- Building
TL;DR
What I shipped DARP: finds the repetitive work in an accounting firm, builds an AI agent for it, and proves whether it saved time. Desktop app (Mac/Windows) plus a live web demo Agent accuracy 80 invoices on HarnessRouter: 97.5% right, 0 dangerous. Inbox and W-9 agents: 40 of 40 each HarnessRouter usage Harness chrn_e03cf193f58143979768bb2b95a2468c, Claude Code base, Sonnet 5.5, about 2.56M input tokensReal API test Intuit's QuickBooks sandbox: 17 of 17 checks passed Code About 27,500 lines of TypeScript in 198 files, 444 unit tests, 35 browser tests, 44 commits How I built it With Claude Code as my coding agent, plus 5 discovery calls with accounting firms Live demo taupe-alpaca-79be2f.netlify.app What I built
DARP, for accounting firms. It finds the repetitive work inside a firm, ranks it by what it costs, builds an AI agent to do it, and proves whether that agent saved time. It runs on the firm's own computers, and a person approves every change. Four steps, which are the app's four pages:
- Discover: records which app and window are in use (never keystrokes or screenshots), finds repeated sequences, and ranks them by hours and dollars per month.
- Build: drafts an agent plan from the observed steps, with checks and hand-offs to a person.
- Run: the agent reads email, PDFs and spreadsheets and proposes changes in QuickBooks Online and Gmail. Nothing is written without approval.
- Prove: conservative metrics from real runs only. Practice runs never count, review time counts against the agent, and there is no verdict under 10 items.
Stack: React, Vite, TypeScript, a small Node server, an Electron desktop app, QuickBooks Online and Gmail over OAuth, MCP over stdio, and DARP's own PDF reader. About 27,500 lines in 198 files, 444 unit tests, 35 browser tests, 44 commits.
Timeline (2026, Pacific time)
When What happened Nov 15 5 discovery calls (5+ hours) to learn what accounting firms actually need. Nov 15-17 Core built, about 30 commits: on-device redaction, workflow mining, a hash-chained audit log, a run engine with an approval gate, the invoice and W-9 agents, and Prove's statistics. Nov 17, 11am Desktop app with a built-in activity recorder. Merged as PR #1: 118 files, about 15,700 lines. Nov 17, evening Accuracy and security audit: blocked DNS rebinding and cross-site writes, split back-to-back repeats into tasks, labelled dollars as estimates, made Prove stricter. Nov 18, 12am First large runs: 60 invoices (95.0%, 2 dangerous) and 80 (98.8%, 1 dangerous). Fixed the causes. Nov 18, 7-8am Agent working from an inbox: 70.0% right, because it searched for the word "invoice". After the fix: 40 of 40. Nov 18, 10:30am W-9 collection agent: 40 of 40 right. Nov 18, 2-3pm Intuit's QuickBooks sandbox: 9 of 14 checks, then 17 of 17 after fixes, including a real bill entered and a duplicate and bank-change request held. Nov 18, evening Moved the agent onto HarnessRouter: 80 invoices, 97.5% right, 0 dangerous. Emailed 18 Bay Area firms for pilots. How DARP uses Claude
Feature What DARP does with it Tool use (agent loop) Read tools run at once; write tools ( propose_bill,propose_email...) become approval cards instead of writes.Plans and write-ups Build drafts the agent's plan from observed steps; the prompt includes the agent's real tools and DARP's enforced limits, so it can't promise what the agent can't do. Scanned PDFs Normal PDFs are read locally. Only scanned ones go to Claude. Claude Code My coding agent for most of the build, and the fallback engine for the test bench. Real runs
A test bench generates invoices with known right answers: 24 kinds (PO match and overrun, duplicates, unknown vendor, bank-detail change, prompt injection, credit memo, foreign currency and more), 5 date styles, 4 layouts. The number I watch is Dangerous: a wrong bill, or one that should have been held, that passes every check. Target: 0. All runs used Sonnet 5.5.
Run (Nov 18) Items Right Dangerous Notes Bill entry, 12:00am 60 95.0% 2 A $0.09 arithmetic slip; a due date from terms Bill entry, 12:06am 80 98.8% 1 "Net 30" due date off 4 reruns, 12:09am 20 each 55-95% 0 CLI failures dropped items silently, which led to the safety net Client reminders, 12:22am 60 100% 0 Never wrote to an unverified "new" address Real PDFs, 12:46am 40 82.5% 0 5 runs failed on the local CLI; every hold was right From the inbox, 4:48am 40 70.0% 1 11 dropped: searched for "invoice" From the inbox, fixed, 7:56am 40 100% 0 W-9 collection, 10:30am 40 100% 0 QuickBooks sandbox, 2:17pm 17 checks 17 passed n/a Real bill entered; duplicate and bank change held Rescoring every saved run with the final rules: 424 invoices, 94.1% right, 1 dangerous mistake, now blocked. Sonnet 5.5 and Opus 5.5 both scored 60 of 60 on the same set, so firms get Sonnet at about half the price.
Running the agent on HarnessRouter
Harness
chrn_e03cf193f58143979768bb2b95a2468c(DARP AP agent, baseclaude-code,claude-sonnet-5.5). DARP uploads abatch.json(procedure, allowed decisions, books snapshot, invoices), starts a background response with anIdempotency-Keyand caps (120 steps, 30 min, $2), polls, and reads backdecisions.json. Every decision then goes through DARP's own checks, so HarnessRouter got no shortcut and DARP lost no safety.Run Invoices Right Dangerous Notes First task 4 n/a n/a 402 no_key: the trial needs your own model keyWith TokenRouter credits 4 100% 0 Over-limit, credit memo and bank change held 3 at a time 24 of 80 95.8% 0 14 batches refused: 429 one_at_a_time2 at a time, with retry 80 97.5% 0 39 of 39 holds right; median 115 s per batch About 2.56M input tokens (2.24M from cache) and 46k output, about 32k input tokens per invoice.
What the runs taught me
- The model is rarely the weak point; the plumbing is. Fixing how the agent searched the inbox took it from 70% to 100%.
- Never trust the model with arithmetic that touches money. Every bill must carry the printed total, and DARP blocks it unless the lines add up to the cent.
- "Nothing dropped silently" has to be enforced in code. Every unfinished item and unopened attachment goes to a person by name.
- Prompts are guidance; checks are guarantees. Duplicates, the $25,000 limit, PO tolerance, unknown vendors and totals are enforced by DARP right before the write, whatever the AI says.
Problems I hit, and fixes
Problem Fix QuickBooks refused invoice numbers over 21 characters, and silently accepted a malformed date. Both checked before approval. The inbox agent missed vendors who don't write "invoice". It reads every message; unopened attachments go to a person. The claudeCLI failed mid-batch and items vanished.An unhandled-item safety net, with throttled batches retried and reported separately. Discover reported 54 hours a month where the truth was 15. Routine mining with a significance test; now within 6-11%. The local server answered a foreign Host header (DNS rebinding). Loopback only, with Host and Origin checks; 403 otherwise. The AI failed when the app was opened from the Dock (no claudeon PATH).Calls the Anthropic API directly when a key is set. - Blocked
-
QuickBooks production access waits on Intuit's review. Until then the agent can only write to sandbox companies, so pilots start with Discover and practice runs.
-
The Mac and Windows apps are unsigned until I buy an Apple Developer account and a Windows code-signing certificate. The first launch shows a warning, and the Windows installer has never been built.
-
Cost per invoice is unmeasured. My test runs used the local Claude Code login, which doesn't report cost per call.
-
- Loved
-
Sonnet 5.5 matched Opus 5.5. Both scored 60 of 60 on the same invoices, so firms get the cheaper model with no loss I could measure.
-
The hard cases held. A prompt injection asking for $21,751.40 instead of $2,175.14 was entered at the true amount and flagged as fraud. Bank-detail changes and spoofed senders were held every time.
-
HarnessRouter's API was easy to build on. Background mode plus Idempotency-Key made polling and retries simple, the files API made a sandboxed agent easy to feed and read back, the one_at_a_time error was clear, and the API reference matched the live service exactly.
-
The QuickBooks sandbox. It found two real problems in an afternoon that 400+ mocked tests never could.
-
- Want
-
A shorter path to production access for small, read-mostly apps with a person approving every write.
-
A lighter verification path for read-only plus drafts access used by a handful of known pilot firms. Until then, every tester must be added by hand as a test user.
-
Say at sign-up that the free trial needs your own model key. My first task was refused with 402 no_key. The fix (TokenRouter credits under Bring Your Own Key) was in a hackathon PDF, not in the API error or the Skill.
-
Report spend when the model runs on your own key. Every session reported spend_usd 0.00 on TokenRouter credits, so I can count tokens but not dollars per invoice.
-
Loading