for agents: this page is markdown at https://switchyard.so/docs/limits.md; every page is listed in llms.txt.
Limits
This page lists Switchyard's ceilings, each from the constant that enforces it, and the most load that a named run measured.
You will have: the number for each ceiling, and the measured load that each number comes from. Works with: the board, yard, MCP, REST. The ceilings are the same on every surface.
A limit keeps one project from slowing other projects down. If a limit stops real work, tell us (see the last section).
Most seen in load tests
Each figure names its run. Every figure here comes from a run of at most 100 agents. We publish a figure past 100 agents once a run at that size passes.
| What | Most measured so far |
|---|---|
| Tasks in one run | 1,022 tasks over four projects in 30 minutes, all merged (100 agents, 1 November 2026, run 41357785721) |
| Agents at once on one project | 100 agents that submitted at once: 301 pushes, each seen at its head, and all 100 tasks merged in 323 seconds (1 November 2026, run 41335178062) |
| Repository with 100 agents | 6 GB: 171 tasks merged, 4.2 minutes from submit to merged at the median (run 41357785721) |
Tasks and agents
| What | Limit |
|---|---|
| Open tasks in a project | 100 on every plan. We can raise it for an organization. Past the limit, a new task whose create carries a key waits in line: yard and the MCP server send one, and over the API it is the Idempotency-Key header or a requestId. A create without a key is refused with 503 and Retry-After. When the project is overloaded, it lets in at most 50 open tasks until it recovers |
| A task that is never used | abandoned after 4 hours with no push, no plan, no claim, no stacked task and no Git key |
| An idle task | abandoned after 14 days with no activity. A submitted task is not abandoned |
| Subagent tasks that one task can start | 20, finished tasks included |
| Tags on a task | 16 |
waitMinutes on yard_submit | 1 to 10. One call waits at most 25 seconds; call again to wait longer |
| Requests for one key | 600 a minute. For an agent group key, 600 a minute for each MCP session |
| Failed requests with one key from one address | 120 a minute. For the next 60 seconds that key's failed requests from that address are refused with 429; its requests that succeed are still served, so agents sharing a group key are not refused for each other's failures |
| Failed requests from one address | 600 a minute, with any keys. For the next 60 seconds the address is refused unless a request carries a key, and at most 600 a minute that carry one are checked |
Checks and the merge queue
| What | Limit |
|---|---|
| Tasks tested together | 8. The number doubles after 4 batches that pass, up to 32, and halves (not below 8) after a failure |
| A check's run time | 10 minutes, unless the rules set timeout. At most 1 hour |
| Check machines | small: 1 vCPU, 6 GiB of memory, 12 GB of disk. large: 4 vCPU, 12 GiB, 20 GB. A checkout can use the disk less 2 GB |
| A check whose rules name no size | runs large when its checkout needs more than 10 GB. On a repository of 1 GB or more whose checkout cannot be measured, it runs on the size it last ran on, else small. A small run that runs out of disk runs once more on large, billed as small |
| Check machines that run at once, for each organization | 4 on Launch, 20 on Scale, 50 on Enterprise. A check past the limit waits for a machine, for up to 6 hours |
| Services that a check starts | each service must answer within 60 seconds of its start |
| Test reports read from one check | 50 files, 4 MiB each, 8 MiB in total. The check's record keeps 200 tests in 16,000 bytes, failed tests first |
| Failed tests that a second run runs again on their own | 100. With more tests, or a rerun command over 16,000 characters, the whole check runs again |
| Paths that one check checks out | 16 files or folders (checkout in the rules). Without checkout, a check gets the whole tree |
| Checks that can start from a saved disk | only a check that takes the whole tree: no checkout, no git = "full" and no services. Every other check uses a fresh setup. Off until Switchyard turns it on for a project |
| A saved disk after its project is deleted, or after Switchyard turns saved disks off for the project | never used again. It cannot be deleted before it expires, and it can take up to 30 days to expire at our hosting provider |
| Where a saved disk is stored | with our hosting provider. Its storage region cannot be chosen |
| Checks drafted at setup | 20 |
| Passes that a project keeps | 20,000 |
| Merging paused | at most 30 days in one pause |
Conflicts and AI suggestions
| What | Limit |
|---|---|
| Tasks compared for clashes on one conflicted file | every pair up to 32 tasks. From 33 to 512 tasks, clashes are found by halving. Above 512 tasks, every task on the file is named |
| A remembered conflict fix | 256,000 characters for each file |
| Remembered fixes sent with each request | 2,000,000 characters, most recently used first |
| AI suggestions on Switchyard's key | 100 a month for each organization, and $150 per credit cycle in total over every organization, each cycle running from the 11th of one month to the 11th of the next. Past either, the conflict goes to a person |
| AI suggestions that one project asks for in a day | 50 |
Repositories and Git
| What | Limit |
|---|---|
| A repository when you connect it | 8 GB as GitHub counts it. Setup refuses a larger repository. GitHub counts every branch, so its size is often more than what Switchyard copies |
| Repository size by plan | 3 GB on Launch, 8 GB on Scale and Enterprise. The pricing page has the plans |
| The first copy | main comes first, at its last commit only, and tasks can start when it is in. The history copies after it in steps of 2 GB. A pack over 4 GB is split. A pack that is still over 5 GB after the split cannot be kept |
| Disk for the history copy | twice the history left to copy, plus 0.5 GB. A project's engine moves to a larger disk when necessary, up to 20 GB. A history that 20 GB cannot hold waits |
| Extra engines that serve clones and reads | up to 8 for one project: 2 for a repository under 1 GB, 4 under 2 GB, 8 from 2 GB. One more is added each minute while the project's clones and fetches wait, up to that number, and one is stopped after each 30 minutes in which none is used (about a third busy, or with a clone waiting more than 5 seconds). From 50 open tasks, one stays while the project is in use |
| An API request body | 1,000,000 bytes |
Accounts and records
| What | Limit |
|---|---|
| Sign-in sessions | 30 days, renewed while you use them |
A device code for yard login | 15 minutes |
| A sign-in link sent by email | 10 minutes |
| Emails to one address (sign-in links and welcomes) | 10 a day |
| Single sign-on connections | 2 for each organization, each for at most 5 email domains |
| A single sign-on DNS proof value | 7 days, then the page shows a new one |
| SCIM users in one page of a list | 100 |
| A task's Git key | 24 hours, or 10 minutes for an agent signed in over OAuth |
| Audit log | 400 days, exportable as CSV or JSON lines |
| Project deletion | scheduled 7 days ahead, and you can cancel it until then. After the deletion, the project's record and Git data are kept 30 days more, then removed (for a project on our older Git host, its record and refs are kept, not its Git objects). Within those 30 days, write to us and we will try to bring it back |
Raising a limit
Write to hello@switchyard.so with the project id, the limit that stops you, and what you are trying to do.