Free preview workspace — no card, no redeploy · Open a workspace →
Agent action control plane Limited preview

Trust your agents. Verify every move they make.

Gavrun sits between your agents and the systems they touch. Every tool call gets a verdict in milliseconds — and the dangerous ones wait for a human.

No card · no redeploy · wraps one call

gavrun · approvals live
held for a human dec_b73f·1a09
stripe.issue_refund
Refund to #5521
$3,180.00
agentpayments-agent · sonnet-4.5 policyrefunds.over_1000 — needs approval asked2 min ago · run_4d7b91e2
Recent decisions
14:32:04db.query · orders by customerallow
14:31:58comms.send_email · unverified domainblock
14:30:12stripe.issue_refund · $72.40allow

Sits in front of the frameworks your agents already use

LangChain CrewAI LlamaIndex OpenAI Agents SDK Anthropic Claude Pydantic AI
The console

One workspace for the whole governed fleet.

Onboard agents, author policies, clear approvals, and read the activity trail — from a single developer control plane. Metering and risk on the front page, decision history one click away.

app.gavrun.dev/overview live
01 The decision

A guardrail checks the input. Gavrun owns the whole call.

One decision engine, three phases — preflight before the call runs, inflight while it runs, and postflight after it returns. A single verdict covers what actually happens, not just what was asked.

01Preflight
before the call runs

Decide on the way in

Match every call against policy before a byte executes. Allow it, block it, or pause it for a person.

stripe.refundhold
02Inflight
while it runs

Watch it live

Track token spend and streamed output in flight — and cut any call that drifts out of bounds.

4,210 tokens · streamingwatching
03Postflight
after it returns

Verify & record

Inspect the result, redact or roll back if needed, then write an entry that can't be edited.

response · 142mslogged
Z One engine — preflight · inflight · postflight tools · models · tokens · outputs · safety
02 Author policies

Write the rules once. Gavrun decides the rest.

Scope a policy to a tenant, a set of agents, and an environment — tool allow-lists, approval-gated actions, model and token ceilings, and built-in defaults for PII handling, prompt injection, and data externalization.

app.gavrun.dev/policies live
03 Capabilities

What it takes to let agents act safely.

Onboard agents & scope credentials

Register each agent and generate scoped SDK credentials in minutes. Every key is tied to an identity, so you always know which agent acted.

SDK keysidentity-bound

Define policies that fit your risk

Set rules per tool for models, approvals, token budgets, and safety actions. Allow, block, or require review — with conditions and thresholds.

toolsmodelstokenssafety

Route risky actions to humans

High-impact calls pause and wait for a teammate to approve or reject — before anything executes. Nothing irreversible happens without sign-off.

human-in-the-loop

Audit every decision by ID

Review approval history and activity by decision ID. Every allow, block, and approval is on the record — searchable and exportable.

decision IDsfull history

Trace rejected retries

When a rejected agent tries again, you see it. Follow the full retry chain on a single timeline to spot agents probing their limits.

retry chains

Email & Google sign-up

Get your team in fast. Sign up with email verification or Google federation, then pick up right where you left off across sessions.

email verifyGoogle SSO
04 Activity log

Every decision, kept on the record.

One searchable trail of registrations, decisions, guardrail hits, and policy changes — keyed by decision ID, metered by tokens, and flagged for risk. From the first call your agent ever makes.

app.gavrun.dev/activity live
05 For developers

Wrap one call. Ship your agent unchanged.

Register an agent, pull its scoped credentials, and configure once. Gavrun intercepts each tool call — deciding, recording, and routing for approval whenever your policy says so.

1

Install & configure

One install, then Gavrun.from_config() with a scoped agent_id and client secret — set once per agent.

2

Run your agent as-is

Every tool call is intercepted automatically — no changes to your agent loop, no redeployment.

3

Set policies in the dashboard

Configure rules per tool — thresholds, token budgets, and human-approval requirements. Changes take effect immediately.

agent.py python
from gavrun import Gavrun, GavrunDecisionDenied

gavrun = Gavrun.from_config(
    config_path="~/.gavrun/gavrun-governance.yaml",
    agent_id="agent_76f70164eb79",
)

# Wrap any tool or model call — Gavrun
# decides before it ever runs.
try:
    result = gavrun.guard_execution(
        action_type="tool",
        action_name="stripe.issue_refund",
        payload={"amount": 3180, "customer": "cus_5521"},
        callback=lambda: issue_refund(amount=3180),
    )
except GavrunDecisionDenied as e:
    logger.warning("Denied: %s", e.reasons)

# → allow · block · held for a human
06 High-impact actions

Made for the moves you can't undo.

Refunds & payments

Block transactions over a threshold. Route high-value refunds to a reviewer before money moves.

require approval

Data deletion

Guard against destructive deletes and schema changes. Sensitive records can't be removed without sign-off.

block by default

Infrastructure changes

Keep agents from touching production. Resource deletes and config changes pause for an operator.

operator review

Outbound email

Stop messages to external or unverified recipients. Enforce tone, pricing, and content rules before send.

policy-gated

Put a verdict in front of every action.

Open a preview workspace, point your agents at Gavrun, and watch the first decisions land in minutes. Sign in to pick up where you left off, or start fresh — no card, no redeploy.

Preview access · no card, no redeploy