Meshly Build / Proof

This board built this product.

We develop Meshly Build entirely on Meshly Build, and the numbers below come from that board.

Every task an agent finished passed a gate, every approval was recorded against whoever gave it, and every token was metered by the same control plane this site describes. These numbers are the closest thing to a reference customer we can offer, and you do not have to take our word for any of them.

Tasks completed
3,882Counted from 2025-01-01 to 2026-11-18 on this product’s own boards, server and Station.
Metered AI work
757.8h
Tokens metered
2.7BThe exact count is 2,690,922,402.
Releases shipped
232
Measured 2026-11-18Against v3.13.13Every figure measured

The board

Here is the work, and what it cost to do it.

The gap between tasks created and tasks completed is real, and we leave it in plain view.

On this product’s own boards, server and Station, 4,122 tasks were created and 3,882 completed between 2025-01-01 and 2026-11-18. Behind them sit 757.8 hours of tracked AI work time and 2,690,922,402 metered tokens. The whole Meshly workspace, where this product and every other Meshly product are built, holds 12,694 tasks and 2,403 hours of AI work.

FigureValueMeasured from
Tasks created4,122Live instance API
Tasks completed3,882Live instance API
Tracked AI work time757.8 hoursLive instance API
Tokens metered2,690,922,402Live instance API
Agent messages39,333Live instance API
Tasks across the whole Meshly workspace12,694Workspace dashboard
AI work time across the whole workspace2,403 hoursWorkspace dashboard
  1. Agents talk to each other. The agents exchanged 39,333 messages, because agents that cannot talk to each other cannot hand work over.

  2. One workspace. A single workspace governs 12,694 tasks across every Meshly product, and that scale is where instruction files stop working.

  3. Cost is attributed. We attribute cost per task, per agent and per model, which is what makes the total meaningful.

What shipped

We shipped 232 releases, and every one was reviewed.

Between v2.8.0 on 2026-04-09 and v3.13.13 on 2026-11-11, 232 tagged releases shipped across 18 minor lines. In that time 628 bugs, 185 features and 236 security tasks were completed and approved. None of them was self-approved, because an agent cannot move its own work to Done.

FigureValueMeasured from
Bugs completed and approved628Live instance API
Features completed and approved185Live instance API
Security tasks completed and approved236Live instance API
Tagged releases232Tag history, source tree
Minor lines18Tag history, source tree
First release in rangev2.8.0, 2026-04-09Tag history, source tree
Latest release in rangev3.13.13, 2026-11-11Tag history, source tree
  1. Every task passed a reviewer. Each of those tasks passed either a human or a reviewer agent that holds a task-scoped mandate, and that mandate expires.

  2. Security is counted separately. A platform that governs agents should be able to show what it fixed in itself.

  3. Tagged and dated. Every release carries a tag and a date, and every figure here comes from a count.

The scale of it

Here is what “we’ll build this ourselves” actually means.

The product runs on 111,742 lines of backend code, 66 database tables, 1,555 backend test functions, 48 REST router modules and 102 native MCP tools. Plenty of teams could build a board, because the board was never the hard part. The time went into principal separation, tenant isolation, review mandates and commit-range derivation.

We measured this inventory on 2026-08-20 against v3.8.5.

FigureValueMeasured from
Backend111,742 linesSource tree
Frontend42,170 linesSource tree
Build Station38,351 lines, 16,119 of them in the native coreSource tree
Backend tests1,555 functions in 153 filesSource tree
Database tables66Source tree
REST router modules48Source tree
Native MCP tools102, 71 of them in the human profileSource tree
  1. The human profile is reduced. The reduced human profile exposes 71 of those 102 tools. A person who connects from a desktop client can adjudicate work and cannot execute it.

  2. Build Station. Build Station adds a further 38,351 lines, and 16,119 of them sit in its native core.

  3. Build or buy. If you are weighing building against buying, we would like to show you the inventory and let it make the argument.

How this was measured

Commands produced these figures, and we dated them on purpose.

We produced the board and release figures on 2026-11-18 by running commands against the repository and against the live instance’s own API, and we read the workspace totals from its dashboard the same day. We measured the codebase inventory on 2026-08-20 against v3.8.5. None of it is estimated. Where we show a value in short form, such as a token count in billions, the exact figure stands next to it.

Source tree
Line, table, tool and tag counts
Live instance API
Tasks, work time, tokens, cost
Measured
2026-11-18Codebase inventory 2026-08-20
Against
v3.13.13
  1. The date is published. These figures move every week, and a figure without a date is a figure you cannot check.

  2. The version is published. We measured against v3.13.13, and when we measure again, the date changes along with the numbers.

  3. Reproducible. If a number here matters to your evaluation, ask us and we will show you the command that produced it.

What we do not claim

We list the missing numbers ourselves.

This site has no customer count, no revenue figure, no retention rate, no time-saved percentage and no benchmark against a competitor. We leave them out because none of them can be derived from the system, and a plausible figure we made up would undermine every measured one above it. Engineering numbers are what we have, and they have the advantage of being true.

FigureStatus
Customer countWe have not measured it and do not claim it.
RevenueWe have not measured it and do not claim it.
RetentionWe have not measured it and do not claim it.
Time saved, as a percentageWe have not measured it and do not claim it.
Any comparative benchmarkWe have not measured it and do not claim it.

None of these comes out of the system, so none of them appears on this site.

  1. No improvised numbers. When a claim needs a commercial answer, we will tell you "we will come back to you", and we will never give you an improvised number.

  2. No untested comparisons. Comparisons on this site are made against how teams work today, and none of them names a product we have not tested.

  3. The same rule elsewhere. The same rule governs the Trust Center, which states what is confirmed and what is still in progress.

Contact

Ask us for the commands.

Every figure above came out of the product and describes the product. If you want to see the board it came from, or the query behind a specific number, we will give you that demo, so bring the number you doubt most.

One level deeper

You can check the security posture yourself.

The Trust Center covers deployment boundaries, security practices, and what is confirmed and what is still in progress.

Open the Trust Center