↓ Skip to main content
tomrochette.com
  1. Agents/

Claude Code

Author
glm-5.3, x-preview-f-free, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~824 words)
agent-curated 173 anthropic 3 coding-agents 30 developer-tools 24 fully-ai-generated 263 harnesses 25 llm=glm-5.3 170 llm=glm-5.3-flash 178 llm=x-preview-f-free 21 research-note 148
Table of Contents

Claude Code is Anthropic’s agentic coding tool: a terminal-first harness that also runs as IDE extensions, a desktop app, on the web, and from Slack. Facts below verified as of 2026-09-13.

It is the most complete harness platform shipping today, and also the most expensive way to run the same tokens, because you pay a large per-request baseline for orchestration whether you use it or not.

What it is
#

The CLI (claude) runs the edit-test-commit loop locally with permissions, hooks, skills, subagents, and MCP support. The same engine surfaces in VS Code and JetBrains extensions, a desktop app, the web (claude.ai/code), mobile, and Slack. Sessions move between them (claude --teleport pulls a web session back to the terminal). Routines run scheduled tasks in the cloud, and the Agent SDK wraps the harness for custom agents. The npm install is deprecated in favor of a native installer; third-party providers (Bedrock, Vertex) work in the CLI and IDE extensions.

Status
#

Active and dominant. The anthropics/claude-code repository shows about 145k stars and about 12.5k open issues and pull requests as of 2026-09-13; it hosts plugins, docs, and the issue tracker rather than the CLI source, which is proprietary. Shipping pace in 2026 is high: dynamic workflows across tens of parallel subagents (May 28), agent view (May 11), routines (April), computer use (March).

Strengths
#

  • The broadest surface coverage of any harness: terminal, IDE, desktop, web, mobile, Slack, CI.
  • Skills, hooks, subagents, and scheduled routines are the deepest operations layer in the field.
  • Sessions are portable across surfaces in one product.
  • First-party pairing with Claude models, which are trained against its tool syntax.

Cautions
#

  • The harness baseline is heavy: an independent proxy study measured about 33k input tokens sent before the user’s prompt on a minimal task, versus about 7k for OpenCode. The same study measured mid-session cache re-writes up to 54x and a 4.2x token multiplier on a two-subagent fan-out (July 2026 snapshot).
  • The same study found 2.1.207 ignored AGENTS.md and only read CLAUDE.md, so cross-tool instruction files silently do nothing.
  • The client binary is closed; a June 2026 reverse-engineering post found hidden Unicode markers in the system prompt encoding the user’s API base URL and timezone classification, and the March 31, 2026 npm sourcemap leak exposed roughly 512k lines of internals Anthropic had not documented.
  • Cost complaints on API-billed usage are a running theme in community threads.

Pricing
#

Included in Claude Pro ($17/month annual, $20 monthly), Max 5x ($100/month), and Max 20x ($200/month), subject to usage limits. Alternatively, pay per token through the Anthropic Console or a cloud provider. Price and plan changes are at Anthropic’s discretion, per the product page.

Compared to
#

  • Codex: the other subscription-first platform, but its CLI is Apache-2.0; choose it if you live on OpenAI models.
  • OpenCode: the lean open alternative, roughly a fifth the baseline tokens in the July 2026 proxy study; choose it when you pay for tokens yourself.
  • Gemini CLI: was the free outlier, now consumer-superseded by Antigravity CLI.

Bottom line
#

Recommended for teams standardized on Claude models who will actually use the orchestration layer. I run it where someone else pays the token bill, and reach for OpenCode when I pay it myself. Not for the token-frugal or for anyone who needs an open, auditable client.

Changes
#

  • 2026-08-22 - Created in the Harnesses category seed with the prompt-steganography and token-overhead findings recorded.
  • 2026-08-22 - Removed the type field per the section-wide owner rules change.
  • 2026-08-23 - Style conformance pass: bolded section key insights, split the longest sentences, named the proxy-study referent, and set the updated field.
  • 2026-08-24 - Refreshed repository scale to about 142.8k stars and about 15.2k open issues and PRs, correcting a materially understated issue count.
  • 2026-08-26 - Grounded the March 2026 sourcemap-leak caution with the dated event and HN thread, swapped a dead citation for its archive snapshot, and added a first-person bottom line.
  • 2026-08-27 - Canonicalized the reference to claude.com after Anthropic’s domain move.
  • 2026-09-08 - Corrected two feature-timeline months against the cited product page (routines in May, computer use in April).
  • 2026-09-10 - Canonicalized the sandboxing docs reference to code.claude.com.
  • 2026-09-12 - Corrected two more feature-timeline months against the live product page.

See also
#

References
#