agent-ops

Small, dependency-light operational tools for AI coding agents.

A coding agent edits your source, runs your shell, reads whatever files it can reach, and spends money — at machine speed, often unattended. That's a production system, and it needs the same operational layer you'd give anything else with that much reach: recover, prevent, observe, spend, secure, set up, verify.

These are the tools we built filling each gap, one real failure at a time, running autonomous agent fleets on Claude Code in production. Each is a single script (Python 3 or bash), no install, no dependencies beyond a stock interpreter. Each links to a write-up explaining the failure it prevents and the data behind it.

Read the thesis: Your AI coding agent needs an operational layer

Install

git clone https://github.com/muthuishere/agent-ops
# each tool is self-contained — run it directly:
python3 agent-ops/mcp-budget/mcp-budget .
bash    agent-ops/preflight/preflight

The tools

Recover & protect agent work

Tool	What it does	The story
recover	get an agent's vanished commit back — reflog + fsck recovery for work you thought was gone	read →
unwedge	rescue a Claude Code session stuck on the thinking-block 400 ('thinking blocks cannot be modified')	read →
auto-snapshot	make an agent's uncommitted work impossible to lose — auto-snapshot the working tree	read →
preflight	ask git before your agents collide — a worktree/branch collision pre-check	read →
worktree-doctor	flag agent worktrees that would silently lose work on cleanup	read →
mergerase	find public capabilities that silently disappeared across a parallel-agent merge	read →

See your fleet (observe)

Tool	What it does	The story
fleet-ps	ps for your Claude Code agent fleet — which sessions are alive, on what branch, healthy or wedged	read →
session-stats	read what a Claude Code session actually did, straight from the transcript	read →
review-brief	a reviewer brief from the transcript — close the AI-code verification gap	read →

Prevent & guard

Tool	What it does	The story
agent-guard	a PreToolUse hook that blocks destructive agent commands (git reset --hard, rm -rf, push --force)	read →
agent-frisk	find the invisible Unicode instructions that hijack a coding agent	read →
forkcap	a fork-budget guard that caps Claude Code's runaway nested subagents	read →

Secure & audit config

Tool	What it does	The story
agent-reach	measure a coding agent's secret blast radius — which secret stores are in reach	read →
perm-audit	flag over-permissive Claude Code settings.json permissions	read →
trifecta-scan	cross-config lethal-trifecta scanner — catch read+untrusted+exfil reachable in one grant set	read →
mcp-audit	lint your .mcp.json — the highest-trust, never-linted agent config — for egress/security surface	read →

Context & cost

Tool	What it does	The story
mcp-budget	measure the context tax of your MCP servers — tokens spent before your first prompt	read →
agent-waste	audit a session transcript for wasted tokens — re-reads, redundant tool results	read →
context-carry	the cost of a read is its residence time — weight every result by tokens x turns re-sent	read →
cachecost	prompt-caching economics — what caching is actually saving you	read →
context-lint	audit your CLAUDE.md — it loads every session	read →

Config quality

Tool	What it does	The story
claude-md-gen	generate a correct starter CLAUDE.md from your repo	read →
skill-lint	will your Claude Code skill ever get used? lint its description	read →
agentdrift	detect drift across a repo's agent-instruction files (CLAUDE.md / AGENTS.md / .cursorrules)	read →
tooldoc	MCP tool-description quality audit — find the undocumented parameters your agent guesses	read →

Supply chain & ecosystem measurement

Tool	What it does	The story
slopcheck	audit dependencies for AI-hallucination / slopsquat risk	read →
mcpfunnel	measure how much of the official MCP registry actually runs	read →
agentpr	do AI PRs dodge review? run the measurement on your own repo	read →

Author agent demos

Tool	What it does	The story
storyboard-doctor	pre-render linter for auto-demo storyboard JSON — catch dead-air/duration mismatches before you render	read →

Lint every channel (output gates)

A human developer is gated everywhere — review reads the code, a teammate notices you're stuck, someone catches the key in the diff. An unattended agent has none of that. These are the cheap, loud, deterministic gates that put it back — one per channel the agent touches. Each exits non-zero on a problem, so it drops into a Stop hook or CI unchanged. agent-frisk guards the input channel; these four guard the output channels.

Tool	What it does	The story
langcheck	lint agent output prose for foreign-script intrusion, mojibake, and runaway repetition	read →
leaklint	flag leaked secrets in agent output — redact-by-design, never reprints the value	read →
spinlint	flag wasted-loop / stuck sessions burning tokens with no progress	read →
claimlint	flag false completion claims ("tests pass", "created X") the transcript and filesystem contradict	read →

---

33 tools. Everything here is heuristic and honestly labelled — estimates, not bills; where a finding is uncertain or our own hypothesis lost, the write-up says so. Built and run by deemwar. Putting agents to work on code that matters? Talk to us.