthropic pipeline: Claude Design handles prototyping; Claude Code handles GitHub, Vercel deployment, and functionality beyond HTML/CSS ^[src-009]
• Can consume Claude Design exports (ZIP or hand-off command) and the Design System skill.md manifest for brand-consistent work ^[src-009]
• Mobile optimisation of Claude Design outputs must be explicitly requested before deploying to production ^[src-009]

Updates from src-010 (Cloud agents & model releases)

• Cloud Routines launched 2026-04-14: configure a prompt once, runs on Anthropic infrastructure on a cron schedule, API POST, or GitHub event. No local machine required ^[src-010]
• Standalone desktop app launched alongside Opus 4.7 — calendar view for routines + local tasks, multi-session management, context window monitor, localhost preview, split views ^[src-010]
• /ultrareview command: spins up a dedicated review session reading all pending changes before committing ^[src-010]

Updates from src-011 (Claude Code power features)

• /btw command: opens a side-question overlay that never enters conversation history
• /re (rewind): jumps back to any previous message and drops everything after it — Anthropic’s #1 recommended context hygiene habit
• /loop command: reruns a prompt on a recurring schedule within a session (up to 3 days max)
• /hooks command: configures notification hooks in natural language
• claude --worktree [name] flag: creates an isolated git worktree for parallel sessions on the same project
• Native /voice command (rolling out) for voice-to-terminal input
• ultra think keyword allocates maximum thinking budget (~32K tokens) — for architecture and hard debugging
• Custom permissions: explicitly allow safe commands and deny destructive ones; deny-list takes priority over allow-list
• Context window startup overhead: 8K–62K tokens consumed before any user message (CLAUDE.md, MCP tool definitions, skills, context files)
• Auto-compaction fires at 95% context — retains only 20–30% of detail — community consensus is this is too late
• Boris Cherny (Claude Code creator) confirmed to start every session in plan mode ^[src-011]
• Output tokens cost more than input tokens; instructing conciseness doesn’t significantly help since most output tokens are invisible (file writes, tool calls) ^[src-011]
• Sub-agents can be assigned cheaper models (Haiku) while the main thread runs Opus — significant cost reduction ^[src-011]
• Agent teams extend sub-agents: agents share a task list and can communicate peer-to-peer ^[src-011]

Updates from src-012 (Video editing & content creation)

• Claude Code is now used as orchestration layer for full video production pipelines: coordinates video-use (editing), HyperFrames / Remotion (motion graphics), FFmpeg (rendering), HeyGen API (avatar generation), ElevenLabs API (voice clone) ^[src-012]
• Custom “make a video” skills and design philosophy markdown docs guide consistent outputs per video type ^[src-012]
• Token consumption for a video editing session: ~125K–260K tokens per session ^[src-012]

Updates from src-016 (10-hour Claude Code course)

• Nate positions Claude Code as a full build surface for automations, websites, apps, APIs, MCP integrations, RAG systems, Google Workspace workflows, executive assistants, skills, sub-agents, agent teams, browser automation, and client-delivery assets ^[src-016].
• The course stresses the Agentic Build / Deploy Boundary: Claude Code can self-heal while actively building and iterating, but scheduled/webhook/cloud deployments usually run deterministic code and tools without the live agent present ^[src-016].
• Practical operating sequence: start with project setup and CLAUDE.md, use plan mode for ambiguous work, keep secrets in .env, structure projects around reusable workflows/tools, then test and deploy only after the system behaves predictably ^[src-016].
• Monetisation advice connects Claude Code capability to delivery discipline: define scope, use client-owned credentials and infrastructure, run QA with real data, document handover, and separate maintenance from new-feature requests ^[src-016].

Updates from src-053 (lean AI tech stack)

• Nate places Claude Code as his number-one S-tier AI tool and describes it as the operating system he lives inside for as much work as possible ^[src-053].
• His preferred surface is VS Code, either through the terminal or the Claude Code extension, though he notes Claude Code can also run in the desktop app, CLI, Cursor, Windsurf, or other IDE-style environments ^[src-053].
• Claude Code remains the exploratory/planning center of Nate’s stack, while Codex (OpenAI) is used as a complementary weekly companion for execution, review, and troubleshooting ^[src-053].
• The video reinforces Agent Harness Portability: Claude Code projects should be organized so other harnesses such as Codex, OpenClaw, Hermes Agent, or future tools can operate in the same directory ^[src-053].

Updates from src-054 (Boris Cherny / Sequoia interview)

• Boris says Claude Code started inside Anthropic Labs in late 2024 as a Product Overhang bet: models were near the point of writing all code, while mainstream coding products were still type-ahead tools ^[src-054].
• Claude Code was not immediately a hit. Boris says it was barely usable for roughly six months, then growth inflected with Opus 4 and subsequent model releases ^[src-054].
• The codebase was deliberately built in TypeScript and React because those were on-distribution for the model at the time; for that codebase, Boris says the model now writes 100% of his code ^[src-054].
• Boris’s personal setup emphasizes massive parallelism: five to ten Claude mobile sessions, hundreds of agents, overnight thousands of agents, and loops for PR babysitting, CI health, flaky tests, rebasing, and Twitter-feedback clustering ^[src-054].
• Future product emphasis includes making loops first-class, making it easier to run many agents, improving batch/sub-agent/team patterns, and leaning on stronger models to reduce the need for heavy harness-side safety mechanisms ^[src-054].

Updates from src-055 (Karpathy / agentic engineering)

• Karpathy uses Claude Code and related tools as examples of the December 2025/early 2026 shift from agentic code chunks that needed correction to coherent workflows where outputs increasingly “just came out fine” ^[src-055].
• In the Agentic Engineering frame, Claude Code is not merely a faster editor; it is a harness for coordinating stochastic agents while the human owns spec, taste, fundamentals, security, and verification ^[src-055].
• Karpathy suggests AI-native hiring should evaluate candidates by giving them larger realistic projects and letting agents attempt to break the result, rather than relying on old puzzle-style coding interviews ^[src-055].

Updates from src-071 (Anthropic Economic Index)

• Anthropic’s March 2026 Economic Index says Claude Code had grown to represent a large share of sampled first-party API traffic by February 2026 ^[src-071].
• Claude Code’s agentic architecture splits coding work into many smaller API calls, making coding growth show up as a broader set of O*NET task categories in API traffic ^[src-071].
• This positions Claude Code as part of API Workflow Migration: coding work is moving from chat-style Claude.ai use into programmatic, more automatable workflows ^[src-071].

Updates from src-072 (Scientific computing)

• Anthropic’s scientific-computing article uses Claude Code as a long-running research agent on an HPC/SLURM setup, launched inside tmux so the user can detach and reattach while the agent works ^[src-072].
• The recommended project pattern is CLAUDE.md for goals and rules, a progress file for session memory, a reference implementation or test suite as oracle, and Git commits after meaningful units of work ^[src-072].
• Claude Code is shown not only as a coding assistant but as a research-lab worker that can make sequential progress on specialized numerical software with occasional human steering ^[src-072].

Updates from src-074 (Hermes Agent course)

• Nate uses Claude Code as the setup and management layer for Hermes Agent and OpenClaw-style assistants: inspect repos, edit environment/config files, maintain shared project instructions, and troubleshoot the agent server ^[src-074].
• The course reinforces the role split: Claude Code remains Nate’s higher-control surface for deep coding, context inspection, and complex project work, while Hermes is the phone-first assistant surface for quick asks, scheduled jobs, and proactive operations ^[src-074].
• Claude Code also becomes the “agent maintainer” in the self-improvement loop: when Hermes repeatedly needs the same behavior, Claude Code can help turn it into a skill, patch memory, or adjust files in the shared repo ^[src-074].

Source references

• ^[src-001] Nate Herk — “Claude Code + Paperclip Just Destroyed OpenClaw” (2026-03-28)
• ^[src-002] Robin Cartier — “Karpathy’s LLM Knowledge Base: A Practitioner’s Verdict” (2026-04-08)
• ^[src-003] Robin Cartier — “What is Agentic AI? A Complete Guide” (2026-03-10)
• ^[src-004] Nate Herk cluster (see summaries/src-004-*.md)
• ^[src-005] Nate Herk cluster (see summaries/src-005-*.md)
• ^[src-009] Nate Herk — Claude Design cluster (2026-04-17 to 2026-04-21)
• ^[src-010] Nate Herk — Cloud agents & model releases cluster (2026-04-14 to 2026-04-17)
• ^[src-011] Nate Herk — Claude Code power features cluster (2026-04-20 to 2026-04-27)
• ^[src-012] Nate Herk — Video editing & content creation cluster (2026-04-15 to 2026-04-23)
• ^[src-016] Nate Herk — “Build & Sell with Claude Code (10+ Hour Course)” (2026-03-12)
• ^[src-053] Nate Herk — “Overwhelmed By AI? Just Copy My Tech Stack” (2026-05-08)
• ^[src-054] Sequoia Capital — “Anthropic’s Boris Cherny: Why Coding Is Solved, and What Comes Next” (2026-05-04)
• ^[src-055] Sequoia Capital — “Andrej Karpathy: From Vibe Coding to Agentic Engineering” (2026-04-29)
• ^[src-071] Anthropic – “Anthropic Economic Index report: Learning curves” (2026-03-24)
• ^[src-072] Siddharth Mishra-Sharma – “Long-running Claude for scientific computing” (2026-03-23)
• ^[src-074] Nate Herk — “Hermes Agent: Zero to Personal AI Assistant (1 Hour Course)” (2026-05-10)