ctxcop

Keep secrets out of AI coding agents' context windows. Supports Claude Code, Codex CLI, Cursor, Pi, OpenCode, and Aider (narrower coverage — see docs/harnesses.md).

When the agent runs a Bash tool call:

Agent → Bash:  aws sts get-session-token

what reaches the model (after ctxcop's PreToolUse hook):

{
    "Credentials": {
        "AccessKeyId":     "<REDACTED:ctxcop-aws-access-key:OLIA>",
        "SecretAccessKey": "<REDACTED:aws-secret-key:fA31>",
        "SessionToken":    "<REDACTED:ctxcop-aws-session-token:7g8K>",
        "Expiration":      "2026-05-21T22:17:48Z"
    }
}
[ctxcop] redacted 3 secret(s) before this output reached the model.

Your own interactive shell is untouched. ctxcop only intercepts the tool calls the agent makes through its harness.

ctxcop hooks into each harness's lifecycle events where a credential might enter (or exit) the conversation and applies the right defense: rewrite the call so the secret never reaches the model, deny it with actionable retry guidance, or warn after the fact when neither is possible. Detection uses betterleaks (gitleaks' successor) plus an embedded high-confidence ruleset, with recursive base64/hex/percent/unicode decoding and user/project overlays.

---

Install

Build from source for now — prebuilt binary releases are on hold until macOS Developer ID codesigning + notarization are wired into the release pipeline (Gatekeeper rejects the ad-hoc-signed binaries a plain go build produces once they've been through a download/quarantine flow; a locally built binary isn't affected).

# 1. go install.
go install github.com/webflow/ctxcop/cmd/ctxcop@latest

# 2. From source.
git clone https://github.com/webflow/ctxcop && cd ctxcop
go build -o /usr/local/bin/ctxcop ./cmd/ctxcop

Wire it up

ctxcop install                       # autodetect ~/.claude, ~/.codex, ~/.cursor, ~/.pi, ~/.config/opencode, ~/.aider.conf.yml; prompt before writing
ctxcop install --harness=cursor      # one harness only; skips prompt
ctxcop install --harness=aider       # aider integration is static-config; see docs/harnesses.md#aider
ctxcop install --yes                 # autodetect, skip prompt (CI)

Idempotent — prior ctxcop entries are replaced cleanly; unrelated hooks, model settings, and MCP config are preserved. A fresh harness session picks up the hooks; the first SessionStart additionalContext block primes the agent.

Verify

$ printf 'AWS_ACCESS_KEY_ID=%sLALEMEL33243OLIA\n' AKIA | ctxcop scan
AWS_ACCESS_KEY_ID=<REDACTED:ctxcop-aws-access-key:OLIA>

Uninstall

ctxcop uninstall                     # prompts; removes from every detected harness
ctxcop uninstall --harness=cursor    # one harness

Run ctxcop uninstall before removing the binary itself. Otherwise each harness exec's a path that no longer exists. Most fail open, but you'll see log noise.

---

Configuration

Runtime behavior via environment variables; install-time behavior via flags. Nothing retained on disk by default.

Runtime env vars

Variable	Default	Purpose
CTXCOP_AUDIT_LOG	unset	Path to an append-only JSONL log. Unset = no logging. One line per event: {ts, tool, action, rules, count, field}. Mode 0600.
CTXCOP_RULES	unset	Path to an extra TOML rule file.
CTXCOP_DISABLE_RULES	unset	Comma-separated rule IDs to remove from the composed ruleset.
CTXCOP_POSTTOOLUSE	empty	Set to off to disable Claude Code's PostToolUse warning hook.
CTXCOP_POSTTOOLUSE_ALLOW	unset	Comma-separated tool-name globs (only * wildcard) whose responses should not trigger Claude Code's PostToolUse notice. The audit log records warned-suppressed instead.
CTXCOP_SKIP_PATHS	unset	Comma-separated glob list of paths where Read/Write/Edit/NotebookEdit hooks should not scan. Adds to baked-in defaults (testdata/, fixtures/, *_test.*, etc.).
CTXCOP_DEV	empty	Set to warn to downgrade Write/Edit/WebFetch/MCP blocks to allow + warning. UserPromptSubmit and Bash/Read paths unaffected.

Install-time flags

Flag	Default	Purpose
--harness=auto|claude-code|codex|cursor|pi|opencode|aider|all	auto (install) / all (uninstall)	Which harnesses to write to.
--scope=user|project	user	user writes to $HOME/.<harness>/. project writes to ./.<harness>/ for repo-local hook configs.
--yes / -y	off	Skip the confirm-before-write prompt. Required for non-TTY invocations — without it, ctxcop fails closed rather than silently confirming a piped "yes".

---

Documentation

• docs/harnesses.md — per-harness hook coverage and steering guidance
• docs/detection.md — rules, overlays, fixtures, project config
• docs/hook-contracts.md — JSON wire shapes per harness event
• docs/architecture.md — code structure and invariants
• docs/known-limits.md — documented bypasses and gaps
• docs/verify-reproducibility.md — rebuild a tag from source and compare sha256
• SECURITY.md — vulnerability disclosure, embargo terms, build verification
• THREATMODEL.md — trust model, in-scope and out-of-scope threats
• THIRD_PARTY_AUDIT.md — dependency review log
• CHANGELOG.md — release notes
• ROADMAP.md — planned work
• testing/ — manual end-to-end fixtures for harness adapters

Development

go test ./...
go vet ./...
go build -o ctxcop ./cmd/ctxcop

See CONTRIBUTING.md for DCO sign-off, signed commits, and pre-merge gates. Maintainer team is @arr-wf plus @webflow/infrastructure-security; routing per .github/CODEOWNERS.

License

MIT — see LICENSE.

ctxcop is built on the betterleaks secret-scanning engine (MIT, © Zachary Rice). Full third-party attributions are in NOTICES.md.