io.github.pendpost/pendpost

Official

by pendpost · JavaScript

Local-first social MCP: an agent drafts and schedules posts; you approve before anything publishes.

Local-first social posting where agents draft and schedule, but you approve before publishing.

Topics

ai-agentsautomationclaudecontent-creationlocal-firstmcpmcp-servermodel-context-protocolschedulingself-hostedsocial-mediasocial-media-automation

README.md13 KBView on GitHub

pendpost

Agent-operated social media with a human approval gate.

pendpost is a free, open-source (MIT), local-first social media planner where an AI agent drafts and schedules posts across Instagram, Facebook, LinkedIn, YouTube, and X behind a human approval gate you control. It is MCP-native: AI agents draft, lint, schedule, and queue your posts, but nothing goes live until a human approves it. It is built for developers, agencies, and technical solopreneurs who want agents to do the work without handing them the keys, and without getting accounts flagged.

Why pendpost is different (not just a scheduler)

Most "AI social" tools are schedulers with an agent bolted on. pendpost is the opposite. It is an operations layer designed around the agent-plus-human workflow, and these are the parts a scheduler does not give you:

Human approval gate. Every post carries an approval state (draft, approved, rejected) and is fail-closed: a post with no approval will not publish. plan_create_post always creates a draft, and only approve_post or reject_post can flip it. There is no self-approval, so the actor who created a post can never approve it (an agent cannot bless its own draft); the owner is exempt.
Anti-ban circuit breakers. A Meta error 368 (an action block) trips a breaker that halts the Meta lane and never auto-resumes, because 368 carries no machine-readable clear time. Health probes send zero Graph traffic while blocked. A cadence cap defers bursts rather than dropping them, and a lane pause kill switch is always available.
Humanizer brand-lint. Captions are checked before publish against editable rules in rules.json. The humanizer layer flags English AI-writing tells. Errors block publish; warnings are advisory.
Honest native scheduling. Where a platform supports it (Facebook scheduled posts, YouTube publishAt), pendpost uses native scheduling, so those posts fire even when your computer is off. Instagram, LinkedIn, and X have no native scheduling, so pendpost must be running to publish them; run it on an always-on host to cover those too. It stays honest about which cover/thumbnail mechanics actually apply per platform.
Dual interface. A web dashboard and any MCP client drive the same contract. A parity test enforces that every capability ships on both faces.
German and Swiss localization. Run the dashboard, digest, and notifications in English or real Swiss German (de-CH, with proper umlauts and 24-hour Swiss dates); set the language in Settings. See the localization docs.

Quickstart

One line, no setup:

bash

npx pendpost
# then open http://127.0.0.1:8090

No account, no signup - pendpost runs locally and starts immediately. You can draft, approve, and schedule a full campaign right away; the first run ships an example "Acme Launch" campaign in data/plans, so you see real content at once. Connect a platform in Setup when you are ready to publish - until then a lane is live-but-unauthenticated, so publishing simply waits for the connection rather than faking it.

Prefer git or docker?

bash

# git
git clone https://github.com/pendpost/pendpost pendpost
cd pendpost
npm start            # or: node bin/pendpost.mjs (builds the dashboard on first run)
# then open http://127.0.0.1:8090

bash

# docker
docker compose up
# then open http://127.0.0.1:8090

The dashboard

The planner and the approval queue are the two screens you live in: draft and schedule on the left, approve or reject on the right. Nothing on the right can approve itself.

Planner	Approval queue

MCP clients

The easiest path self-boots: pendpost speaks MCP over native stdio, so the client launches the server for you - nothing to start first.

Claude Desktop (one-click): install the pendpost .mcpb bundle attached to each GitHub release. It self-boots npx -y pendpost --stdio and opens the approval dashboard in the same process.

Any stdio MCP client:

json

{
  "mcpServers": {
    "pendpost": { "command": "npx", "args": ["-y", "pendpost", "--stdio"] }
  }
}

Advanced / dev (HTTP transport): pendpost also serves MCP over streamable-HTTP at /mcp. This needs the server already running via npx pendpost:

bash

claude mcp add --transport http pendpost http://127.0.0.1:8090/mcp

AI-assisted setup (Claude for Chrome)

Going live on a real platform means creating a developer app in each vendor's portal and running one OAuth ceremony. pendpost makes that agent-drivable: on each incomplete card in the dashboard's Setup page, a Copy AI prompt button copies a ready-to-paste, secret-safe prompt for Claude for Chrome that drives the portal for that one platform. You authenticate at every login/consent gate, and the credential is minted locally by the CLI - it never passes through the agent or the chat. The same setup contract is documented for any agent in AGENTS.md.

What the MCP tools do

Every capability is an MCP tool, and the dashboard mirrors it. Read-only tools can never publish; write tools create drafts and are gated by the approval rules above. Grouped by what they do:

Read and inspect: plan_list, plan_get, account_status, assets_list, activity_log, validate_media, platform_validate, pendpost_health, publish_preview, brand_lint, generate_digest, config_get, clients_overview.
Compose: plan_create_post, plan_update_post, plan_delete_post, campaign_create, campaign_set_active.
Approve (the human gate): approve_post, reject_post. An agent can never approve its own draft.
Schedule and publish: scheduler_set, publish_due_run, reschedule, unschedule, mark_posted, verify_post.
Covers and assets: set_cover, clear_cover, asset_upload, rename_asset, delete_asset.
Insights and safety: fetch_insights, token_refresh, pendpost_record_block, health_recheck, meta_lane_set.
Config and clients: config_set, client_create, client_update, client_archive, client_list, client_set_active.

The authoritative count and the read/write split are derived from lib/mcp.mjs and verified by test/parity-check.mjs (see the MCP docs). Each tool also carries readOnlyHint/destructiveHint/title annotations in tools/list.

Going live

When you are ready to publish, connect each platform in Setup (or copy .env.example to .env and fill in only the platforms you use). A platform publishes once its credential is present; until then it is live-but-unauthenticated and publishing waits for the connection. (PENDPOST_MODE=mock exists only as a test/demo fixture that routes every lane through the credential-free mock driver.)

Each platform has an interactive setup ceremony that writes to .env:

bash

# Meta (Facebook + Instagram). A System User token is preferred for automation.
node scripts/meta-social.mjs setup-system-user \
  --system-user-token <T> --page-id <ID> --app-id <ID> --app-secret <S>
# (or `node scripts/meta-social.mjs setup` for a long-lived Page token)

# LinkedIn (after creating an OAuth app)
node scripts/linkedin-social.mjs auth

# YouTube (after creating an OAuth client)
node scripts/yt-social.mjs auth

# X (Twitter), OAuth 2.0 PKCE (browser). OAuth 1.0a needs no command - see below.
node scripts/x-social.mjs auth

X (Twitter) supports two auth paths; OAuth 1.0a is recommended because it needs no browser and sidesteps the ERR_TOO_MANY_REDIRECTS that some apps hit on X's OAuth 2.0 consent screen.

OAuth 1.0a (recommended, zero browser). In the X developer portal: (1) set the app's User authentication settings to Read and Write FIRST; (2) then under Keys and tokens, generate/regenerate both the API Key/Secret and the Access Token/Secret; (3) paste all four into .env as X_API_KEY, X_API_SECRET, X_ACCESS_TOKEN, X_ACCESS_TOKEN_SECRET. The lane goes live the moment X_ACCESS_TOKEN_SECRET is set - no auth command. Verify with node scripts/x-social.mjs probe (prints connected as @handle). Gotchas, in order:
- All four values must come from the same app and the same generation.
- Regenerating the API Key/Secret invalidates the Access Token/Secret - so always regenerate the Access Token/Secret after the API Key/Secret, never before.
- One wrong character yields 401 "Could not authenticate you" (code 32), which masquerades as an access-token problem. Copy the values from the portal; do not hand-type them.
OAuth 2.0 PKCE (browser). Create an OAuth 2.0 confidential client with callback http://localhost:8087/callback and scopes tweet.read tweet.write users.read media.write offline.access, then run node scripts/x-social.mjs auth. The access token lasts ~2h and the refresh token rotates; pendpost refreshes both automatically before each run.

See .env.example for the full list of environment variables per platform.

Brand-lint + humanizer

Captions run through a brand-lint pass before they can publish. The rule set lives in rules.json, which you edit directly. Each rule is an object with an id, a severity (error blocks, warn is advisory), a matcher (a regex or a platform-aware built-in like captionLength, hashtagCount, allCaps, brokenLink), and a hint. The shipped default also includes humanizer rules that flag English AI-writing tells (AI vocabulary, em dashes, negative parallelism, reflexive rule-of-three padding, filler and hedging, promotional puffery). To disable a rule, delete its object; to add one, append an object. A rules.json at your workspace root overrides the shipped default. brand_lint is both an MCP tool and the dashboard composer's lint panel.

Architecture

pendpost is one zero-dependency Node process (server.mjs) with four faces: a REST API at /api, an MCP server at /mcp (streamable-HTTP, JSON-RPC 2.0, 43 tools), a /media face that range-streams local files under data/, and /, which serves the built React dashboard from app/dist. Backend logic lives in lib/*.mjs. There are four publish engines in scripts/: meta-social.mjs (Facebook and Instagram), linkedin-social.mjs, x-social.mjs, and yt-social.mjs, each spawned as a subprocess on a scheduler tick or on demand and each emitting a JSON envelope. Plans and state are local JSON. The workspace root holding .env, config.json, state.json, and data/ is overridable via PENDPOST_ROOT (default: the install dir).

Platforms

Facebook, Instagram, LinkedIn, X, YouTube. Each engine handles its own auth, publishing, native scheduling, cover/thumbnail mechanics, and read-only insights. X has no native scheduling, so due tweets are published by the scheduler at their scheduled time (like Instagram and LinkedIn).

Security + privacy

pendpost binds 127.0.0.1 (loopback) by default, never phones home, and keeps secrets only in your own .env (which is gitignored). See SECURITY.md for the full posture and how to report a vulnerability.

Status

pendpost is maintained part-time by a small team and is early. Expect rough edges, and set your response-time expectations accordingly.

Docs

The full reference lives at docs.pendpost.com (source in site-docs/). Marketing, pricing, and every download option live at pendpost.com.

License

Install

Configuration

Arguments

--stdio

claude_desktop_config.json

{
  "mcpServers": {
    "pendpost": {
      "command": "npx",
      "args": [
        "-y",
        "pendpost@1.2.0",
        "--stdio"
      ]
    }
  }
}