Claude Code
Type: full-code · Vendor: Anthropic · Language: TypeScript, Shell · License: proprietary (closed-source CLI; no LICENSE in repo) · Status: active · Status in practice: mature · First released: 2025-02-22
Anthropic's first-party agentic coding tool — a single CLI/IDE/desktop/web surface that turns Claude into a tool-using engineer with persistent project memory, structured subagent delegation, hooks, skills, and scheduled cloud routines.
Description. Claude Code lives in the terminal, IDE, desktop app, and browser, sharing the same engine across surfaces. It reads the repository, edits files, runs commands, drives git, and exposes a documented agentic loop with explicit primitives for memory (CLAUDE.md, auto memory), skills, subagents, hooks, MCP, plan mode, todo lists, and scheduled routines. Anthropic positions it as both a developer tool and the reference implementation of an autonomous coding agent (the Claude Agent SDK is built on the same engine).
Agent loop shape. ReAct-style think→act loop where Claude reasons, calls a tool (Read/Edit/Bash/etc. or MCP tool), observes the result, and continues until done. Plan mode runs a read-only exploration first; once accepted, execution runs the tool-use loop with checkpointing, todo tracking, and optional subagent forking. Hooks fire at SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, SessionEnd. Sessions persist as JSONL under ~/.claude/projects/ and can be resumed, branched, or forked.
Primary use cases
- agentic coding across terminal/IDE/desktop/web with shared engine
- long-running autonomous coding sessions with checkpointed JSONL transcripts
- scheduled or webhook-triggered Routines on managed cloud
- subagent-delegated coding tasks in isolated contexts
Key concepts
- CLAUDE.md memory → cross-session-memory (docs) — Persistent project/user instructions, read at session start.
- Skills → agent-skills (docs) — Packaged SKILL.md workflows loaded on demand, replacing custom commands.
- Subagents → subagent-isolation (docs) — Specialised agents in isolated context windows, delegated to via description.
- Hooks → event-driven-agent (docs) — Shell/HTTP/LLM handlers that fire on lifecycle events.
- MCP → mcp (docs) — Connect external tools/data via Model Context Protocol.
- Routines → scheduled-agent (docs) — Cloud-scheduled or webhook-triggered Claude Code sessions on Anthropic infra.
Patterns this full-code implements —
- ★Agent-Computer Interface
Single ACI across terminal, IDE, desktop, and browser surfaces; shared engine + CLAUDE.md + MCP across all.
- ★★Agent Resumption
Sessions saved continuously to local JSONL transcripts; resumable via --continue or --resume.
- ★Agent Skills
SKILL.md files extend Claude's toolkit; conforms to the Agent Skills open standard.
- ★★Approval Queue
Permission rules gate skill/tool execution; allowed-tools pre-approves; deny rules block; managed settings can enforce centrally.
- ★★Cross-Session Memory
CLAUDE.md (user-authored) + auto memory (agent-authored) carry knowledge across sessions; both load at the start of every conversation.
- ★★Event-Driven Agent
Hooks fire on lifecycle events with JSON context passed to user-defined handlers.
- ★★Extended Thinking
Extended thinking + `ultrathink` trigger phrase + effort levels.
- ★★Model Context Protocol
Full MCP client: stdio/SSE/HTTP transports, three scopes (local/project/user), OAuth, dynamic tool updates, automatic reconnect, tool search defaults on.
- ★★Multi-Model Routing
opusplan alias splits Opus (plan) / Sonnet (execute); skill frontmatter `model` field overrides per skill; supports Anthropic/Bedrock/Vertex/Foundry providers.
- ★★Plan-and-Execute
Plan mode runs a read-only exploration; on accept, execution mode runs the tool loop.
- ★★ReAct
ReAct shape is implicit throughout — Claude plans, writes, verifies; bug-fix loop traces issue → identifies root cause → implements fix.
- ★★Scheduled Agent
Routines run on Anthropic-managed infra; cron, API, or GitHub-event triggered.
- ★Subagent Isolation
Each subagent runs in its own context window with custom prompt, tool access, and permissions.
- ★Todo-List-Driven Autonomous Agent
Built-in TaskCreate/TaskGet/TaskList/TaskUpdate tools manage the session task checklist; TodoWrite is the prior primitive, still toggleable.
- ★★Tool Use
Documented tool table: Read/Edit/Bash/Glob/Grep/WebSearch/WebFetch/Monitor/Agent/MCP tools etc., each with permission rules.
- ★★Step Budget
Effort-level scale (low/medium/high/xhigh/max) is a token/cost budget; MAX_MCP_OUTPUT_TOKENS and tool-search threshold are explicit context budgets; no documented step-count cap.
- ★★Tool Result Caching
Honest downgrade — Claude Code uses prompt caching (system prompt + tool defs + history) and WebFetch 15-min cache, not tool-result deduplication.
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.
Instantiates
Alternatives & relatives
- full-code · framework
Aider
Peer terminal coding agent — single-model loop, no subagents/skills.
- full-code · framework
Codex CLI
Peer first-party (OpenAI) terminal coding agent.
- full-code · framework
Cursor
Peer IDE-first coding agent.
- full-code · framework
Cline
Peer open-source IDE+CLI agent, similar plan/act split.
- full-code · framework
Continue
Peer IDE extension.
- full-code · framework
Claude Agent SDK
The SDK exposing the same engine as a programmable library.
Listed as alternative by (14)
- full-code · framework
- full-code · framework
- full-code · framework
- full-code · framework
- full-code · framework
- full-code · framework
- app · closed product
Devin
- app · closed product
GitHub Copilot Coding Agent
- full-code · framework
Goose
- full-code · framework
JetBrains Junie
- full-code · framework
Open Interpreter
- full-code · framework
OpenHands
- full-code · framework
- full-code · framework
Windsurf
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: verified