Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/wz/agent-gate/claude-mdgit clone --depth 1 https://github.com/WZ/agent-gateWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.04691 | $0.04691 |
| Opus 5 | $0.02346 | $0.02346 |
| Sonnet 5 | $0.00938 | $0.00938 |
| Haiku 4.5 | $0.00469 | $0.00469 |
Grade A, and why
agent-gate CLAUDE.md scanned grade A with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Makes network callslowCapability
Not a fault in itself. Listed so you know the mod talks to something, and to what.
│ launcher ──spawn──▶ agent (claude, curl, gh, ...) │ How it starts
The opening of the file, as written. The whole thing — 341 lines — stays where its author put it; the contents beside it link to each section on GitHub.
CLAUDE.md — agent-gate
Project memory for AI assistants working on this repo. Read this before making non-trivial changes.
What this is
agent-gate is a personal, single-user audit gate for Claude Code's outbound
HTTPS. It TLS-intercepts every request the agent (and its subprocesses) makes,
parses + flags + persists the flow, and serves a local web dashboard for
review. It does not ship as a service, ship to a backend, or run as
multi-tenant. Everything is loopback, single-user, local disk.
Three subsystems share one binary:
┌──────────────────────── agent-gate ────────────────────────┐
│ │
│ launcher ──spawn──▶ agent (claude, curl, gh, ...) │
│ (sandbox/jail) │ │
│ │ all egress forced here │
│ ▼ │
│ proxy :8888 ──parse──▶ policy ──▶ store (jsonl+sqlite) │
│ │ │
│ dashboard :7878 ◀────────────────────────┘ │
└────────────────────────────────────────────────────────────┘
The launcher physically forces egress through the proxy via a per-platform network jail (macOS sandbox-exec, Linux user namespace, Windows scaffolded for Plan 4). The proxy is a TLS-MITM forward proxy. The dashboard is read-only against the store.
Critical rules — never break these
- Audit-log completeness. Every captured flow MUST land in the JSONL +
SQLite store. The proxy emits on a buffered channel; if the consumer falls
behind, the proxy slows down (and upstream may time out). Drop-on-full is
a correctness bug, not a knob. See
internal/proxy/proxy.go. - Fail-closed posture. If the proxy crashes during
agent-gate run, the agent's network goes dark. The supervisor's teardown path may not silently swallow this; the child must die when the proxy dies. Seeinternal/launcher/supervisor.go. - Loopback only. The proxy refuses to bind a non-loopback address; same
for the dashboard. See
isLoopbackAddrincmd/agent-gate/proxy_cmd.goandisLoopbackininternal/dashboard/server.go. Never relax this. - File modes. CA private key + JSONL + dismissals.json must be 0600;
config + data dirs must be 0700. The CA package refuses to start if
key.pemis wider than 0600 — gated onruntime.GOOS != "windows"since Windows ignores unix mode bits. macOS + Linux enforce; Windows skips both in the production check (internal/ca/ca.go) and the unit test (internal/ca/ca_test.go). - No secrets in commits.
git addfiles by name, nevergit add -Aorgit add .. The user's data dir contains real captured prompts and responses; an accidentaladd -Awould commit them. - Don't push without permission. Local merge + tag is fine; pushing to
originalways needs explicit user authorization (per the user's global CLAUDE.md). - Always execute in a git worktree. Every code change, test run with
side effects, and binary build happens inside
.worktrees/<branch>/, never in the main checkout. The main checkout is reserved for reading, git operations, and merging PRs. The worktree is created BEFORE work starts — never skip it "just for a small change" and never infer it later. When invoking subagents to execute a plan, instruct them to cd into the worktree as their first step.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 341 lines · 4,691 tokens per session scan A de9742d13727
agent-gate CLAUDE.md is an instructions file published in the GitHub repository WZ/agent-gate (2 stars, last pushed 3mo ago), licensed Apache-2.0. It adds 4,691 tokens to every session, about $0.0235 per session on Opus 5. A static security scan graded it A with 1 finding (makes network calls). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other instructions, from other repositories
fkf AGENTS.md
Instructions for fmind/fkf, covering agents.md (project), status, workflow, product boundaries and trust and execution.
codex-tg AGENTS.md
Instructions for mideco-tech/codex-tg, covering agents, repository purpose, working mode, agent workflow kernel and context pull map.
corresync AGENTS.md
AGENTS.md instructions for nkiyohara/corresync, covering repository instructions, scope, architecture invariants, working agreement and agent and model operating guide.
Zenith AGENTS.md
AGENTS.md instructions for bigduu/Zenith, covering repository guidelines, project structure & module organization, shared memory via jiandu mcp, build, test, and development commands and coding style & naming conventions.
Zenith CLAUDE.md
Claude Code instructions for bigduu/Zenith, a project described as: Nine-submodule workspace for a local-first AI system: Bamboo runtime, Lotus/Bodhi UX, Jiandu memory, Nova computer use, Magpie connectors, optional hosted services, and docs.
fkf CLAUDE.md
Instructions for fmind/fkf, a project described as: Local-first Go CLI that turns plain JSON and Markdown work history into reproducible context packs for coding agents.