oneshot-gtm
Open-source GTM agent for technical founders. Pay-per-result. Signed receipts. Founder-led discipline encoded. Two surfaces: terminal CLI + local web dashboard, both backed by the same SQLite ledger.
# Dashboard-only (published, no clone needed):
bunx oneshot-gtm-server # opens http://127.0.0.1:3030
# Full install (CLI + dashboard, repo clone — see below):
bun run cli -- init # one-time setup
bun run cli -- ui # opens http://127.0.0.1:3030
What this is
A focused, opinionated wrapper around OneShot — a pay-per-use API toolbox for email, SMS, voice, deep research, person enrichment, browser automation, and website build, all settled per-call in USDC on Base with cryptographically signed receipts for every action.
OneShot is the toolbox. oneshot-gtm is the strategy wrapper: it encodes the canonical PMF + founder-led-sales playbook (Mom Test, Sean Ellis 40%, Predictable Revenue, do-things-that-don't-scale, multichannel cadence, signed-receipt CAC) as a set of named plays you actually run from the terminal or the dashboard.
It's open source (MIT) so you can read every prompt, fork every play, and trust what's running.
Two ways to use it
Terminal — for power users + scripting
bun run cli -- intel advise # interactive coach
bun run cli -- discover icp interview-prep "your hypothesis"
bun run cli -- find watch --once # poll all due triggers, enqueue candidates
bun run cli -- find drain podcast-guest --dry-run # preview approved /queue rows
bun run cli -- cadence advance # daily tick: poll inbox + fire follow-ups
~30 commands across 9 groups. See bun run cli -- --help or jump to the Command map.
Dashboard — for visibility + non-technical co-founders
bun run cli -- ui # default: serves prebuilt React app on 127.0.0.1:3030
bun run cli -- ui --dev # vite hot-reload + API server in parallel
bun run cli -- ui --port 4000 # custom port
bun run cli -- ui --no-browser # don't auto-open
Eight pages, all reading the same ~/.oneshot-gtm/ledger.sqlite:
- Home — spend (7d / 30d), reply rate trend, in-flight cadences, recent receipts
- Queue — triggers table (enable, edit JSON config, fire) + target queue (status + play filters, bulk approve). Per-play Drain opens a modal that hands off to
/run/<play>with approved rows pre-loaded — every draft + lint flag streams live, and the latest draft persists on the queue row so you can re-read it later by expanding the row (subject + body + flags + receipt links). Per-row spinner + locked button while a trigger is running. - Replies (
/inbox) — every reply matched to its prospect + play + cadence status by sender address, merged across all sender identities, with a match-status filter (all / matched / no match) to cut newsletter + bounce noise. Expand a reply to answer it in place: write the draft yourself or generate with the LLM (founder voice, primed with the inbound message + your prior touches), edit, and send. Replies go out from the identity that received them and thread on both transports — Gmail via In-Reply-To / References, OneShot via the platform'sreply_to_email_id(SDK 0.22). Sends carry an idempotency key, so a retry after a timeout can't double-send - Cadences — table view with inline Stop + Log outcome buttons; outcome modal supports
meeting_booked / sql_qualified / deal_won / deal_lost / ghosted. Logging an outcome — or a reply landing — tags the cadence's value back to OneShot (RoCS), so/measureand/receiptsreflect what each cadence earned. Per-row chevron exposes the next-step draft preview before send; bulk select for batch preview/send; sent-step history collapsible inline; pulsing "sending" badge while a fire-and-forget send is in flight. - Receipts — paginated table with a memo column (the "why" of each call) and a value chip plus an all / valued / unvalued filter; click a row → modal with the signed receipt payload alongside its memo + structured
decisionContext - Plays — cards with channel badges + Run button (for
show-hn/job-change/post-funding/accelerator-batch/hiring-signal/podcast-guest/stack-consolidation/repo-interest) + Copy CLI button - Measure — CAC + RoCS tables filterable by time range, plus a RoCS by cadence section: per-cadence spend vs tagged value vs RoCS multiple, grouped by goal
- Setup — editable wizard: founder profile, LLM provider/model, OneShot wallet keys (hidden inputs), telemetry toggle. Saves to chmod-600
~/.oneshot-gtm/.env.
The Run a play form (/run/$playName) takes editable target rows + a dry-run toggle and streams drafted emails back via Server-Sent Events with lint flags + clickable receipt links. When arriving from /queue via the Drain button, target rows auto-hydrate from approved queue rows and each generated draft persists back to its originating row on completion — re-readable from /queue at any time.
A floating strategist dock is mounted on every page. Open it to chat through trigger config: it reads your ICP + product one-liner and proposes JSON configs as confirmation chips you click to apply. Endpoint: POST /api/strategist/stream (SSE).
Discovery — where targets come from
Motion plays don't require hand-curated JSON anymore. Ten finders auto-discover prospects, ICP-filter them, and enqueue into /queue for one-click approve / reject:
show-hn— HN Algolia poller, surfaces same-day Show HN postspost-funding— webSearch by ICP-derived industry × round (auto), or a TC/Crunchbase URL listjob-change— webSearch for"joined as <persona>"announcements with persona + company filtershiring-signal— Greenhouse / Lever / Workable / Ashby ATS searchpodcast-guest— recent-guest discovery across Latent Space, Lenny's, 20VC, Acquired, Invest Like the Bestaccelerator-batch— yc-oss directory + websearch fallback for non-YC cohorts (Techstars, Antler, 500 Global, AI Grant)github-topics— GitHub-API manifest scan (package.json,pyproject.toml,requirements.txt) detects vendor stack deterministically; finds repos stitching together N agent vendors as competitor-switch targetsgithub-stars— recent stargazers of repos you watch, routed per repo: tag a repocompetitor(→ competitor-switch) oradjacent(→ repo-interest, a "you're into X, my product helps" intro)luma-events— upcoming events from Luma's own city pages (genuinely upcoming, not search-indexed leftovers), gated per event by a topic + ICP check before any spend; pitches the hosts + featured guests Luma exposes publicly per event — with their LinkedIn/website, so contact resolution actually lands. Each queue row is taggedHostorGuestand drafted accordinglybreakup-revive— scans the local ledger for prospects cold for 60-90 days
Each finder runs as a trigger with its own interval + spend cap — click the interval in the /queue triggers table to change cadence (presets 1h–7d, or revert to the default). Captured per-prospect signals (LinkedIn URL via webSearch + phone via passive enrichment when surfaced) show next to the email + company in /queue. Approved rows ship via bun run cli -- find drain <play> or the per-play Drain button on the Queue page.
The dashboard server runs an in-process scheduler that fires enabled triggers on their interval automatically — open bun run cli -- ui, enable a trigger, and it polls without you needing a separate find watch daemon. The CLI watch command stays useful for cron + headless deployments where you don't want the dashboard.
/home surfaces a Scheduler section per trigger — state pill, last-run summary (the cand=N · kept=M · icp=K · $X.YY line you see on /queue), last polled, next due — so "is the scheduler alive?" is a glance, not a grep events.jsonl. Overdue triggers show in oxblood; disabled ones collapse behind a chevron.
A trigger whose stored config is missing required inputs (e.g. accelerator-batch without a cohort) reads as not ready on /queue — the Enable toggle and Run Now button are disabled with the reason in a tooltip. Edit config via the pencil icon to clear it; nothing fires while a trigger is unready. The same gate returns 409 on POST /api/triggers/:name/enabled and :name/run, so scripted callers can't bypass it.
Before any findEmail call, a pre-flight check skips dud domains (free-tier subdomains like *.vercel.app/*.github.io, social hosts, link aggregators, personal email providers) and inputs where the "name" is obviously a username (samaralihussain, no whitespace or period). On a 50-candidate Show HN run that historically dropped ~37 of 50 at the SDK, the prescreen now eliminates the wasted spend at ~$0.05/call — roughly $1–2 saved per run. Skipped rows log a finder.skipped_findemail event with the reason for later blocklist tuning.
Sender rotation — OneShot domains or your own Gmail
Outbound ships through a sender identity pool: any mix of OneShot wallet-owned sending domains — multiple domains, and multiple mailboxes within one domain (e.g. jane@acme.com + sales@acme.com) — plus your own Gmail / Google Workspac