Skip to content
GEOstack

The agent control plane

Stop your AI agents before they do real damage.

Arc is the guardrail between your agents and production. Every action passes an allow / ask / block policy, waits for a human when it’s risky, and runs only as a signed (ES256) request your app verifies — with a cumulative spend cap so a runaway loop can’t quietly burn your budget.

Start free Read the docs

Free to start — no credit card. Add the @geostack/arc SDK when you’re ready to integrate.

arc · decision stream live
12:04:01 read_account ( id=usr_8841 ) matched policy · read-only allow
12:04:02 issue_refund ( $240.00 ) held for a human — over auto-approve limit ask
12:04:09 increase_ad_budget ( $500,000,000 ) cap exceeded — agent budget is $2,000 / day block
12:04:11 delete_production_db ( name=orders ) default-deny — action not in policy block
$
Approval required ask
action
issue_refund
amount
$240.00
agent
support_ops_agent

Why now

Reportedly ~$500,000,000 of Claude usage in one month — because no one set a cap.

Figure reported by Axios from a consultant’s account. The company is unnamed and unconfirmed.

A consultant told Axios an enterprise client ran up roughly half a billion dollars in Claude token spend in a single month after handing out employee licenses with no usage limits — thousands of people running agentic workflows at once. The figure is unconfirmed, but the pattern is on the record: Microsoft cancelled most internal Claude Code licenses ($500–$2,000/eng/mo); Uber burned through its 2026 AI budget by April.

The cap didn’t exist, or wasn’t turned on. Arc is the cap — plus the approval and the kill-switch for the action you can’t undo.

agent budget · today $2,000 / $2,000
cap reached next action → blocked

How it works

Four checks between the agent and the thing it can’t take back.

  1. 01

    Policy

    allow

    Each action is allow, ask, or block. Anything you didn't declare is blocked. Default-deny, not default-trust.

  2. 02

    Approval

    ask

    Risky actions pause for a human. So does any spend over the cap — the loop waits instead of running.

  3. 03

    Signed execution

    signed

    Approved work ships as a short-lived ES256 JWS your app verifies before any side effect. The agent never holds a broad API key.

  4. 04

    Audit

    audit

    Every attempt, approval, block, and budget hit lands in a redacted, hash-chained log you can replay.

guard.ts @geostack/arc 
arc.defineActions({
  issue_refund: {
    risk: "high",
    defaultDecision: "ask",   // pause for a human
    cost: { mode: "field", field: "amount" },  // charge against a budget
  },
  delete_production_db: { defaultDecision: "block" },
})

// the agent calls this; Arc decides, then signs.
const runtime = arc.createArcAgentRuntime({ agentToken })
const res = await runtime.invoke(appId, "issue_refund", { amount: 240_00 })
if (res.status === "pending_approval") notifyApprover(res.approval_id)   // $240 → held

TypeScript-first, fail-closed by default. Sync policy with the arc CLI or the MCP adapter.

What you get

A receipt for every action — even the ones that never ran.

When something runs, it’s a signed request you can verify. When something is blocked, that’s logged too. The chain is tamper-evident: each entry carries the hash of the one before it, so a deleted row breaks the chain.

Audit event signed
timestamp
2026-06-04T12:04:02Z
action
issue_refund
payload
••••
approver
user_421
prev 9b4c…2f11 hash d821…84aa
ES256
signed execution, with nonce replay protection
default-deny
unknown actions are blocked, not waved through
hash-chained
redacted, tamper-evident audit trail
spend caps
cumulative budget guardrails, enforced per agent

Works with any agent or MCP server

  • OpenAI
  • Anthropic
  • Gemini
  • Cursor
  • Windsurf

Provider-agnostic. Arc wraps the action, not the model — these mark compatibility, not partnerships.

Put one risky agent action behind Arc today.

Wrap your most dangerous action this afternoon. Sign up for a free hosted workspace, integrate the SDK, and guard your first action — free, no credit card.

Start free → Talk to us

Next on the roadmap: Arc Auth — scoped, revocable identity for agents.