Architecture & guarantees
How “nothing goes out until you approve it” is enforced
The product pages make one promise. This page shows where the code keeps it: which component decides whether an action may run, where a draft waits, what can move it, and how the credentials that finally send it are stored.
- A queued draft is sent only by an approval that claims it from pending
- Autonomy is resolved per workspace, action and network when the action runs
- An API key can make the dial stricter, never looser
- Connected-account tokens are sealed with AES-256-GCM, bound to their workspace
Overview
The path of a write
Every outbound action — a scheduled post, a reply, a DM, a comment under someone else’s post — has the same three possible fates, decided by one controller. The figure above is the whole model; the sections below take it apart.
Off off
The action is skipped. Discovery and drafting still run; nothing is sent, and the controller’s reason is kept on the item.
Approve approve
The draft is written to the approval queue as pending and waits there. This is where every action starts.
Auto auto
Sent without review — but only after guardrails and moderation pass. Any violation sends it to the queue instead.
Queue
A draft leaves the queue only by being claimed
Writes that need a person are stored as queue items with status pending. Approving one is a two-step operation, and the first step is what makes it safe to approve from two places at once — the dashboard, a teammate’s phone, a reviewer agent.
- Written as
- pending — by the scheduler, a growth run, an MCP write tool, or a v1 brief
- Approve
- Refused unless the item is pending. Then one conditional update moves it pending → approved; the caller that gets no row back sends nothing.
- Edit
- You may change the text while approving. What ships is your text, and that is what gets classified and logged.
- After the claim
- Plan gates run (publishing needs a paid plan; the Free reply allowance is reserved atomically and refunded if the send fails), then the connector sends.
- Ends as
- sent, with the platform’s id — or failed, with the reason. Rejected items never send.
-- approveAndSend(id) — before any gate or send
UPDATE approval_queue
SET status = 'approved', error = NULL
WHERE id = $1
AND status = 'pending'
RETURNING id;
-- no row back → "already being processed", nothing is sentWho can approve
Autonomy
The dial is resolved when the action runs, not when it was planned
The controller answers one question for a (workspace, action, network): off, approve or auto. The scheduler asks it for every post at the moment it would publish, and a growth run asks it for every comment — so turning the dial down stops what was already planned.
// lib/autonomy.ts — decide(workspace, action, platform)
platformDefault = PLATFORM_CAPS[platform][action] ?? 'approve'
setting = autonomy_settings(workspace, action) // may be absent
mode = setting ?? minMode(DEFAULT_MODES[action], platformDefault)
beyondPlatformDefault = setting is looser than platformDefault- Your setting
- Stored per workspace and action, and returned as you set it. The network’s policy is a default, not a ceiling.
- No setting
- The safer of our default and the network’s own policy — so an absent opinion never lands outside a network’s rules.
- Looser than policy
- Allowed, and flagged: the decision carries a beyond-platform-default flag and a reason naming the rule it crosses.
- Before auto
- Your banned topics, link and length rules, and a moderation pass. A violation — or a moderation check that could not run — queues the action for a person.
| Network | Post | Reply | DM | Comment on others |
|---|---|---|---|---|
| approve | approve | approve | off | |
| approvepolicy: auto | approvepolicy: auto | approve | offpolicy: approve | |
| approvepolicy: auto | approvepolicy: auto | approve | offpolicy: approve | |
| approvepolicy: auto | approve | approve | off | |
| X | approvepolicy: auto | approvepolicy: auto | approve | offpolicy: approve |
| Bluesky | approvepolicy: auto | approvepolicy: auto | approve | offpolicy: approve |
| YouTube | approve | approve | approve | offpolicy: approve |
| Stack Exchange | approve | approve | approve | offpolicy: approve |
| Lemmy | approvepolicy: auto | approvepolicy: auto | approve | offpolicy: approve |
| Discourse | approvepolicy: auto | approvepolicy: auto | approve | offpolicy: approve |
Nothing starts on auto on any network. Where a network’s own policy would allow more, it is shown as “policy”: that is the line a looser setting of yours is measured against.
Keys
A key can lower the dial, never raise it
An fr_live_ key is bound to one workspace when it is minted, carries only the scopes it was minted with, and is stored as a SHA-256 hash — the plaintext is shown once and never again. Only an owner or admin can mint one.
- read · default
- List and read: drafts, queue items, analytics, the registry. Also what keeps an approval webhook firing.
- draft · default
- Create drafts and queue items — every social write tool enqueues — run audits and research, reject a queue item. Default on a new key.
- publish
- Opt-in. Tools that act on your own published content or a browser task directly: cross-posting an already-published article, refreshing or editing a live article, running a goal on a web page. Never social posts, replies or DMs.
- admin
- Opt-in. Workspace configuration — syndication targets, relationships, and set_autonomy, which can still only lower a mode.
- approve
- Opt-in. Approving a queue item sends it, so it is never on by default.
- -32001
- The key lacks the scope this tool needs. The message names the scope and where to mint a key that has it.
- -32002
- No key may do this as asked: raising an autonomy mode, or approving without the approve scope. The error carries the dashboard link.
- HTTP 401
- Missing or revoked key. Revoking takes effect on the next request.
- Webhook
- A key can register one “something is waiting” webhook. It is signed with HMAC-SHA256 and carries no draft text — only what is waiting and where to read it.
{
"jsonrpc": "2.0",
"id": 7,
"error": {
"code": -32002,
"message": "set_autonomy: an API key can only lower autonomy (reply is 'approve'; 'auto' would raise it). Raise it in the dashboard: …/dashboard/autonomy",
"data": { "tool": "set_autonomy", "human_only": true, "dashboard_url": "…/dashboard/autonomy" }
}
}Credentials
Tokens are sealed, and sealed to their workspace
Your network passwords never reach FounderReply — you sign in to each network yourself, in a real browser session, and the browser service keeps that session. Any account token or secret FounderReply does store is encrypted at rest with AES-256-GCM before it is written. The associated data binds each ciphertext to its table, column and workspace, so a blob copied into another tenant’s row fails authentication instead of decrypting.
v2.<base64( 0x02 ‖ kid ‖ iv[12] ‖ ciphertext ‖ tag[16] )>
AES-256-GCM, Web Crypto (runs on the Workers runtime)
iv 96-bit, crypto.getRandomValues, fresh per encryption
aad 0x02 ‖ kid ‖ "brand_accounts:<column>:<workspace_id>"- Cipher
- AES-256-GCM through Web Crypto: 256-bit keys, a fresh 96-bit random IV per encryption, a 128-bit tag.
- Bound to
- The version byte, the key id and the table:column:workspace context, all authenticated.
- Rotation
- A key ring by id. New writes use the active key; older rows still decrypt under theirs until they are re-encrypted, and only then is the old key retired.
- API keys
- Never encrypted, because never stored: only the SHA-256 hash is kept.
Honesty
Where the queue is not in the path
A guarantee is only useful if its edges are stated. These are the five ways something goes out without a person approving that exact item — each one a choice someone in your workspace made.
You raised the dial
An action set to auto in the dashboard sends without review, after guardrails and moderation. Only a signed-in member can raise it.
You published it yourself
A signed-in member sending finished content to the publish route without asking for approval — the API twin of “Publish now”. No API key reaches that route.
A key with the approve scope
An owner or admin can mint a key that approves queue items — for a reviewer agent. The decision is recorded against the key.
Press pitches that earned it
After fifteen consecutive pitches approved without an edit, pitches send directly; about one in twenty still goes through the queue, and any edit or rejection turns it off. A key’s approvals never count toward it.
A key with the publish scope
Opt-in tools that act on your own blog and browser tasks directly — cross-posting an article that is already published, refreshing or editing a live article in place, driving a goal on a web page. None of them posts, replies or sends DMs on a social network; those always go through the queue.
Everything else — every MCP write tool, every v1 brief, every scheduled action on a gated setting — waits in the queue. See the MCP tool registry for what each tool does.
Build on it
Automate the work. Keep the approval.
Mint a key, point your agent at the MCP endpoint, and every write it makes waits in the same queue you already approve from.
Free to start · Approval on by default · No card required