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/moberghr/mtk-agent-toolkit/claude-mdgit clone --depth 1 https://github.com/moberghr/mtk-agent-toolkitWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/instructions/moberghr/mtk-agent-toolkit/claude-md)<a href="https://agentmods.dev/instructions/moberghr/mtk-agent-toolkit/claude-md"><img src="https://agentmods.dev/badge/instructions/moberghr/mtk-agent-toolkit/claude-md.svg" alt="Measured on agentmods" height="20"></a>What 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 5.1 | $0.03031 | $0.03031 |
| Opus 5 | $0.01515 | $0.01515 |
| Sonnet 5 | $0.00606 | $0.00606 |
| Haiku 4.5 | $0.00303 | $0.00303 |
Grade A, and why
mtk-agent-toolkit CLAUDE.md scanned grade A with 0 findings 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 5d 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.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 107 lines — stays where its author put it; the contents beside it link to each section on GitHub.
claude-helpers — MTK Standards
Version history and per-release notes: see CHANGELOG.md. (Kept out of this file so the always-loaded prefix stays prompt-cache-stable across releases.)
This file +
.claude/rules/are the source of truth for AI agents. Detailed standards live in.claude/rules/. Reference docs live in.claude/references/(shared) and.claude/references/{stack}/(stack-specific).
Skill Routing
| What you need | Command | When |
|---|---|---|
| First-time repo setup | /mtk-setup |
Bootstrap — detects tech stack, pulls guidelines, generates CLAUDE.md and architecture-principles.md |
| Re-run audit | /mtk-setup --audit |
Refresh .claude/references/architecture-principles.md after architectural change |
| Merge multi-repo audits | /mtk-setup --merge |
Unify per-repo audits in .claude/references/audits/ into a team-wide doc |
| Everything else | /mtk <description> |
Natural language — routes to fix / implement / pre-commit-review / repo-health / context-report |
| Periodic readiness check | /mtk repo-health or bash scripts/repo-health-score.sh |
12-asset scorecard + PR review mining (last 10 merged PRs) |
| Validate toolkit | bash scripts/validate-toolkit.sh |
Before every commit — structural check of manifest, plugin.json, and skill anatomy |
| Install health check | /mtk-doctor |
PASS/WARN/FAIL diagnostics across core files, components, hooks, and environment fit; --json for CI, --fix for safe auto-repairs |
| Promote a lesson | /promote-lesson |
Promote a personal lesson from .claude/lessons/personal.md or Claude Code native memory (~/.claude/projects/<sanitized-cwd>/memory/) to team-wide tasks/lessons.md; optionally open a validated contribute-back PR to the toolkit |
| Mine lessons from past sessions | /mtk mine lessons |
Sweep recent session transcripts for durable lesson/memory candidates (reject-by-default rubric, suggest-only) |
| Disable tier-2 hooks | MTK_HOOKS_TIER2=0 in .claude/settings.local.json env |
Silences skill-invoking hooks (queue + drain) without touching shared settings |
| Enforce spec scope (hard deny) | MTK_SCOPE_GUARD_ENFORCE=1 in .claude/settings.local.json env |
Upgrades hooks/scope-guard.sh from advisory to a hard PreToolUse deny (exit 2) when an Edit/Write targets a file outside the approved spec's change_manifest/test_manifest. Default (unset) stays advisory. See docs/competitive-analysis-2026-07.md P0#1 |
| Auto-approve safe plans | MTK_AUTO_PROCEED=1 in .claude/settings.local.json env |
Skips Phase 2.5 prompt only when spec has no open decisions and no plan-gap BLOCKING findings |
| Disable artifact publishing | MTK_ARTIFACT_PUBLISH=0 in .claude/settings.local.json env |
Stops workflow skills publishing spec/plan/handoff/health to a claude.ai Artifact (data-egress opt-out for regulated repos); disk output is unaffected. See .claude/references/artifact-publishing.md |
| Enable compaction snapshots (plugin installs) | MTK_COMPACT_SNAPSHOT=1 in .claude/settings.local.json env |
Opts a plugin-installed repo into pre-compaction git-stash snapshots; always on in this dev checkout |
| Pin MTK to a checkout | MTK_HELPER_ROOT=/path/to/claude-helpers in .claude/settings.local.json env (or the shell) |
Makes MTK resolve from that checkout first, before the project copy and the plugin cache — covering both the ## MTK File Resolution block every entry-point skill opens with (skill + reference reads) and the inline script resolvers (workflow-artifact.sh, learnings.sh). For dogfooding MTK from a separate clone with $CLAUDE_PLUGIN_ROOT unset, or pinning one version out of a multi-version plugin cache. Note: target-repo scripts (spec-archive.sh, lint-ears.sh) resolve the project root from $CLAUDE_PROJECT_DIR/git, so their output always lands in the target regardless of where the script lives |
| Tune the mtk-compress nag | MTK_COMPRESS_MAX_NAGS=N in .claude/settings.local.json env |
Per-session budget for hooks/compress-monitor.sh's "pipe this through mtk-compress" tip (default 1). The tip is worth one read; repeating it on every large output trains you to ignore every hook, so it self-silences after the budget. 0 silences it while leaving the hook wired; MTK_COMPRESS_MONITOR_DISABLED=1 disables it outright, and MTK_COMPRESS_WARN_CHARS moves the 5,000-char trigger |
| Allow interactive-prone shell commands | MTK_INTERACTIVE_GUARD=0 in .claude/settings.local.json env |
Disables hooks/interactive-guard.sh, the PreToolUse hard deny on Bash commands that can block on a prompt (S4.12) — gh pr merge with no --delete-branch/--no-delete-branch decision, and prompt-capable commands piped through tail/head where the prompt is buffered out of sight. Read-only pipes (gh pr view … | tail) are never blocked |
| Declare that subagent dispatch is unavailable | MTK_SUBAGENT_DISPATCH=0 in .claude/settings.local.json env |
Tells implement's Phase 2.9 pre-flight up front that this session cannot dispatch implementer subagents — the harness does not expose Agent/Workflow, or a standing instruction forbids subagents the engineer did not ask for. HIGH/MAX then runs the inline-MAX profile (compensations C1–C3) instead of prescribing a path that cannot run and logging the same ceremony reduction once per run. It does not lower the rigor level: the level reflects the change's blast radius, the path reflects the session's tooling. Default (unset) probes per run |
| Force or skip the pre-flight baseline | MTK_BASELINE_CAPTURE=1 / =0 in .claude/settings.local.json env |
implement Phase 2.9 captures build/test/typecheck state at the base commit before the first edit, so a failure found at batch 3 is attributable instead of needing a throwaway worktree to prove it pre-existing. On by default at rigor HIGH/MAX; =1 forces it at any level, =0 opts out — and then every later checkpoint must state that no baseline exists rather than implying one |
| Tune the collateral-churn thresholds | MTK_COLLATERAL_WS_MIN / MTK_COLLATERAL_WS_RATIO / MTK_COLLATERAL_GEN_MIN / MTK_COLLATERAL_ASSET_PCT / MTK_COLLATERAL_ASSET_MIN / MTK_COLLATERAL_FMT_MIN / MTK_COLLATERAL_FMT_RATIO in .claude/settings.local.json env |
Thresholds for hooks/collateral-guard.sh, which flags churn that is not the change you made: whitespace/EOL-only rewrites (a small edit to a CRLF file rewrites the whole file), generated artifacts riding along undeclared (one npm install drags a lockfile into a feature commit), asset directories regenerated wholesale, and structured files re-serialized around a small real edit (a big JSON diff whose parsed content barely changed — the class a whitespace comparison cannot see). Wired into pre-commit-review step 4.6 and the spec-drift-detection collateral axis |
| Write a tracked run receipt | MTK_RUN_RECEIPT=1 in .claude/settings.local.json env |
implement Phase 7.5 writes docs/specs/<date>-<slug>.receipt.md — a tracked sibling of the spec holding the run's evidence (baseline vs final figures, gates and their reasons, dispatch path and compensations, ceremony reductions, drift/coverage/collateral verdicts, reviewer lane outcomes). The workflow artifact under .mtk/ is gitignored, so without this the only tracked residue of a run is tasks/lessons.md and the spec. Fields never recorded are written as not recorded, never reconstructed |
| Calibrate context-budget nags to your window | MTK_CONTEXT_WINDOW_TOKENS=1000000 in .claude/settings.local.json env |
Rescales context-budget.sh's file/mod/op nudges from their 200k baseline to your model's real context window, so a large-context (e.g. 1M) model isn't nudged to checkpoint or hand off mid-task. Per-threshold overrides: MTK_CTX_FILES_WARN / MTK_CTX_MODS_WARN / MTK_CTX_OPS_WARN; MTK_CONTEXT_BUDGET_PCT sets the read-bytes reset percentage |
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.
- 5d ago First seen · 107 lines · 3,031 tokens per session scan A d12d443455d6
mtk-agent-toolkit CLAUDE.md is an instructions file published in the GitHub repository moberghr/mtk-agent-toolkit (7 stars, last pushed yesterday), licensed MIT. It adds 3,031 tokens to every session, about $0.0152 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other instructions, from other repositories
vscode buildNext.instructions.md
Working notes and architecture documentation for the new esbuild-based build system in build/next. Use when making changes to the new build pipeline (transpile/bundle commands, NLS plugin, source-map handling, resource copying, or self-hosting watch tasks).
spec-kit AGENTS.md
AGENTS.md instructions for github/spec-kit, covering agents.md, about spec kit and specify, quickstart — add a new integration in 5 steps, integration architecture and integrationmanifest — file tracking.
next.js AGENTS.md
AGENTS.md instructions for vercel/next.js, covering next.js development guide, codebase structure, monorepo overview, core package: packages/next and other important packages.
codex AGENTS.md
AGENTS.md instructions for openai/codex, covering rust/codex-rs, the codex-core crate, code review rules, crate api surface and model visible context.
langchain AGENTS.md
AGENTS.md instructions for langchain-ai/langchain, covering global development guidelines for the langchain monorepo, corridor security analysis, project architecture and context, monorepo structure and development tools & commands.
vscode oss-third-party-notices.instructions.md
Instructions for microsoft/vscode, covering vs code oss third-party-notices pipeline, architecture, pipeline flow in ci, applying the notice (cutover) and fallback chain (never fail the build).