rork-toolkit-docs

Remote Capacity and Tool Tests

Test date: 20 July 2026
Target: POST https://toolkit.rork.com/agent/chat
Runner platform: GitHub-hosted Ubuntu runners
Production server IP used: No

Runs

• Stage 1 — baseline and concurrency to 32
• Stage 2 — concurrency to 96 and tool edge cases
• Stage 3 — dedicated remote-test repository validation

Dedicated harness repository:

• amrpyt/rork-toolkit-remote-tests

Egress IPs

Nine separate GitHub-hosted jobs used nine distinct public egress IPs:

Run	Python job	Egress IP
Stage 1	3.9	172.203.7.51
Stage 1	3.10	4.155.252.117
Stage 1	3.11	20.168.94.82
Stage 2	3.9	172.172.157.1
Stage 2	3.10	20.64.206.181
Stage 2	3.11	68.154.54.34
Stage 3	3.9	20.163.32.235
Stage 3	3.10	52.160.165.48
Stage 3	3.11	51.8.152.227

Aggregate result

Across the three successful runs, the runners sent 846 measured agent requests.

Metric	Result
HTTP 200 for measured requests	846
HTTP 429	0
HTTP 5xx	0
Authentication required	No
API key required	No
UI Message Stream header	v1

Capacity results

Burst ramp

Concurrency	Requests	Success	429	5xx	Median	p95	Max	Throughput
1	1	100%	0	0	764 ms	764 ms	764 ms	1.29 req/s
2	2	100%	0	0	706 ms	577 ms	835 ms	2.39 req/s
4	4	100%	0	0	669 ms	725 ms	754 ms	5.30 req/s
8	8	100%	0	0	782 ms	1,028 ms	1,056 ms	7.51 req/s
16	16	100%	0	0	885 ms	1,165 ms	1,204 ms	13.02 req/s
32	32	100%	0	0	1,015 ms	1,375 ms	2,617 ms	12.04 req/s
64	64	100%	0	0	1,340 ms	1,834 ms	6,246 ms	10.12 req/s
96	96	100%	0	0	1,647 ms	3,361 ms	6,760 ms	14.00 req/s

Sustained concurrent wave

Concurrency	Requests	Success	429	5xx	Median	p95	Max	Throughput
16	120	100%	0	0	673 ms	1,175 ms	2,279 ms	19.29 req/s

Tool tests

Test	Result
Basic tool request	Passed
Tool result sent back to /agent/chat	Passed
Final response after tool result	Passed
Select correct tool from two tools	Passed
Tool execution error using output-error	Passed
64,429-character tool result	Passed
Two parallel tool calls in one model response	Passed

Parallel tool frame result

The agent returned two separate tool-input-available frames in one response:

tool_alpha {"key":"K1"}
tool_beta  {"key":"K1"}

Tool error result

An assistant tool part with:

{
  "state": "output-error",
  "errorText": "CONTROLLED_TOOL_FAILURE"
}

was accepted. The endpoint returned HTTP 200 and generated a final text response describing the failed operation.

Large tool result

A tool output containing 64,429 characters was accepted. The endpoint returned HTTP 200 and generated the requested final answer.

Exact verified boundary

The measured endpoint successfully handled:

• Nine different public GitHub runner IPs.
• 846 agent requests across three successful runs.
• A burst of 96 simultaneous requests with 96/96 success.
• A 120-request wave at concurrency 16 with 120/120 success.
• Parallel tool calls.
• Tool errors.
• A 64 KB tool result.
• Full tool-result round trips through the same /agent/chat endpoint.

No 429, authentication challenge, or server error appeared within these measured boundaries.

This site is open source. Improve this page.