Skip to content
in/guard/out
Sign up
in/guard/out

Guardrails for AI apps and agents

Everything your AI sends and receives, checked first.

in/guard/out sits between your app and its model. Private data is masked before it reaches the model, wrong numbers are caught before they reach a person, and risky agent actions are stopped before they run.

Setup is two lines of code.

Sign upSee how it works

Every request & response is checked on the way through

What goes wrong when AI does real work

Newer models cost more, and even the best of them still make these mistakes – cheaper ones more often.

  1. 01

    Customer data goes to a third party

    Every prompt that mentions a customer sends their name, email or account number to the model provider.

  2. 02

    Answers quietly change the numbers

    A figure drifts from the source, or a total doesn’t add up – and it reads just as confidently as the right one.

  3. 03

    Agents follow instructions they shouldn’t

    A line hidden in an email or a web page can steer an agent into sending data out, or paying the wrong account.

  4. 04

    Nobody sees what the agent did

    Dozens of steps, tool calls and model calls per task – with no record of what ran, what it touched, or what it cost.

How a request becomes a safe answer

in/guard/out checks traffic in both directions: what your app sends to the model, and what comes back.

  1. 1Your app

    Your AI agent sends a request

    Same SDK, same prompts – it just calls in/guard/out instead of the model.

  2. 2in/guard/out

    Checked on the way in

    Private data becomes placeholders, and the request is scanned before a token is spent.

    mask private datainjectionbudget
  3. 3Your model

    LLM model answers

    OpenAI, Claude, OpenRouter or a local model – it receives the masked request, not the real data.

  4. 4in/guard/out

    Checked on the way out

    Figures are checked against the source and every tool call against your rules – then the real data is put back.

    fact-checktool rulesformat & safetyrestore data
  5. 5Your app

    Delivered and recorded

    Your app gets the answer – fixed, or blocked with the reason. Every step is on record, with its cost.

How it works, in detail →

What it does

Protect data

Private data stays private

Names, emails, cards and IDs are masked before anything reaches the model provider, and restored on the way back. They can’t slip out through a tool call either.

Data protection →
Verify answers

Wrong numbers get caught

Every answer is checked against its source: the maths is recomputed, figures are matched, unsupported claims are flagged. Safe fixes are applied automatically.

Fact-checking →
Control actions

Risky actions don’t run

An agent’s tool call has to pass your rules first: which tools, how risky, within what budget. A payment to an account the agent never read about doesn’t go out.

Agent security →

Try it: see what gets caught.

Pick a case – a prompt carrying private data, an answer with wrong numbers, or an agent action – and see every check’s verdict.

In Block mode anything that fails is stopped. In Fix & log mode what has one safe fix is repaired, and the rest is recorded and let through.

Try it · pick a case, see every check
Pick a case
Check resultsBlocked
Source: July $2.4M · August $1.8M · prior quarter $4.6M

July–August revenue was $4.5M, with August alone at $1.9M – up 14% on the prior quarter.

  • blockgrounding · arithmetic2.4 + 1.8 = 4.2, not 4.5
  • blockgrounding · figuresAugust is $1.8M in the source
  • blockgrounding · direction“up 14%”, but $4.6M → $4.2M is a decrease – no single safe rewrite

Block stops anything that fails a check and returns the reason, so the caller can’t miss it. Private data is masked in both modes.

See what your agent actually did.

Every step is recorded: the model calls, the tool calls and how risky each one was, what was stopped, and what the run cost.

Agents resend their whole history on every step, so in/guard/out can rebuild the full run as it happens – and enforce limits on the run as a whole, like budgets and loops.

How agent security works →
Runs · reconstructed run graph
1 blocked
SESSION run-4f9c21STEPS 4TOOL CALLS 3COST $0.0143COVERAGE 1/1
  1. #1openai/gpt-4ook
    readsearch_orders{"order_id":"4471"}↻2
  2. #2openai/gpt-4ook
    injection
  3. #3openai/gpt-4ook
    networksend_email{"to":"<EMAIL_ADDRESS_1>"}
    taint
  4. #4checkpointblocked
    paymentpay_via_bank{"iban":"DE89…"}
    ungrounded
What it caught
  • injection
    poisoned tool result, step 2
  • taint
    screened email leaving via send_email
  • ungrounded
    IBAN absent from the run’s sources
  • ↻ repeat
    same call fingerprint as step 1

Live in three steps

1

Change two lines

Point your SDK’s base URL at in/guard/out and use your in/guard/out key.

client = OpenAI(
−   base_url="https://openrouter.ai/api/v1",
−   api_key="sk-or-…",
+   base_url="https://api.inguardout.com/v1",
+   api_key="gr-…",
)
2

Watch first

Checks start in Fix & log: safe fixes are applied, everything else is recorded, nothing is blocked.

3

Block when ready

Switch the checks that matter to Block – per key, per user, or per request.

Works withOpenAIClaudeOpenRouterLangChainn8nMCP serversLocal models

See it on your own traffic

Put in/guard/out between your AI and its model.

Start in Fix & log, see what gets caught, then decide what to block.

Sign up