More intelligence per engineer

Same coding agents, more results. Elbow helps your team build software more efficiently and at a fraction of the cost.

Get started with Elbow

Works withClaude CodeCodexConductorT3 Code

Ship more with

Your engineers won't notice the switch

They keep their harness, their prompts and their settings. Elbow decides how hard to think at every step, and the work comes back the way it always has.

~/fieldline/payments-apiCodex
  1. codex
  2. Partial refunds round the tax the wrong way on multi-currency orders. Fix it and add a test.
  3. Explored└ Read src/refunds/partial.ts
  4. Search roundTax in src/
  5. Read src/money/currency.ts
  6. Edited src/refunds/partial.ts (+9 -4)58 - const tax = roundTax(refund.tax);58 + const tax = taxPerLine(refund, order.currency);
  7. Edited test/refunds.test.ts (+22 -0)
  8. Ran npm test└ 212 passed
  9. Tax is now rounded per line in the order's currency before the refund is split, so a ¥ order no longer loses a unit. The new test covers JPY, EUR and a mixed basket.
  10. Open a PR
  11. Ran gh pr create --fill└ github.com/fieldline/payments-api/pull/2184
  12. Opened #2184, "Fix partial refund rounding on multi-currency orders".

Ask Codex to do anything

? for shortcuts88% context left
Fix partial refund rounding on multi-currency ordersECEmma Clarkepayments-apiCodex
0:00FastRead$0.01
0:21FastSearch$0.00
0:34FastRead$0.01
1:02ThinkingEdit$0.38
2:47BalancedEdit$0.09
3:30FastRun$0.01
3:58BalancedReply$0.04
4:31FastRun$0.01
4:44FastReply$0.01
9 steps, 4m 51s$2.24$0.56
Open#2184Fix partial refund rounding on multi-currency orders+31 -4

Opus level results, four times the work per dollar

On Terminal-Bench 2.1, the public test of real work in a terminal, Elbow solves 86% of tasks to Opus 5.5's 88%, at $0.06 per solved task against $0.27. Sonnet 5.5 solves 78% for more than one and a half times Elbow's cost.

ModelTasks solvedCost per solved task
Opus 5.5medium effort88%$0.27
ElbowBalanced86%$0.06
Sonnet 5.5medium effort78%$0.09
Terminal-Bench 2.1, 80 tasks, the same agent for every model at the effort shown. Scores are pass@1, one try at each task, averaged over three runs for Elbow and five for the others; the lines show each score's 95% margin of error. Cost per solved task is everything spent, failed runs included, divided by the tasks solved. October 2026.

Say yes to the bigger jobs

Your AI bill drops from the first session for the same work, so the jobs that were waiting for budget get started. You see where every person and team puts their AI, give each a monthly budget, and hear about it before one runs out.

Members (46)Top 8 by spend
PersonTeamModesSpend, this month
LKLena Krauslena.kraus@fieldline.devData$624.60 of $600.00
ECEmma Clarkeemma.clarke@fieldline.devPayments$612.40 of $750.00
TRTomás Reyestomas.reyes@fieldline.devPlatform$588.15 of $750.00
JWJun Weijun.wei@fieldline.devPlatform$471.90 of $600.00
SOSam Okaforsam.okafor@fieldline.devMobile$433.25 of $600.00
MDMarco Diazmarco.diaz@fieldline.devWeb$318.75
ABAisha Belloaisha.bello@fieldline.devPayments$296.40 of $500.00
NBNoah Bergnoah.berg@fieldline.devMobile$254.10 of $500.00
#eng-leadsSlack
E
Elbow 9:02 AM

Payments has used 82% of its $3,300.00 October budget with 8 days to go. Emma Clarke is closest to her budget, at $612.40 of $750.00.

Raise the budget Open Payments

Invoice, September 2026Issued Oct 1
Claude Code
$12,904.18
Codex
$1,377.40
Total, one invoice
$14,281.58

Lines by team, agent and mode, and as CSV.

Everything a team needs to run its AI

Spend, people, sessions and security in one place, so the whole team can lean on its agents with nothing to chase.

Intelligence

  • Thinks as hard as the step needsQuick on the reading and searching, deliberate on the calls that decide whether the change is right.
  • Fast, Balanced and Thinking modesElbow picks the mode for each step, and anyone can hold a session in one mode when they want to.
  • Failover between providersWhen a provider slows down or refuses, the request moves on and the session carries on.
  • Checked on every releaseEach release runs against held-out coding benchmarks before it ships.

Spend

  • Monthly budgetsOne for the org, each team and each person. Each warns or stops at 100%, your choice.
  • Alerts by email and SlackAt 80% and 100%, to your admins and to the person whose budget it is.
  • Spend by repoSee which repositories and teams the money goes to, down to a single session.
  • One invoice a monthEvery engineer and every agent on one bill, with lines by team, agent and mode.

Visibility

  • Session historyEvery session: what was asked, each step and its mode, how long it took and what it cost.
  • InsightsAgent hours against people's time, and the repos, tools and agents your team leans on.
  • Search everythingPress ⌘K to jump to any person, team, repo or session.
  • Usage export and APICSV for finance, and a read-only key for your own dashboards.

Team management

  • Orgs and invitesCreate your org, invite engineers by email, and let your company domain join on its own.
  • TeamsGroup people into teams, each with its own budget, spend and top repos.
  • RolesAdmins set budgets and policy; engineers just code.

Agents

  • Claude Code, Codex, Conductor, T3 CodeOne install starts any of them through Elbow.
  • Self-updating CLIEngineers install once; updates arrive on their own.
  • Keys stay with usEngineers never hold a provider key, so there is nothing to leak or rotate on their laptops.

Security

  • Audit logEvery sign-in and admin change, searchable and kept.
  • DevicesSee every signed-in laptop and sign one out in a click.
  • Transcripts on your termsKeep what was asked for session history, or turn it off for the whole org.

Where software teams get their AI.