Scan your agent
Scan your agent, free

For AI agents that touch money, customer data or productionFor AI agents that touch money and data

Your AI agent says it did the right thing. We check if it really did, and stop it if it didn't.

Scan your agent, free Watch it catch one

Free and read-only. Nothing leaves your machine.

Deedglass

The agent says

billing-agent 12:04

Refunded $20 for order 4821 and let the customer know.

What really happened

Stripe read just now
Action
Refund, order 4821
Amount
$200.00
Status
Sent
Doesn't match

Stopped. The agent is paused. Its next refund waits for your OK.

  1. Claimed
  2. Checked
  3. Proven

Works with the tools your agents already use

  • Stripe
  • Postgres
  • GitHub
  • Slack
  • Gmail
  • Notion
  • MCP
  • Claude Agent SDK
  • OpenAI Agents
  • LangChain
  • CrewAI
  • Cursor
  • Claude Code

How it works

One action, from ask to record. Follow it down the page.

  1. 01Ask

    Your agent asks first.

    Before it touches money, data or code, the agent sends the action to Deedglass. One line in its setup, no rewrite.

    billing-agent wants to

    Refund $120 for order 4822

    Tool
    stripe.refund
    Order
    4822
    Amount
    $120.00

    Asking Deedglass first

  2. 02Rules

    Your rules decide.

    Allow it, block it, or ask a person. Rules are plain lines you can read. Risky ones wait for a tap in Slack or email.

    Your rules

    • Refunds up to $50Allow
    • Refunds over $50Ask a person
    • Drop or empty a tableNever
    • Customer data in bulkNever

    Approval needed. Refund $120 for order 4822.

    ApproveDeny

  3. 03Run

    Then it runs.

    Only after a yes does the action reach Stripe, Postgres, GitHub or your own API. A no never reaches them.

    Approved by you, 12:05

    Stripe Refund, order 4822, $120.00 Sent

    Only now does the action reach Stripe. Nothing ran before the yes.

  4. 04Check

    We check what really happened.

    Deedglass reads the result from the system itself and compares it with what the agent said. Matches, Doesn’t match or Can’t check.

    The agent says

    $120 refunded

    Stripe shows

    $120.00 refund

    Matches Read from Stripe itself, not from the agent.

  5. 05Record

    It’s sealed in a record.

    Who asked, who approved, what ran and what we found, signed, so any later edit shows.

    Deed#1,284
    Action
    Refund, order 4822
    Asked by
    billing-agent
    Decision
    Approved by you
    Result
    Matches
    Level
    Proven

    Sealed 12:05:14

  6. 06Chain

    And linked to the last one.

    Every record points to the one before it, so a change anywhere before the newest one shows. Anyone on your team can check the history in a browser.

    1. #1,280
    2. #1,281
    3. #1,282
    4. #1,283
    5. #1,284

    Each record holds a fingerprint of the one before it. Change one byte anywhere and every check after it says Doesn't match.

Try it

Try to sneak one past.

You are the agent. Send any move. The rules are on the right, and Deedglass checks every result.

Your moves

The rules

  • Refunds up to $50Allow
  • Refunds over $50Ask a person
  • Drop or empty a tableNever
  • Customer data in bulkNever

What happened

  1. Pick a move on the left.

Sneaky moves that got past: 0 of 4

What you get

Seven parts. All free to run yourself.

  • Rules

    Plain lines that allow, block or ask a person. Start from a pack for Stripe, Postgres or GitHub.

  • Check

    Reads the result from the real system and compares it with what the agent said.

  • Records

    Every action sealed and linked to the last, so anyone can check the history later.

  • Approvals

    Risky actions wait for a person. Approve in Slack or email with one tap.

  • Stop

    One switch pauses every agent at once, across every server.

  • Undo

    Roll back what an agent changed, where the system allows it, and record that too.

  • Insights

    See which agents act, how often, what they touch and what gets blocked.

  • One scan to start

    See what your agents can touch today, graded A to F, before you change anything. It sends nothing.

    Scan your agent

Measured, not promised

Numbers we can show you.

  • 70 of 70

    attacks in our public suite end the way they should.

    See all of them

  • 23.5 ms

    the longest it took Stop to reach a second server, over 20 tries. Target under 2 s, met.

  • 1.53 s

    to check 100,000 records and every link between them. Target under 5 s, met.

    Check a record

  • 0 bytes

    that the free scan sends anywhere. It reads your machine and stays there.

    Run it

Timings measured on an Apple M1 laptop, plugged in. The attack suite last ran on October 13, 2026.

Pricing

Free to run. Pay when it earns its keep.

  • Free

    $0

    Rules, checks, records, approvals, Stop, Undo, Insights and the scan, on your own servers.

    Start free

  • Founding partner

    $99 once

    Pro for six months, a direct line to Arjun, and a say in what ships next.

    10 places in total

    Take a founding place

  • Pro

    $49 a month

    The Action on private repos, the full report, Pro policy packs and the fix writer.

    See Pro

Full scan report, $29 once (opening soon). AI Agent Safety Setup, $299 to $750: we set it up for you. Cloud beta in November.

Questions

Good questions.

Anything else: hello@deedglass.com. A person answers.

What does Deedglass do, in one line?

It sits between your AI agent and the tools it uses. It decides what may run, checks what really happened, and keeps a record you can check later.

Do I have to rewrite my agent?

No. MCP agents point at Deedglass instead of the tool. Python and TypeScript agents add one wrapper, with ready adapters for LangChain, OpenAI Agents, the Claude Agent SDK and CrewAI.

What does “Can’t check” mean?

We couldn’t read the result from the real system, so we say so instead of guessing. Connect that system and the same action gets checked.

What leaves my machine?

Nothing, by default. Deedglass runs on your own servers with your own Postgres. The free scan sends nothing. Pro checks its license at most once a day, and only when a Pro feature runs.

What does it cost?

Running it yourself is free. Pro is $49 a month for the extras. The founding place is $99 once for six months of Pro, and we can set it all up for you.

Who is behind it?

Arjun Prabhakaran, a solo founder in India. Founding partners get a direct line to him and a say in what ships next.

Know what your agents really did.

Start with the free scan. It reads your setup and sends nothing.

Scan your agent, free See pricing