substack-ops

Standalone Substack CLI + 26-tool MCP server. Your IDE drafts the replies. Zero AI API keys.

Site → substack-ops.chavan.in · Source → 06ketan/substack-ops

Posts, notes, comments, replies, reactions, restacks, recommendations, search, profiles, feeds, automations, MCP server, Textual TUI. One Python install, one binary, MIT licensed.

TL;DR — MCP-native (no API key, one command)

uvx substack-ops mcp install cursor          # or claude-desktop, claude-code, print
# Restart your host. Then in chat:
#   "list unanswered comments on post 193866852"
#   "draft a warm reply to comment 12345"
#   "post that draft"

Your host's LLM (Cursor's, Claude's) does the drafting via the propose_reply / confirm_reply tools. No ANTHROPIC_API_KEY / OPENAI_API_KEY needed.

Setup (dev / from source)

git clone https://github.com/06ketan/substack-ops && cd substack-ops
uv sync
uv sync --extra mcp     # mcp SDK for the MCP server (recommended)
uv sync --extra tui     # textual for the TUI
uv sync --extra chrome  # pycryptodome + keyring for Chrome cookie auto-grab

Auth defaults to ~/.cursor/mcp.json's mcpServers.substack-api.env. Override with env or .env. Or use one of the auth flows in auth login / auth setup.

uv run substack-ops auth verify
uv run substack-ops quickstart   # 20-step tour

Command surface

Grouped by intent. Every write defaults to --dry-run; flip with --no-dry-run (and --yes-i-mean-it for the irreversible ones). All writes land in .cache/audit.jsonl and are dedup-checked against .cache/actions.db.

Auth (4)

Command	What it does
`auth verify`	Confirm the cookie works; print authed user/pub.
`auth test`	Same as verify, exit non-zero on failure (CI-friendly).
`auth login --browser chrome\|brave`	Auto-grab cookie from local Chromium browser via macOS Keychain.
`auth login --email me@x.com`	Email magic-link → paste-the-link interactive flow.
`auth setup`	Interactive paste of `connect.sid` cookie.

Read — Posts (8)

Command	What it does
`posts list [--pub] [--limit] [--sort new\|top]`	List posts from a publication (yours by default).
`posts show <id\|slug> [--pub]`	Post metadata (title, dates, reactions, comment count).
`posts get --slug <slug> [--pub]`	Same as `show` but slug-only.
`posts content <id> [--md] [--pub]`	HTML body (auth-aware for paywalled). `--md` converts to Markdown.
`posts stats <id>`	Engagement counts — reactions, comments.
`posts search <query> [--pub] [--limit]`	Substack-side full-text search.
`posts paywalled <id> [--pub]`	Boolean: is this post paywalled?
`posts react <id> [--off] [--pub]`	Add (or remove with `--off`) a reaction. Defaults to ❤.
`posts restack <id> [--off]`	Restack a post (Substack does not support unrestack).

Read — Notes (5)

Command	What it does
`notes list [--limit]`	Your published Notes.
`notes show <id>`	One note + its reply tree.
`notes publish <body> [--no-dry-run]`	Publish a top-level Note.
`notes react <id> [--off]`	React on any Note.
`notes restack <id> [--off]`	Restack a Note.

Read + Write — Comments (5)

Command	What it does
`comments tree <post_id> [--pub]`	Full nested comment tree as table.
`comments export <post_id> --out file.json [--pub]`	Same tree as JSON.
`comments add <post_id> <body> [--pub] [--no-dry-run]`	New top-level comment.
`comments react <id> --kind post\|note [--off]`	React on a comment.
`comments delete <id> --kind post\|note [--no-dry-run]`	Destructive — your own comments only.

Reply engine (6)

Command	What it does
`reply template <post_id> --template thanks`	Rule-based replies (no LLM).
`reply review <post_id>`	LLM drafts each, you `[a]ccept / [e]dit / [s]kip / [q]uit`.
`reply bulk <post_id> --out drafts.json`	Draft every comment to a file. Edit, set `action: "approved"`.
`reply note-bulk <note_id> --out drafts.json`	Same for replies under a Note.
`reply bulk-send drafts.json [--no-dry-run]`	Posts only `approved` rows. Dedup-checked.
`reply auto <post_id> --no-dry-run --yes-i-mean-it`	Draft + post immediately. 30s rate limit.

Read — Discovery (8)

Command	What it does
`feed list --tab for-you\|subscribed\|category-{slug}`	Reader feed (the Substack app feed).
`profile me` / `profile get <handle>`	Profile.
`users get <handle>` / `users subscriptions <handle>`	Public user info + their subs.
`podcasts list [--pub]`	Audio posts.
`recommendations list [--pub]`	Pub's recommended publications.
`authors list [--pub]`	Pub's contributor list.
`categories list` / `categories get --name <X>`	Substack's category taxonomy.

Automations (3)

Command	What it does
`auto presets`	List built-in YAML rules.
`auto run <name>`	One-shot run a preset.
`auto daemon <name> --interval 60`	Loop forever; logs to audit.

Operations + safety (3)

Command	What it does
`audit search [--kind] [--target] [--status] [--since 7d]`	Query the JSONL audit log.
`audit dedup-status`	Counts in the dedup SQLite DB.
`quickstart`	20-step interactive tour.

MCP server (3)

Command	What it does
`mcp install <cursor\|claude-desktop\|claude-code\|print> [--dry-run]`	Auto-merge config into your host.
`mcp serve`	stdio MCP server (26 tools).
`mcp list-tools`	Print the tool registry.

Other (1)

Command	What it does
`tui`	Textual TUI — 6 tabs (Notes, Posts, Comments, Feed, Auto, Profile).

Multi-publication

Every read command accepts --pub <subdomain|domain>. Defaults to your own publication.

substack-ops posts list --pub stratechery --limit 5
substack-ops posts search "ai" --pub stratechery
substack-ops recommendations list --pub stratechery

Reply modes

Mode	What it does	Safety
`template`	YAML keyword/regex rules under `src/substack_ops/templates/*.yaml`	dry-run default
`review`	LLM drafts each reply, you `[a]ccept / [e]dit / [s]kip / [q]uit`	dry-run default + manual gate per comment
`bulk`	LLM drafts every comment to `drafts.json`. Edit file, set `action: "approved"`	offline review, dedup-checked on send
`bulk-send`	Posts only items with `action: "approved"`	dry-run default; dedup DB prevents the M2 31-dup-replies regression
`auto`	LLM drafts and posts immediately	requires `--no-dry-run --yes-i-mean-it`, 30s rate limit

After every live note-reply the engine re-fetches the new comment and asserts ancestor_path is non-empty. If empty, the audit row's result_status is flipped to "orphaned" (the M2 bug where parent_comment_id was silently dropped — now caught).

Automations

Built-in presets (auto presets):

like-back — when someone reacts to your note, react to their latest note.
auto-reply — same trigger, but post a templated thank-you.
auto-restack — when a watchlist handle posts a new note, restack it.
follow-back — when someone follows you, follow them back.

Custom YAML rules under ~/.config/substack-ops/auto/*.yaml. Loop with auto daemon <name> --interval 60.

MCP server

substack-ops mcp install cursor              # auto-add to ~/.cursor/mcp.json
substack-ops mcp install claude-desktop      # auto-add to claude_desktop_config.json
substack-ops mcp install claude-code         # uses `claude mcp add` under the hood
substack-ops mcp install print               # print the snippet only
substack-ops mcp install cursor --dry-run    # preview without writing
substack-ops mcp serve                       # stdio server
substack-ops mcp list-tools                  # 26 tools

Manual config snippet (if you prefer):

{
  "mcpServers": {
    "substack-ops": {
      "command": "substack-ops",
      "args": ["mcp", "serve"]
    }
  }
}

If the mcp SDK is not installed, the server falls back to a minimal stdin/stdout JSON-line dispatcher that's still useful for scripting:

echo '{"tool":"list_posts","args":{"limit":3}}' | substack-ops mcp serve

MCP-native draft loop (no API key)

3 tools designed to let your host LLM draft for you:

Tool	What it does
`get_unanswered_comments`	Returns the worklist: comments where you have not yet replied (any depth).
`propose_reply`	Dry-run only. Returns a `token` + payload preview. No write.
`confirm_reply`	Posts a previously-proposed reply by token. Idempotent via dedup DB. Token TTL 5 min.

Differentiator tools (the safety + drafting stack that makes the unattended mode safe): bulk_draft_replies, send_approved_drafts, audit_search, dedup_status, get_unanswered_comments, propose_reply, confirm_reply.

LLM strategy

Two layers, both free:

MCP-native (default). Host LLM drafts via propose_reply / confirm_reply. No env vars, no API key. Use this for interactive replies.
Subprocess CLI (daemon path). For reply auto / auto daemon when no human is in the loop. Auto-detects claude (Claude Code), cursor-agent, or codex on PATH. Override with SUBSTACK_OPS_LLM_CMD.

There is no paid-API-key path. If you want one, vendor the old _anthropic / _openai methods from substack-ops v0.2.0 yourself.

Textual TUI

substack-ops tui

6 tabs: Notes / Posts / Comments / Feed / Auto / Profile. Sub-tabs: 1=mine, 2=following, 3=general. Keys: tab, 1-3, ↑/↓, enter, r, l, s, o, q/esc.

Auth methods

substack-ops auth verify                  # uses mcp.json or env
substack-ops auth login                   # auto-grab cookies from Chrome (macOS Keychain)
substack-ops auth login --browser brave
substack-ops auth login --email me@x.com  # email magic-link, paste-the-link mode
substack-ops auth setup                   # interactive paste cookies

Architecture

mcp.json | env | Chrome | OTP  →  auth.py / auth_chrome.py / auth_otp.py
                                            │
                                  .cache/cookies.json
                                            │
                                  SubstackClient (httpx)
                                            │
   ┌──────┬──────┬───────┬───────┬───────┬──────┬──────┬─────┬──────┐
   ▼      ▼      ▼       ▼       ▼       ▼      ▼      ▼     ▼      ▼
 posts  notes  comments  feed  profile  users  recs  cats  ...   reply_engine
                                                                       │
                                                       ┌───────────────┼────────────┐
                                                       ▼               ▼            ▼
                                                  template       ai_review     ai_bulk + ai_auto
                                                       └───────────────┬────────────┘
                                                                       ▼
                                                            base.post_reply / post_note_reply
                                                                       │
                                                              ┌────────┼────────┐
                                                              ▼        ▼        ▼
                                                            dedup    audit  ancestor_path
                                                            (SQLite) (jsonl)  guardrail
   auto/engine.py ────────────────┐
   mcp/server.py  ──── 23 tools ──┼─── all share SubstackClient
   tui/app.py     ──── 6 tabs   ──┘

Endpoints used

Action	Method + URL
Auth check	`GET https://substack.com/api/v1/subscriptions`
List posts	`GET {pub}/api/v1/archive`
Post by id	`GET {pub}/api/v1/posts/by-id/{id}`
Post by slug	`GET {pub}/api/v1/posts/{slug}`
Post content	same as above; `body_html` field
Post search	`GET {pub}/api/v1/archive?search=`
Comments	`GET {pub}/api/v1/post/{id}/comments?all_comments=true`
Reply to comment	`POST {pub}/api/v1/post/{id}/comment` body `{body, parent_id}`
Add top-level comment	same with `parent_id: null`
React to post	`POST {pub}/api/v1/post/{id}/reaction` body `{reaction}`
Restack post	`POST https://substack.com/api/v1/restack` body `{post_id}`
Restack note	`POST https://substack.com/api/v1/restack` body `{comment_id}`
Delete post-comment	`DELETE {pub}/api/v1/comment/{id}` (PUB host)
Delete note	`DELETE https://substack.com/api/v1/comment/{id}` (BARE host)
My notes	`GET https://substack.com/api/v1/reader/feed/profile/{user_id}`
Note thread	`GET https://substack.com/api/v1/reader/comment/{note_id}`
Note replies	`GET https://substack.com/api/v1/reader/comment/{note_id}/replies`
Publish note	`POST https://substack.com/api/v1/comment/feed` body `{bodyJson}`
Reply to note	same with `{bodyJson, parent_id}` (NOT `parent_comment_id` — known M2 bug)
React to comment	`POST {host}/api/v1/comment/{id}/reaction` (host = pub for post-comments, substack.com for notes)
Recommendations	`GET {pub}/api/v1/recommendations/from/{publication_id}`
Authors	`GET {pub}/api/v1/publication/users/ranked?public=true`
Categories	`GET https://substack.com/api/v1/categories`
User profile	`GET https://substack.com/api/v1/user/{handle}/public_profile` (auto-redirects on 404)
Reader feed	`GET https://substack.com/api/v1/reader/feed/{recommended\|subscribed\|category/{slug}}`

Tests

uv run pytest -q     # 43 tests, ~0.6s, no live network

Coverage today: auth, client (read+write+engagement+delete), reply engine, dedup DB, audit log search, MCP tool registry & dispatcher, automation engine preset loader, the M2 parent_id regression test, the M2 host-mismatch regression test.

GSD workflow

.planning/ scaffold for Get Shit Done under ~/.claude/skills/gsd-*. Roadmap at .planning/ROADMAP.md, per-phase plans at .planning/phases/M*/PHASE.md.

Known gaps

Full email stats (opens/clicks/views) — needs dashboard CSRF flow. Fallback: Playwright MCP scrape.
Reactions endpoint shape on POST/DELETE not yet probed live; current shape is a best-guess from upstream tool catalogs.
Auto-engine new_follower / new_note_from triggers are stubbed (return note: "trigger not yet implemented").
TUI sub-tabs (1/2/3) and reply/like/restack key bindings are scaffolded but not wired to the client yet.
Chrome cookie auto-grab tested only for macOS Chrome; Brave path included; Linux/Windows not supported.

License

MIT. See LICENSE.

The vendored httpx-port helpers under src/substack_ops/_substack/ are derived from the MIT-licensed NHagar/substack_api package — kept here so this repo ships zero runtime dependencies on third-party Substack libraries. Attribution preserved in each file's module docstring.

substack-ops

Quick Install

substack-ops

TL;DR — MCP-native (no API key, one command)

Setup (dev / from source)

Command surface

Auth (4)

Read — Posts (8)

Read — Notes (5)

Read + Write — Comments (5)

Reply engine (6)

Read — Discovery (8)

Automations (3)

Operations + safety (3)

MCP server (3)

Other (1)

Multi-publication

Reply modes

Automations

MCP server

MCP-native draft loop (no API key)

LLM strategy

Textual TUI

Auth methods

Architecture

Endpoints used

Tests

GSD workflow

Known gaps

License

Reviews