Benchmarks

The current benchmark evidence comes from the production broker harness for the qualified Context7 upstream. It supports a narrow resource-consolidation claim for identical credential-free, workspace-free contexts. It does not support a call-latency improvement claim.

Measured resource results

Three repetitions produced these medians and ranges:

Clients Mode Upstream instances OS processes Resident memory Startup to first tool list
1 Direct 1 (1-1) 3 (3-3) 252.1 MiB (252.0-252.5) 746 ms (699-891)
1 Broker 1 (1-1) 3 (3-3) 263.3 MiB (262.6-263.5) 3,007 ms (2,899-3,261)
5 Direct 5 (5-5) 15 (15-15) 1,258.5 MiB (1,258.5-1,260.5) 1,361 ms (1,338-1,460)
5 Broker 1 (1-1) 3 (3-3) 264.1 MiB (263.7-264.7) 3,777 ms (3,577-3,786)
20 Direct 20 (20-20) 60 (60-60) 5,042.0 MiB (5,039.1-5,044.5) 4,286 ms (4,244-4,953)
20 Broker 1 (1-1) 3 (3-3) 266.9 MiB (266.7-267.6) 3,789 ms (3,668-3,886)

At five clients, the broker reduced Context7 instances by 80% and resident memory by 79.0%. At 20 clients, it reduced instances by 95% and resident memory by 94.7%. One client had no process saving and used 4.4% more resident memory.

Every completed trial returned its Context7 process count to baseline, and every broker report reconciled with external process roots.

Latency limitation

The benchmark’s Context7 calls were upstream-throttled. Their recorded first-call and steady-state timings are not valid performance evidence.

An unthrottled 20-client diagnostic produced repeated 30-second timeouts and triggered configured degradation, so Irigate makes no high-concurrency latency claim from the current evidence.

Context limitation

The measured sharing case is Context7 with identical credentials and workspace assumptions. Context7 has neither client credentials nor workspace state, so no distinct-credential or distinct-workspace saving is claimed.

code-review-graph remains isolated because it retains context-bound state. shadcn remains isolated because no reviewed qualifier exists.

Compatibility snapshot

Client Result
Hermes Validated one direct Streamable HTTP broker call
Kilo/OpenCode Validated one direct Streamable HTTP broker call
Codex Validated one direct Streamable HTTP broker call
Claude Code Installed, but unavailable because the local CLI was not authenticated

Evidence gaps

The next evidence work is a quota-backed latency rerun and normal-session observation. The remaining questions are whether repeated unthrottled calls improve or degrade developer experience, whether realistic distinct contexts eliminate savings, whether authenticated Claude Code works directly, and whether the broker runs normally without orphan processes or state leakage.