Testing nowReplaces Zapier
sheetsmith
github.com/sample/sheetsmith
Chains spreadsheet triggers to AI steps from one YAML file, no per-task billing.
Playbooks
Tools you can run yourself. Tested before you try them.
Open-source AI tools that claim to replace paid software, run on real jobs. What they cost, what broke, and where DIY stops.
Sample entries, shown while the first tests are written up.
Testing nowReplaces Bland AI
voicedesk
github.com/sample/voicedesk
Outbound reminder calls for a list of about 100 people from a single config.
QueuedReplaces Docparser
docsieve
github.com/sample/docsieve
Pulls invoice fields with a local model, so documents never leave the machine.
QueuedReplaces Buffer
postline
github.com/sample/postline
Drafts a week of posts from one long article and schedules them itself.
QueuedReplaces Apollo
crawlkit
github.com/sample/crawlkit
Builds a prospect list from public pages and scores fit against your own criteria.
QueuedReplaces Front
replybench
github.com/sample/replybench
Sorts a shared inbox and drafts replies, with a review queue built in.
leadloom
github.com/sample/leadloom
VerdictUse with caveats
Solid enrichment for short lists, but it stalls past a few hundred rows without a queue.
ReplacesClay$149 / month
Cost per run50 leads: $0.45500 leads: $4.60, about 22 minutes
Setup time40 minutesComfortable with a terminal and API keys
Tested on2 Oct 2026v0.8.2
- Repo
- sample/leadloom
- Base model
- Claude Haiku 4.5
- APIs
- Search API, email verifier
- Hosted on
- $6 / month VPS
- Clone the repo and copy the example settings file.
- Add your model and search API keys to the settings file.
- Put your leads in a CSV with a company name and website column.
- Run the enrich command and point it at the CSV.
- Open the output file and spot-check ten rows before trusting the rest.
- Rate limits stopped the run at row 180.Added a two-second pause between rows in the settings file.
- About one in twelve job titles were guessed, not found.Turned on the option that leaves a field blank when no source is found.
When the list passes a few thousand leads a week, needs to write straight into your CRM, or a wrong enrichment costs you a deal. That needs a queue, retries, evaluation against a checked sample, and monitoring.
Book a callcallnote
github.com/sample/callnote
VerdictUse it
Accurate summaries with clear next steps, and it ran on the first try.
ReplacesFireflies$19 / seat / month
Cost per run1 hour of audio: $0.38Transcription runs locally, so only the summary is billed
Setup time15 minutesCan follow install steps, no coding
Tested on29 Sep 2026v1.3.0
- Repo
- sample/callnote
- Base model
- Whisper large-v3, Claude Sonnet 5.5
- APIs
- Model API only
- Hosted on
- Your own laptop
- Install with the one-line command in the readme.
- Add your model API key when it asks.
- Drop a recording into the watch folder.
- Wait for the summary file to appear next to it.
- Edit the summary template once so it matches how your team writes.
- Speaker names came out as Speaker 1 and Speaker 2.Added the attendee list to the prompt file, which fixed most calls.
- A 90-minute recording ran out of memory on an 8 GB laptop.Switched to the medium transcription model in settings.
When calls must be captured automatically for a whole team, pushed into the CRM, and kept under a retention policy. That is integration and compliance work, not a script.
Book a callinboxpilot
github.com/sample/inboxpilot
VerdictSkip
It answers confidently from the wrong document too often to put in front of customers.
ReplacesIntercom Fin$0.99 / resolution
Cost per run200 tickets: $0.90Cheap to run, costly when it is wrong
Setup time2 hoursNeeds someone who can debug a webhook
Tested on24 Sep 2026v0.4.1
- Repo
- sample/inboxpilot
- Base model
- Claude Haiku 4.5
- APIs
- Embeddings API, helpdesk webhook
- Hosted on
- Free tier container host
- Deploy the container with the button in the readme.
- Upload your help documents on the admin page.
- Connect the helpdesk webhook and send a test ticket.
- Run it in draft-only mode so nothing goes out unreviewed.
- Compare its drafts with what your team would have sent.
- 31 of 200 test answers cited the wrong help article.No fix found. Smaller document chunks helped slightly.
- The webhook dropped tickets during a restart.Unresolved. There is no retry in this version.
Straight away, if customers will see the answers. Support needs retrieval you can measure, a tested hand-off to a person, and reliability this repo does not have yet.
Book a callOutgrown the playbook?
If the workflow is proven and now needs scale, integrations or reliability, that is an FDE engagement.
See FDE engagements