Teamday
Claude logo

Harness · Anthropic

Claude Code.

Anthropic's official agentic coding harness. The one we reach for when a mission needs to run for hours, verify its own work, and not lie about being done.

Read the Opus 4.7 launch →All harnesses

What it is

Claude Code is the agentic harness Anthropic ships alongside the Claude family. It wraps the model with the things you actually need to ship work end-to-end: a planning loop, parallel tool calls, file I/O, shell execution, and a self-verification step that re-reads its own output before claiming a task is complete.

Originally released as a CLI for developers, Claude Code now runs inside TeamDay — around the clock, on every agent that selects it as a harness.

Models it runs

  • Claude Opus 4.7 — the newest frontier model, released April 16, 2026. 1M context window, deepest reasoning. Read the model card.
  • Claude Opus 4.6 — the previous flagship. Still a great choice when 4.7 is overkill.
  • Claude Sonnet 4.6 — near-Opus quality at a fraction of the price. The volume horse for content and routine engineering loops.

What makes it distinct

  • Agentic loop. Plan → act → verify results, with parallel tool calls and the ability to re-plan when a step fails.
  • Long-horizon work. Comfortable with missions that span hours and dozens of tool calls without losing the thread.
  • Computer use (macOS, research preview). Screen, keyboard, and mouse control for tasks that have no API — available on Pro and Max plans.
  • 1M context on Opus 4.7 and Sonnet 4.6. Read entire codebases, multi-thousand-page PDFs, or a quarter of CRM history in one prompt.

Capabilities at a glance

  • Sub-agents: yes — define your own specialized agents with their own tool allowlists; a lead agent coordinates and delegates.
  • Skills: yes — packaged reusable instructions with progressive disclosure, stored in .claude/skills/ with YAML frontmatter.
  • MCP servers: yes — full client over stdio, SSE, and HTTP, with OAuth and env-var expansion.
  • Hooks: yes — including PreToolUse, PostToolUse, Stop, SubagentStop, SessionStart/End, UserPromptSubmit, PreCompact, Notification, and many more.
  • Slash commands: yes — user-defined in .claude/commands/ or .claude/skills/ with YAML frontmatter, allowed-tools lists, and arguments.
  • Permissions / sandboxing: fine-grained allow/ask/deny per tool; managed settings for enterprise (MDM); strict-security and bash-sandbox profiles.
  • Plugins: yes — installable via /plugin from official and community marketplaces.
  • Multi-model: Anthropic API, Amazon Bedrock, Google Vertex AI, Microsoft Foundry; per-session model override.
  • Sessions: persistent with claude --resume; --fork-session branches the timeline; Remote Control moves a session across devices.
  • Surfaces: CLI, VS Code, JetBrains, macOS/Windows desktop, claude.ai/code web — all sharing the same engine.
  • Headless / SDK: yes — Claude Agent SDK for Python and TypeScript; -p for non-interactive mode.
  • License: proprietary engine under Anthropic ToS; plugins, hooks, skills examples are open.

How TeamDay uses it

Selecting Claude Code as a TeamDay agent's harness unlocks every Claude model in the dropdown. New Anthropic releases (4.8, 5.0, future Mythos) light up automatically — no TeamDay-side integration sprint, no waiting for a connector update.

  1. Open an agent → Settings → Harness → Claude Code.
  2. Pick the model. Default is Sonnet 4.6 for cost; Opus 4.7 for high-stakes missions.
  3. Attach MCP servers (e.g. Postgres, GitHub, the media MCP) so the agent can act on your stack.
  4. Run a mission. The agent plans, executes, verifies, and reports back.

When to pick Claude Code

  • Multi-hour engineering missions where "looks done" must equal "is done."
  • Schema migrations, security review, customer-facing copy with legal stakes.
  • Long-context analysis (full codebase, full CRM history, full discovery thread).
  • Anything where the cost of a wrong claim is high.

How it's benchmarked

Coding harnesses are measured on Terminal-Bench (tbench.ai) — the standard evaluation suite for end-to-end terminal tasks: file edits, shell commands, builds, test runs, error recovery. When you see vendors quoting "X% on Terminal-Bench Pro," it's the same harness-and-model combination you'd run inside TeamDay.

When to pick something else

  • Codex — for unified multimodal missions (text + screenshots + audio + video in one call) and computer-use heavy work.
  • Gemini CLI — when you need 2M context, Google Workspace integration, or BigQuery-native work.