v0.2.0-pilot.17 · Apache-2.0 · self-hosted
Your backend sends a message and Dispach runs the agent: memory, tools, schedules, channels and approvals, on any OpenAI-compatible model, on infrastructure you control.
› npm install --global dispach
Dispach matches or beats Hermes on 8 of 11 models, and costs the least on 10. OpenClaw scores lowest on every one.
Tasks passed out of 125: 25 multi-turn tool tasks, run 5 times each, graded on what the tools were left holding. Every harness runs in its own container against the same tools and pinned model providers. Cost is the OpenRouter bill for all 125 runs.
| Model | Dispach 4f3e00a | Hermes 0.19.0 | OpenClaw 2026.10.5 |
|---|---|---|---|
| Claude Opus 5.5 | 125$2.05 | 119$2.24 | 105$7.10 |
| Claude Sonnet 5.5 | 125$0.80 | 120$1.76 | 105$4.21 |
| Claude Haiku 5.5 | 124$0.06 | 120$0.07 | 102$0.21 |
| GPT-6.1 Sol | 125$0.54 | 120$0.53 | 95$1.73 |
| Gemini 4 Argon | 119$2.29 | 123$2.79 | 108$5.47 |
| GLM 5.3 | 125$0.80 | 123$1.18 | 98$4.86 |
| DeepSeek V4 Pro | 125$0.31 | 120$0.50 | 96$2.83 |
| Kimi K3 | 115$2.24 | 115$2.36 | 98$10.27 |
| Gemma 4 26B-A4B | 112$0.24 | 108$0.41 | 91$2.95 |
| qwen3.5-9b | 104$0.18 | 112$0.42 | 60$0.78 |
| gpt-oss-20b | 97$0.05 | 111$0.10 | 51$0.12 |
One benchmark of our own design, each harness on its own documented defaults. With 125 runs an arm a gap of a few runs is noise. Method, intervals and where Dispach loses.
5 of 11
On 25 multi-turn tool tasks run 5 times each, Dispach matches or beats Hermes on 8 of 11 models and costs the least of three harnesses on 10.
One benchmark of our own design, each harness on its own defaults. Hermes leads on the two small open models.
Benchmarks →202
A turn is detached from the request that started it. Drop the connection, reattach from any client, and the stream replays from the start.
The recording above is that, unedited apart from shortened request bodies.
Turn lifecycle →91 ms
No network I/O before readiness, by rule. Channels and remote catalogues connect afterwards and report status as events.
Median of 7 boots, Apple M4 Pro. Boot, not response latency. CI fails above 1,200 ms.
Evidence →178 MiB
One runtime hosts many agents, and scoped keys fence each off. Adopt and remove agents live, with no restart.
Idle resident memory, linux/arm64 image, one machine.
Evidence →Every number on this site has a row in the claim register, negative results included.
experimental · where it's going
The multiplayer agent runtime: an isolated runtime for every user and team in your product, with agents that share rooms and memory with the people they work for, and stand in for them, openly, when they are away.
In 0.2.0-pilot.17 it is on the HTTP API only, with no screens yet, and it may change within 0.2.x. Everything above is stable.