COUNTERSEAL

Every irreversible action gets a second seal.

Your AI agent never holds a real API key, and anything it cannot take back waits for your passkey.

Website · Try it, no account · Quickstart · How it works · Limits · Security · Self-hosting · Full guide

Illustrated flow: the repository name is an example. The 428 response, the rule id and the approve link are what the server returns.

Try it without signing up

• In your browser. The demos on the home page run in the tab, with no account and no cookie: an agent's signed Stripe refund held with HTTP 428, approved and recorded (a button stands in for the passkey), the real shell map and the MCP wrapper's real policy. Type your own command into the shell map.
• In your terminal. guard check reads one command against the shell map on your machine and sends it nowhere (npx only fetches the package from npm):

$ npx counterseal guard check "terraform destroy -auto-approve"
HELD  sh.terraform-destroy
      Destroys every resource this configuration manages.
      terraform destroy -auto-approve

It exits 10 when a command would be held and 0 when it is not on the map. The hosted app, the CLI sign-in and real held calls need an account: a name and a passkey, no email.

Quickstart

Three commands (the CLI is the counterseal package, Node 20 or later, zero dependencies):

npx counterseal login                                    # prints a code; type it at /app#device, check where it started, approve with your passkey
npx counterseal add github                               # paste the key once; it goes into the vault
npx counterseal token --label "coding agent, laptop"     # a proxy token for the agent (with one key in the vault no id is needed)

Prefer no terminal? Create an account at /app (a name and a passkey) and press Try a held call: a pretend call you approve with your passkey (nothing real is called and nothing is made public). Then connect your agent.

---

Why: no approval, no key

In April 2026, an AI coding agent hunting for a staging credential found a Railway API token in an unrelated file and called volumeDelete. PocketOS had its production database and the backups on the same volume deleted in nine seconds. The token had been made to manage custom domains, and the prompt-level guardrails were on. (AI Incident Database)

A prompt is not a permission system. Counterseal moves the decision out of the model:

• The agent holds a scoped proxy token (lsh_...), never the real key. The real key stays in an encrypted vault and is added to the request on the server.
• A call that cannot be undone (a delete, a force push, a refund, a DROP TABLE) is held until a human approves that exact request with a passkey. No approval, no key.
• The decision is written down. With receipts on, each decision on a write is also an entry in a public log that anyone can check.

The agent's signed request is the seal. Your passkey approval is the counterseal. The public log keeps the impression.

Connect your agent

After the three commands in the quickstart, give the token to your agent as an MCP server (the leash_request tool calls the APIs through Counterseal):

{
  "mcpServers": {
    "counterseal": {
      "command": "npx",
      "args": ["-y", "counterseal", "mcp"],
      "env": { "LEASH_TOKENS": "github=lsh_..." }
    }
  }
}

Or point any SDK at https://counterseal.gautamkhosla.com/p/github with the token as its bearer. When the agent tries something irreversible it gets 428 held_for_approval and an approve link, and your phone or browser asks for a passkey.

All commands: login, logout, add, creds, tokens, token, revoke, freeze, holds, watch, agent-key, agent-keys, mcp, mcp wrap|pin|unpin|status, guard install|uninstall|status|check|map. Run npx counterseal help.

Features

	What it does
Approve	Irreversible calls are held until a passkey approves them, on your phone or laptop. An approval covers that exact request, once, within 10 minutes.
Record	With receipts on, every decision the broker makes on a write (allowed, held, approved, executed, denied) becomes an entry in a public, append-only Merkle log (RFC 6962, signed checkpoints). Allowed reads are not recorded. A receipt holds the method, the provider host, the decision and the rule, never a path, body, key, token, name, email or IP. Verify in the browser or offline with sdk/receipts.mjs.
Prove who	Bind a token to an agent key (Ed25519). It then only works on requests that key signed (RFC 9421), so a stolen token alone does nothing.
Guard	counterseal guard is a hook for a coding agent's shell commands. A command on the shell map (29 rules: terraform destroy, git push --force, DROP TABLE through psql, rm -rf outside the project, and more) waits for your passkey. A seatbelt, not a sandbox.
MCP gateway	counterseal mcp wrap sits in front of any local MCP server: destructive tool calls wait for your passkey, and the server's tool definitions are pinned so a changed tool is hidden until you review it.
Two-person seal	A team can require two different people, each with their own passkey, before an irreversible action goes ahead. Roles, one-time invites, and nobody can lower the rule quietly: lowering it takes 24 hours and tells everyone.
Deploy Gate (preview)	A GitHub Actions step that holds a deploy job until you approve that exact run. It uses the OIDC token GitHub signs for the job. Not yet tested with real GitHub tokens, only with a stand-in that signs like GitHub, so it is not part of the launch. The action is in this repository, not on the Marketplace.
Brakes	Money caps for Stripe refunds and payouts, a breaker for too many or repeated writes, and Freeze, which stops every agent token at once. They cover calls through Counterseal, not LLM tokens.
The app	An installable web app with live holds, alerts to your devices that carry no payload, the audit log, a privacy center, a data export and an evidence pack.

The irreversible map covers 10 APIs: GitHub, Cloudflare, Railway, Stripe, Supabase, Vercel, Neon, Fly.io, Resend and Shopify. Management APIs only; SQL sent straight over a database connection string never passes through Counterseal.

Details for each are in the full guide and the policy reference.

How it works

flowchart LR
    A["AI agent<br/>token + signature"] --> W
    subgraph W["Counterseal (a Cloudflare Worker)"]
      direction TB
      S["Agent signature check"] --> P["Policy + irreversible map"]
      P --> V[("Vault<br/>real keys, encrypted")]
      P --> H["Held for approval"]
      P --> AU[("Audit log<br/>hash-chained, private")]
      AU --> LOG[("Public log<br/>Merkle tree")]
    end
    W -- "real key, added server side" --> U["GitHub, Stripe, Railway, ..."]
    H -- "alert" --> You["You: passkey on phone or laptop"]
    You -- "approve or deny" --> H
    LOG --> Anyone["Anyone, offline: entry + proof + signed checkpoint"]

One request, end to end:

sequenceDiagram
    participant Agent
    participant C as Counterseal
    participant You as You (passkey)
    participant API as Upstream API
    Agent->>C: DELETE /p/github/repos/acme/app (token + signature)
    C-->>Agent: 428 held_for_approval + approve link
    You->>C: approve (WebAuthn)
    Agent->>C: the same request again
    C->>API: DELETE /repos/acme/app (real key)
    API-->>C: 204
    C-->>Agent: 204

Honest limits

• It is a broker, not a sandbox. It protects calls that go through it. An agent that still has the real key in a .env file can bypass it, so put the real key in the vault and give the agent a Counterseal token.
• The guard and the MCP wrapper are seatbelts. They run on the agent's machine as the same user. An agent that can edit its own hook settings or run a script around the hook is not stopped. They match lists; they do not understand intent. For a hard stop, give the agent a token instead of the real key.
• The map is a reviewed list, not every dangerous action. Everything not on it is allowed on those hosts.
• The public log has one signing key. Split views are detectable, not prevented. Cosigning by independent outside parties is on the roadmap.
• Receipts prove what was decided, that an agent key signed, and that a passkey approved. Publicly they do not say which person approved, and the time of each receipt is public.
• Two people who agree to act together are out of scope for the two-person seal. Break glass, by design, lets one person loosen protection for up to an hour.
• The Deploy Gate has not run against real GitHub tokens yet; the tests use a stand-in that signs like GitHub.
• Brakes cover brokered calls only, not LLM spend and not SQL over a database connection.
• The hosted beta has a daily request budget. It is sized for Cloudflare's Workers Free plan, which counts every request (pages and assets too) against 100,000 a day for the whole account. If that runs out, calls through the hosted service fail until 00:00 UTC (threat model). A self-hosted copy has its own budget.
• Not independently audited. The security standards page is a self-assessment. Beta limits are in docs/DEPLOY.md.

Security model

SECURITY.md (guarantees, controls, how to report a vulnerability) and the threat model. The wire formats are in docs/PROTOCOL.md.

Self-hosting

Counterseal is one Cloudflare Worker, one D1 database and one Durable Object, and it fits the Workers Free plan. Follow docs/DEPLOY.md (secrets, migrations, wrangler deploy, settings). To work on it locally:

cd server
node --test test/*.test.js ../sdk/*.test.mjs ../cli/test/*.test.mjs   # the whole suite
node dev/local.js 8790                                                # a local server on http://localhost:8790

Repository layout

server/        the Worker: API, proxy, vault, passkeys, policy, audit, the public log, approvals, teams, gates
cli/           the counterseal CLI, MCP server, guard and MCP wrapper (zero dependencies)
sdk/           receipts (record and verify offline), Web Bot Auth sign and verify, the shell map
action/        the Deploy Gate GitHub Action and sample workflows
plugins/       the coding-agent plugin (hook, skill, pinned scripts)
packages/      leashcli, the old package name, a thin alias for counterseal
site/public/   the website and the installable app
docs/          guide, threat model, protocol, policy reference, deployment, release checklist

Contributing

Issues and pull requests are welcome. Read CONTRIBUTING.md and the Code of Conduct first. Report a vulnerability privately, as SECURITY.md describes, never in a public issue.

Status

The CLI, counterseal 0.4.0, is on npm. The hosted service lives at counterseal.gautamkhosla.com and is free during the public beta; there is no billing and no paid plan. What each release adds is in CHANGELOG.md, and what is deployed where is in docs/DEPLOY.md. Self-hosting runs everything in this repository.

Next: AWS (SigV4 re-signing), PlanetScale, SSO for teams, outside cosigning of the log checkpoints, SCITT receipts.

License

Apache-2.0.