The lockfile for agent tools

Know what your agents loaded. On every machine.

Every developer on your team now runs coding agents that load skills, plugins, MCP servers, hooks and instruction files nobody re-checks. kova pins all of it in a lockfile, verifies it at every session start across Claude Code, Codex, Gemini CLI, Cursor and the rest, and gives the people responsible for the fleet one answer to the question they cannot answer today: what changed, what did the agents install, and why.

Open-source CLI, MITRolls out like a lockfile, not like an agentA tenth of a second at session start
One lockfile across the harnesses your team already runs
Claude CodeCodex CLIGemini CLICursorGitHub CopilotClaude DesktopWindsurfCline / Roo
Rollout

Start where nobody feels it.

No security control was ever adopted because individual developers chose it. Lockfiles, branch protection, secret scanning: one person turned them on for the org, and they were quiet enough that nobody fought to turn them off. kova is built to spread the same way.

1 · PULL REQUESTS

A check that blocks the Deadbugz class where it arrives

Commit a kova.lock once. From then on a PR that adds or changes an agent instruction or config file (CLAUDE.md, nested rules, .mcp.json, hooks) without updating the lockfile fails the check. No account, no hooks, nothing on laptops.

two lines
# .github/workflows/kova.yml
on: [pull_request]
jobs:
  kova:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: ian-hickey/kova@v0.1.18
2 · LAPTOPS, BY POLICY

Session hooks that developers tolerate

Once the check has earned trust, the CLI goes on machines the way the rest of the toolchain did: it was there. Day one produces nothing a human must act on. Verification costs about a tenth of a second. Edits an agent makes through its own tools pin themselves; only changes nobody watched are reported.

$ kova init
Created ~/.kova/kova.lock · 47 artifacts
$ kova login <team key>
3 · THE FLEET VIEW

What security and platform teams actually buy

Every machine, every harness, when it last verified, what drifted, what agents installed and the prompt that caused it, which connectors are attached to which accounts. Approve or block a hash once for everyone. Alerts to Slack within seconds.

drift events · acme-platform
  critical  mcp_tools_changed    mbp-ian
  high      connector_added      mbp-priya
  info      instruction_file_edited  47 today
Why runtime

Every documented 2026 agent supply-chain incident passed static review.

Scanning a tool at install time answers one question: was it clean when someone looked? None of these were caught that way. All of them change what the agent sees after you stop looking.

IncidentWhat happenedWhat kova sees
Deadbugz
Aug 2026
MCP server behaved normally for three tool calls, then returned credential-hunting instructions from tools/list. Delivered via 23 pull requests.mcp_tools_changed from the proxy, critical. Enforce mode blocks the drifted list.
postmark-mcp
2026
Fifteen clean releases, then v1.0.16 added a hidden BCC to every email the tool sent.Tools hash drift on the next session; description scan flags hidden_bcc.
drp-compliance-sdk
2026
Payload fired during the MCP handshake, before any prompt was sent.Launch config and tool pin at init; new network destination during the handshake.
ClawHavoc
2026
Malicious skills rewrote MEMORY.md. The directives persisted after the skills were deleted.Memory files hashed every session; added text scanned for persistence and hidden-from-user instructions.
How it works

Pin. Verify. Enforce.

Three commands, then it runs itself through the hooks your harness already supports.

PIN

One command inventories everything the agent will load

Skills, plugins, MCP servers (connected once, tool list hashed), hook configs, account connectors, and every instruction file from / down through node_modules. Each gets a content hash in a lockfile.

$ kova init
Created ~/.kova/kova.lock · 47 artifacts
VERIFY

Every session starts with a check the model can read

A session-start hook diffs the live state against the lockfile in about a tenth of a second. Drift goes straight into the agent's context, with the diff of what changed, so the model is told its tools changed before it trusts them.

kova: DRIFT DETECTED — 1 change
  [high] instruction_file_added
  ./node_modules/left-pad/CLAUDE.md
ENFORCE

Changed tools are blocked, not just reported

The MCP proxy replaces a drifted tools/list with an error. The guard denies reads near an unreviewed instruction file and prints the one command that resolves it. A human accepts intended changes; the dashboard shows who did.

$ kova wrap --all --enforce
$ kova approve mcp:markdown-helper
Runtime enforcement

The two places an attack actually lands.

  • tools/list, every call. The proxy relays JSON-RPC byte for byte and re-hashes each tool list against the pin. A server that adds a tool, rewrites a description or changes a schema after install trips a critical event, and in enforce mode never reaches the model.
  • Instruction files, wherever they appear. Harnesses load CLAUDE.md, AGENTS.md and friends from every ancestor directory and from any folder the agent reads a file in. The guard denies the read that would pull in one you have not reviewed.
  • Hidden text. Every instruction file is scanned for Unicode Tag characters (invisible to you, legible to the model), bidi overrides, and a short list of instruction patterns from real incidents.
  • Who installed what, and why. Every install your agent runs is logged with the session and the prompt that caused it. Every edit an agent makes to an instruction file is logged with the diff.
Agent harnessClaude Code · Cursor · …kova proxyhash every tools/listMCP servernpx -y some-serverRead / Edit / Bashtool callkova guardunknown CLAUDE.md nearby?Project filesincl. node_modules/*~/.kova/kova.lock · team policy
Quiet by design

The standard is zero findings a human must act on, on a messy real machine.

Noise is how security tools get uninstalled. kova treats every false positive as a bug, and the rules below came from real machines, not a whiteboard.

Agent edits are not drift

An agent has to edit CLAUDE.md and its memory notes to work. A post-tool hook records what the agent wrote and checks the file holds exactly the edit it declared; such edits pin themselves and land in the dashboard as an info event with the diff, for an asynchronous look rather than a blocked session.

Only unwatched changes are reported

A CLAUDE.md that appeared in node_modules after an install, hooks that changed between sessions, a rules file rewritten by a git pull: nobody saw it happen, so it waits for a human. That is the supply chain, and it is the whole point.

A second opinion, calibrated

Keyword rules cannot tell curl | sh in a README from the real thing. Optional AI triage scores the diff of a change with a classifier that returns probabilities instead of prose: confident malicious goes high, confident benign clears the keyword hit. Measured: recall 0.70 at a 4% false-positive rate, against 0.39 at 19% for rules alone.

Known content stays quiet

Files byte-identical to what a popular package ships, identical copies in git worktrees, content you approved under another path, plugin lock files and registry timestamps: never reported.

Denials carry their own exit

A guard denial names the file and prints the exact kova approve '<key>' that resolves it, so a developer deep in a bug has a one-command path forward and no reason to rip the tool out.

Events close themselves

Approving in the CLI, or a clean verify, closes the matching events on the dashboard. The default view is "needs review": open, medium and above. Everything else is an activity log you can consult, not a queue.

Coverage

One lockfile, every harness.

Teams rarely run a single agent, and the harness vendors will each secure their own corner. The durable part is the lockfile that spans them and the fleet view on top. kova reads each harness's own configuration and hooks into each one's own extension points.

Claude CodeCodex CLIGemini CLICursorGitHub CopilotClaude DesktopWindsurfCline / Roo
Skills (every file in the tree)✓✓✓✓————
Plugins / extensions✓—✓——✓——
MCP launch config✓✓✓✓✓✓✓✓
MCP tool definitions (tools/list + instructions)✓✓✓✓✓✓✓✓
Hooks config✓✓✓✓✓—✓—
Instruction files, incl. ancestors and dependency folders✓✓✓✓✓—✓✓
Session-start verification hook✓—✓✓————
Guard: deny reads near unreviewed instruction files✓—✓✓————
Install ledger and observed-edit hooks✓—✓✓————
MCP proxy wrapping✓—✓✓✓✓✓✓

Codex CLI has no hook system yet; verification runs from your shell or CI there. GitHub Copilot, Claude Desktop, Windsurf and Cline / Roo Code are pinned and verified the same way, but kova does not install into their hook systems yet: verification runs from your shell, CI or the MCP proxy.

The fleet view

For the person who worries about every agent-enabled laptop.

The CLI is free forever. Connect it to a team and every machine's drift, installs, connectors and tool pins land in one place, with a policy the whole fleet enforces.

Fleet view

Every machine, which harnesses it runs, when it last verified, and whether it is clean, partial or drifting.

Install ledger

Which agent session ran npx skills add, on which machine, and the prompt that led to it. The question a security lead cannot answer today.

Drift feed with diffs

What changed in a CLAUDE.md, which hook commands were added, which tool descriptions moved. Not two hashes.

Shared policy lockfile

Approve or block a hash once. Every machine treats blocked as a critical failure and approved as clean on its next sync.

Account connectors

Remote tools attached to a developer's claude.ai account reach every seat and live nowhere on disk. kova inventories them and alerts when one is added or re-pointed.

AI triage, org-wide

One toggle. Verdicts are cached per content hash for the whole org; only the changed text leaves the machine, with credential-shaped strings masked first.

Alerts

Critical and high events go to Slack, a webhook or email within seconds of the hook firing.

Audit artifact

A signed report of your team's agent tooling you can hand to a security reviewer or a customer.

drift events · acme-platformlive
criticalmcp_tools_changedclaude-code:mcp_server:markdown-helpermbp-ian · 2 min ago
highconnector_addedclaude-code:connector:Google Drivembp-priya · 11 min ago
highhooks_changedcursor:hooks:~/.cursor/hooks.jsonmbp-sam · 1 h ago
mediuminstruction_file_added./packages/api/CLAUDE.mdci-runner-04 · 3 h ago
infoinstruction_file_editedclaude-code:memory:~/work/acme/CLAUDE.mdmbp-ian · 4 h ago
Security and data

Built to be reviewed by the people who review things.

What leaves the machine

Artifact names, paths, hashes, tool names and event metadata. For teams, the diff of a changed instruction file and the shell command behind an install event, with credential-shaped strings masked first. Never your prompts, never tool-call contents. The free tier sends nothing at all.

Open source where it matters

The CLI that runs on developer machines is MIT licensed. Read the proxy, the guard and the scanner before you install them.

No new daemon, no kernel module

One static binary invoked by the harness's own hooks. Remove it with kova hooks uninstall and your configs are restored from backups kova kept.

Self-host or hosted

The team API and dashboard run from a single container image against Postgres. Run it in your VPC or use ours.

Honest about limits

Enforcement on a developer's own machine is cooperative: an agent running as the user could, in principle, tamper with kova's state. kova denies and records the attempt; the dashboard is the copy it cannot reach. The network check around tool calls is a heuristic; the tool-hash check is exact.

Billing on our domain

No marketplace, no revenue share, no third party between your card and your invoice.

Pricing

Free for every developer and every repo. Paid when a team needs one view.

Free
$0
  • Full CLI, every harness
  • CI check, lockfile, hooks, proxy, guard
  • Local events and ledger
  • MIT licensed
Install from GitHub
Business
$99 / month
  • Up to 50 seats, 150 machines
  • Everything in Team
  • Enforce mode across the fleet
  • Priority support
Start free, upgrade in-app
Audit one-time
$499
  • We run kova across your team's harness configs
  • Signed report of every artifact, pin and flag
  • Remediation list
  • Hand it to your security reviewer
Request an audit

Enterprise: self-hosting, SSO, invoicing and a security questionnaire are available. Write to sales@kovaverified.io.

Questions

Things engineers ask first.

Will developers tolerate it?
That is the design goal, measured. The session-start check runs in about a tenth of a second (findings and hashes are cached by content; unchanged plugin trees are never re-read). Day one on a real machine produces nothing a human must act on. An agent's own edits pin themselves. KOVA_TRACE=1 shows where any time went.
Will it flag my own CLAUDE.md edits?
Not when an agent makes them through its file tools: those are recorded, checked against the edit the tool declared, pinned, and logged with the diff. Edits made by something kova did not see (a script, a pull, another tool) are reported and wait for kova approve.
What happens when a tool legitimately updates?
You get a drift event naming the server and the tools that changed, with the new descriptions. If it was intended, kova approve mcp:<name> re-pins it. Teams can approve once for everyone.
Where do we start?
The pull-request check. Two lines in a workflow, a committed kova.lock, no account. Laptops follow by org policy once the check has earned trust; the dashboard is where the fleet becomes visible.
Can my agent install kova itself?
Yes. There is a skill on skills.sh and a Claude Code plugin, kept in step with the binary on every upgrade. The skill teaches the agent to run kova verify before trusting a new tool and never to approve drift on its own.
How do I remove it?
kova unwrap --all && kova hooks uninstall. Your harness configs are restored; ~/.kova is yours to delete.

Add the check to one repo today.

Two lines in a workflow for the repo. One line on a laptop. One click to see the fleet.

$curl -fsSL https://kovaverified.io/install.sh | shCreate a team