More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Coworker Charter: Permissions, Approvals, Audit for Agents

    1

    Write a charter for an AI coworker, get an allow, ask or deny verdict before every risky action, and keep a tamper-evident audit log. Policy and evidence for agents that send, spend, delete or publish. Not a sandbox.

    $9.99

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Coworker Charter: Permissions, Approvals, Audit for Agents

    Coworker Charter: Permissions, Approvals, Audit for Agents

    Example session with this skill installed

    My support agent wants to do four things. Check them against its charter, ask me for anything that needs me, and log it all.

    • Read your context and instructions
    • Compiled the coworker charter

    I ran each action through the charter (example-support-agent.json). One was allowed, one needed Sam, and two are refused.

    ActionDecisionRuleWhy
    Refund order 1182, 8 USDAllowA4Refunds of 10 USD or less are routine goodwill refunds the coworker may issue.
    Reply on ticket 4412 to customer@example.orgAskA3A customer email cannot be recalled, so a person reads the final text first.
    Refund order 1190, 140 USDDenybudget.per_action_max140 USD would take spend over the per-action cap of 100.
    Delete all tickets from 2024DenyN2Tickets are deleted one at a time, never in bulk.

    I issued the 8 USD refund. For the reply I showed Sam this message, and sent it after Sam answered "yes, send it" (09:09). I did

    not do the other two.

    Approval needed before I act.
    What will happen: Reply on ticket 4412: the replacement ships Monday and the refund of 8 USD is on its way
    Tool: helpdesk.send_reply (send tier)
    Recipient: customer@example.org
    Why I am asking: A customer email cannot be recalled, so a person reads the final text first. (rule A3)
    Who can approve: Sam Rivera (sam.rivera@example.com)
    Reply "approve" to go ahead or "decline" to stop. I will not do anything until you answer.
    

    The two denials stand. Only Sam Rivera can change what is allowed, by editing the charter, and I will not try another way round either of them. I told Sam about both and logged it as an escalation.

    All of it is in the audit log (five entries: four decisions and the escalation). verify says:

    OK: 5 entries, chain intact.
    head: 5:cd97a53c09f65171da1306b31a610c91424be4ee8c92953778a4d2e79612e559
    

    If someone deletes line 3 of that log, verify names the line and the reason:

    BROKEN: 1 problem(s). First break at line 3.
      line 3: seq is 4 but 3 was expected (entries are missing before this line); prev hash does not match the line above (entries missing, replaced or reordered)
    head: 5:cd97a53c09f65171da1306b31a610c91424be4ee8c92953778a4d2e79612e559
    

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Businesses want an AI agent that can use tools like a coworker, but with a person approving anything risky, clear limits on its tools, a way to take over, and a record of what it did. This skill makes your agent, and you, write that down and stick to it.

    Your agent interviews you about the job, then drafts a charter: the role, an owner, every tool sorted into a risk tier (read, write, send, spend, delete, publish), approval rules, a list of what must never happen, data boundaries, spend caps and an escalation contact. A linter then refuses the usual mistakes: wildcard tool grants, tools with no tier, risky tiers with no approval rule, missing or non-numeric spend caps, no owner, secrets in the text and vague rules. Before the charter goes live the agent runs five sample actions through it and you check the answers are what you meant.

    How it works. Three bundled Python scripts (standard library only) do the deciding and the record keeping. Before each risky action the agent runs check, which returns allow, ask or deny, the rule that decided it, and a plain-English message to show you. The same inputs give the same verdict every time, so the agent never decides for itself. Every decision goes into a hash-chained log. verify names the line that broke if one is edited, deleted or reordered, and report summarises decisions, asks, denials and spend by tool and by day. The skill also gives the agent a take-over and escalation procedure.

    Where it stops. This is a policy-and-evidence layer. It works only while the agent calls it. It cannot stop an agent that ignores it, it sees only what the agent says it will do, and it is not a security sandbox. The log is tamper evident, not tamper proof.

    • Policy and evidence only. It cannot stop an agent that does not call check, and it is not a sandbox. Give the agent least-privilege credentials as well.
    • check sees the action text the agent writes. If the agent describes an action falsely, the charter cannot know.
    • Phrase matching is whole words, case-insensitive, with look-alike letters from other alphabets not unified. A phrase list is a net with holes.
    • The hash chain is tamper evidence, not tamper proof. Anyone who can write the log can rebuild it, and entries cut off the end are noticed only if you kept the head hash somewhere the agent cannot write.
    • approved_by is text the agent records. The log cannot prove a person said yes; it keeps their quoted words so you can check.
    • The log lock works on POSIX systems on a local disk. On Windows or a network drive, run one writer at a time.
    • The scripts never delete anything. Retention is a policy field for you to apply, not something they enforce.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed today

    Requires

    Python 3.10+
    Python

    Permissions

    File Scopes

    Reads
    only
    the
    charter
    and
    log
    files
    you
    name
    Appends
    to
    one
    audit
    file
    writes
    nothing
    else

    Runs three Python scripts from temporary files. No network access, no environment variables. The scripts never delete anything and never change a charter.

    Listedtoday

    What's inside

    Frequently Asked Questions

    Only if the agent routes the action through it. It is policy and evidence, not a sandbox. If the agent ignores it, nothing stops that, but the gaps show in the log.

    No. The scripts read the files you name and append to one log you name.

    The script, from the charter, the same way every time. The agent is told never to override it.

    The skill tells it never to. The log flags a self-approval, an approval from someone not named in the charter, and an action that ran after a deny.

    `check` gives no verdict and the agent must treat that as "ask a person" for anything above a read.

    Any agent that supports SKILL.md and can run a Python script.

    New in Security