Allowly Allowly
Start free

Shared permissions · prompts and agentsThe control plane for agent permissions.

Manage permissions in Allowly. Each agent connects through a line in its prompt. confirm and escalate go to Slack for a human decision.

Start free See how it works

No payment method required.

ALLOWLYsupport agentCLEARED · ALLOWtriage promptHOLD · CONFIRMfinance agentHOLD · ESCALATEresearch agentSTOP · DENYTHE TOWERyour policy in AllowlyEVERY AGENT ASKS FIRSTallow · deny · confirm · escalateHOLDS GO TO PEOPLEreviewers decide in Slack

Why a control plane

Stop copying rules into every prompt.

Copied rules drift apart. Keep the policy in Allowly and connect each prompt with the same line.

BEFORE · RULES COPIED INTO EVERY PROMPTWITH ALLOWLY · A SHARED POLICYsupport promptrefunds: ask a managertriage promptrefunds: allowedfinance agentrefunds: ask financeweekend sessionrules from MarchFOUR COPIES, FOUR ANSWERSevery prompt edited by hand; rules driftPOLICY · publishedsupport prompttriage promptfinance agentweekend sessionSHARED POLICY, SAME LINE IN EACH PROMPTrules live in Allowly; the line stays the samesame line in each

How it works

Write the rules. Add a line to each agent’s prompt. Approve in Slack.

  1. 01

    Write the rules

    Describe permissions in plain language. Review the drafted actions, decisions and reviewers, run examples, then publish.

  2. 02

    Add a line to each prompt

    Copy the instruction into each agent’s prompt. Credentials stay in the agent’s secret settings, never in the prompt.

  3. 03

    Handle approvals in Slack

    confirm goes to the agent owner. escalate goes to the reviewer the policy names. Work resumes only after a fresh allow.

The policy

Rules in plain language, shared by every prompt.

Write the policy once in Allowly and connect the prompts and agents that should follow it. Each connected prompt gets the same rules and the same four decisions.

How policies work →
Policy · support deskIllustrative example

Read support tickets. Ask me before sending a reply. Ask the finance reviewer before issuing a refund. Never delete a customer record.

Drafted actions

support.ticket.readallow
support.reply.sendagent ownerconfirm
billing.refund.issuefinance reviewerescalate
customer.record.deletedeny
Shared bysupport prompttriage promptweekend session
Agent prompt · support agentIllustrative
  1. You are the support agent for a software company.
  2. Answer customers clearly and keep replies short.
  3. Use the ticketing tools to look up orders.
Before preparing each new logical external action, use your private credential HTTP tool to GET /v1/agent-instructions with agent_name set to the URL-encoded configured ALLOWLY_AGENT_NAME at the trusted configured ALLOWLY_API_BASE_URL (normally https://api.allowly.ai); follow the complete returned instructions, map the request to their typed JSON, execute only allow and pause confirm or escalate for Slack review; stop on fetch failure or a redirect to another origin, never expose the key, and keep the exact instructions, policy reference and request body unchanged once a Check starts or review is pending.
the same line in every prompt

The prompt line

The same line in every agent’s prompt.

Hosted instructions keep this line stable. Before each new action, the agent fetches its assigned policy instructions, then prepares a typed JSON Check. Started checks and Slack reviews keep their original policy and request.

Both Hosted instructions and Copied instructions need configured HTTP tools, a registered agent name and a stored runtime key. The line does not install tools. Keep credentials in secret settings, outside prompts and URLs; failed instruction fetches stop the action.

A prompt line relies on the agent following it. For enforced tool boundaries, put the same check in the tool handler or MCP middleware.

Follow the setup guide → Open the check instructions →

Four decisions

Routine work moves. People decide the rest in Slack.

allow
The agent proceeds.The requested action is allowed under the bound policy and current limits.
deny
The agent stops.The action is refused. A failed or unavailable check also leaves the action stopped.
confirm
escalate
The agent pauses.The assigned reviewer approves or rejects in Slack. After approval, the agent checks the same request again and proceeds only on allow.
confirm · escalate go to Slack

Slack records the review decision. The connected agent remains responsible for waiting, checking again and carrying out an allowed action.

A clear boundary

Allowly decides. Agents act. People approve.

Allowly

Keeps the rules and returns allow, deny, confirm or escalate.

Your agents

Do the work, check before external actions and follow the decision.

Your reviewers

Decide confirm and escalate in Slack. The agent re-checks before it acts.

The fine print

A shared place to manage access.

Keep permission rules out of every task prompt. Manage them in Allowly and point connected prompts at the configured permission set.

Policy changes

Review and publish changes centrally. Publication may replace authorization IDs; update the shared connection with the returned IDs before agents resume.

Coverage

A prompt line instructs an agent to check. It cannot prevent a skipped check or another route to a tool. Add tool-level enforcement where your workflow requires it.

Decision records

Allowly records each completed check. Signed receipts let you verify the issuer and detect changes to the signed content; they do not prove execution or a complete activity history.

How the control plane works → Product questions → Data handling →
More ways to use Allowly

Observe reviews submitted activity after it is reported. Execute links a permission check to the outcome reported by your customer-run integration. SDKs, MCP and n8n support tool-level integrations.

Observe → Execute → Integrations → Signed records → Compare approaches → Approval workflow example →

Start with your first agent.

Write the rules, connect Slack and add the prompt line. Reuse the permissions across the prompts that need the same access.

Start free Read the setup guide

Free includes a single active agent and 1,000 lifetime decisions. No payment method required.

View plans and pricing →