Claude Code

Anthropic's coding agent for repo files, commands and git workflows; optional headless runs use configured permissions, not a universal human gate.

Choose it when

You want a repo-aware coding agent for interactive work or a bounded headless script, with tool permissions chosen for the job.

Skip it when

You need offline-only model use, a guaranteed approval before every action, or unattended deployment without a human owner for failures.

01 / Classification

Autonomy evidence

This is our classification, not a vendor badge. The record says what the product normally does. The gaps stay visible.

How we classify it

Tool use + multi-step plans, still human-gated per session.

Runs unattended

Yes — that capability is documented in this brief.

Trigger

Normally a user starts a coding session; scripts/CI can call claude -p non-interactively. The overview separately documents optional cloud scheduled tasks (formerly routines), Desktop local schedules and within-session /loop, not one universal always-on CLI default.

State

Project CLAUDE.md/AGENTS.md instructions and auto memory supply continuity; headless callers can continue or resume a conversation by session ID. This is documented, not a tested recovery guarantee.

Tools

Reads and edits repo files, runs commands and works with git. Headless --allowedTools pre-approves selected tools, subject to permission mode and special checks; it doesn't mean every tool call requires a fresh human approval.

Human control

Rules and permission mode govern execution; auto uses a classifier, acceptEdits accepts file edits, and bypass skips most prompts. With v2.1.302+, headless runs fall back to Manual when no flag or configuration sets a mode; calls needing approval are denied unless approved through rules, hooks or a permission host. An unattended caller must inspect exit status, permission denials and structured_output presence and own failures; no human-prompt or automatic-recovery guarantee is implied.

02 / Operations

Control reality

Useful agents eventually touch something real. This is the operational picture we have, not a promise hidden in a sales page.

Where it runs

cli · ide

Data boundary

Local CLI does not mean offline inference or local-only data: the overview lists subscription, Console and supported third-party access, and the repository describes usage/conversation feedback collection. Provider-specific retention and export controls aren't established here.

Identity & permissions

Not recorded yet.

What it can write

repo files · shell · git

Surfaces

Terminal · VS Code / Cursor extension · Desktop · Browser

Export / lock-in

Not recorded yet.

03 / Receipts

Evidence ledger

Here is what this brief is built on. “Documented” means we found a source. “Tested” means this desk actually ran it. Those are different things.

Evidence status

documented

Source set

8 primary sources

Evidence review

2026-10-21

Sources on file

Corrections

2026-10-14 · Updated stale release metadata to upstream v2.1.299. Replaced blanket human-gated/no-unattended wording with mode-dependent permissions and optional headless automation. Distinguished headless calls from separately documented scheduling surfaces and local CLI location from provider/data boundaries; no runtime test was performed.
2026-10-21 · Refreshed upstream release metadata to v2.1.306 using API published_at, not an installed-version observation. Added the v2.1.302+ headless Manual fallback, explicit configuration and mod/credential caveats, structured-output checks and current cloud scheduled-task naming. Upstream reports fixes for cloud-session Not logged in reporting and many-MCP-tool error/compaction handling; not runtime-tested and not a universal recovery guarantee.

04 / Change log

Related coverage

The brief tells you what the product is. Payloads track the changes that may alter the decision.