Technical

Cambriyn is a self-contained, Rust-native agent harness, work-management system, and model-efficiency stack.

Inside Mission Control

The whole fleet works on its own, managed in Cambriyn’s Mission Control like running Asana or Linear rather than a chatbot. Only the decisions that need you reach the Desk, each with a recommendation ready.

Plans that run

A goal stays live until the cards under it are done. You see the progress, what done means, and which card is waiting on you.

Cambriyn goals, Active: live goals with their progress and their cards.

One board of everything in motion

The team’s work lives on one board, and you are the decider: Scheduled, Running, In review, Blocked, Done, and Failed.

Cambriyn work board with Scheduled, Running, In review, Blocked, Done, and Failed columns.

Agent teammates, not one session

Durable agent profiles created for specific jobs, built to fill roles and lead projects. Memory system built in.

Cambriyn agents as tiles, each showing its current work.

An agent beside every screen

Talk to any agent without leaving the work. Multiple threads can sit side by side as well.

Cambriyn chats, two agent threads side by side.

Built for companies that want their own AI workforce

Anyone whose frontier-model bill grows faster than the work it buys, organizations whose data cannot leave the building, and security teams that need defense running around the clock.

Teams who already live in Claude Code, Codex, Hermes, Cursor, or Devin and want a working team of agents without handing the stack to a host.

Cambriyn Mission Control, Desk / Needs you: the decisions waiting on you, with the open recommendation and a chat beside it.

The core daemon

The agent runtime and the work store, behind one API. The runtime gives every agent tools, a guard and a sandbox, and keeps the work moving.

The work store holds the board, the goals, the decision desk, and the activity log.

The guard

Dangerous commands are blocked, and tool output is scrubbed. Replies stream, an approval sits in the thread before the run continues, and any run can be cancelled.

The model broker

The broker sends each job to the model that fits it: local models on hardware you own, or the frontier models and subscriptions you already have. When one path is maxed, the call falls through to the next, so work fails over instead of vanishing.

Cambriyn model lanes: each backend with its state, and the map from role to lane.

Two kinds of model do the work

Support models are small and fast, and run on every device. Brain models do the thinking, on your own hardware or in the cloud.

Your agents

Every call goes through the broker.

Cambriyn broker

Picks the model for each job.

Background work goes to

On every device

Support models

Decisions and memory on one base model, made specialists by runtime kits. More roles to come.

Thinking goes to

On your hardware

Local brains

CyberFrost, Qwen3.8-27B, DeepSeek V4.1 Flash, GLM-5.3-Flash

or, when you run hybrid

In the cloud

Frontier brains

Claude Opus 5.5, ChatGPT 6.1 Sol

Agents send every call to the Cambriyn broker. The broker sends background work to support models on every device, and thinking to local brain models, or to frontier cloud models in hybrid mode.

Support models

A decision model makes routing, classification, and yes-or-no calls. Memory keeps what the fleet learns. A skill graph loads only the skills a task needs. This is the constant background work that would otherwise go to an expensive model.

Runtime intervention kits

Our own innovation. A kit is a small set of adjustments applied inside a base model while it runs, never merged into its weights, that makes the base a specialist. Today one base serves decisions and memory writing at once; more roles move onto kits next.

Brain models

The thinking goes to a bigger model: a local brain such as CyberFrost, Qwen3.8-27B, DeepSeek V4.1 Flash, or GLM-5.3-Flash, on Apple Silicon and other local hardware with no cloud; or, in hybrid mode, a frontier model such as Claude Opus 5.5 or ChatGPT 6.1 Sol.

Runs on the hardware you choose

Three ways, from the laptops your team already has to a machine built for Cambriyn. Or bring your own: we set it up on the hardware your firm owns.

Three ways to run Cambriyn
HardwareWhat runs on itSetup
Local inference hardwareMac Studios, DGX Sparks, GPU clusters, or a powerful workstation.Strong local brains and the support models.We set it up for you.
Cambriyn’s own machineOur own hardware, built for Cambriyn.Strong local brains and the support models.Arrives ready to run fully local.
Everyday laptops and desktopsThe Dell, Windows, and MacBook machines your team already has.Optimized support models on the device, with frontier models in the cloud.Hybrid.

A private cloud for the whole company

Put a handful of inference machines in your building and connect everyone’s everyday laptops to them. The cluster runs the brains, each laptop runs its own support models, and the work stays on your network.

Your building, your network

A handful of machines

Inference cluster

Mac Studio, DGX Spark, or Cambriyn’s own machine. Runs the brains and the support models.

Every desk

Everyday laptops and desktops

Dell, Windows, MacBook. The Cambriyn app and support models on the device.

Optional, when you run hybrid

Outside

Frontier cloud models

Inside your building, an inference cluster and everyday laptops form a private cloud. Frontier cloud models outside the building are optional.

Every figure we have measured

Timed or counted on our own Mac Studios, from each build’s test and benchmark logs. “Before” is Hermes Agent, the Python agent harness Cambriyn replaces, or our earlier build.

How much faster

BeforeCambriyn

  • Board reads23.2×Before: 3,627 µs. Cambriyn: 156 µs.
  • Safety check per command8.7×Before: 27.0 ms. Cambriyn: 3.1 ms.
  • Sandboxed file write2.9×Before: 25.1 ms. Cambriyn: 8.5 ms.
  • Mission Control cold start1.3×Before: 473 ms. Cambriyn: 350 ms.
How much less memory

BeforeCambriyn

  • Web tools server43.3×Before: 191.9 MiB. Cambriyn: 4.4 MiB.
  • Agent processes, all29.1×Before: 1,925 MB. Cambriyn: 66 MB.
  • Coding-session bridge10.9×Before: 134 MB. Cambriyn: 12.3 MB.
  • Agent runtime9.2×Before: 475 MB. Cambriyn: 51.6 MB.
  • Support models2.0×Before: 23.45 GiB. Cambriyn: 11.5 GiB.
Every figure we have measured
MeasuredResultBefore → after
Agent runtime memory, vs Hermes Agent9.2× less475 MB → 51.6 MB
Prompt tokens per agent call, vs Hermes Agent3.2× fewer24,121 → 7,425
Board reads, vs Hermes Agent23.2× faster3,627 µs → 156 µs
Coding-session bridge memory, vs Hermes Agent10.9× less134 MB → 12.3 MB
Agent processes on one machine, vs Hermes Agent93% fewer15 → 1
Memory of those processes29.1× less1,925 MB → 66 MB
Web tools server memory43.3× less191.9 MiB → 4.4 MiB
Safety check per command8.7× faster27.0 ms → 3.1 ms
Dangerous commands still blocked100%1,443 of 1,443
Sandboxed file write2.9× faster25.1 ms → 8.5 ms
Mission Control cold start26% faster473 ms → 350 ms
Skill listing per turn3.3× fewer tokens1,296 → 392
Support models memory50% less23.45 GiB → 11.5 GiB
Shared kit server vs two servers46% less memory16.76 GiB → 9.04 GiB
Core turn with one shell call5.6 msp50
Core daemon at idle6.8 MBmemory
Decision model call176 mstypical
Memory recall912 msp50

Cambriyn vs Hermes

Cambriyn gives you everything Hermes, OpenClaw, Grok Bot, Muse and Dots do, running on your own machines and tuned to be fast.

Measured head-to-head vs Hermes, on the same Mac Studios.

  • 23.2×faster board reads
  • 9.2×less memory
  • 1process instead of 15 (93% fewer)
  • 10.9×less memory for coding sessions
  • 3.2×fewer prompt tokens
Cambriyn against Hermes Agent

Hermes AgentCambriyn

  • Agent runtime memory9.2× lessHermes Agent: 475 MB. Cambriyn: 51.6 MB.
  • Prompt tokens per agent call3.2× fewerHermes Agent: 24,121. Cambriyn: 7,425.
  • Board reads23.2× fasterHermes Agent: 3.6 ms. Cambriyn: 0.16 ms.
  • Coding-session bridge memory10.9× lessHermes Agent: 134 MB. Cambriyn: 12.3 MB.
  • Agent processes on one machine93% fewerHermes Agent: 15. Cambriyn: 1.

Questions

How does Cambriyn compare to Hermes Agent?

Hermes Agent is the leading open-source agent harness. Here is Cambriyn against it, measured head-to-head on the same Mac Studios: 23.2× faster board reads (3.6 ms to 0.16 ms); 9.2× less agent runtime memory (475 MB to 51.6 MB); 93% fewer processes, 1 instead of 15; 10.9× less memory for coding sessions (134 MB to 12.3 MB); 3.2× fewer prompt tokens per agent call (24,121 to 7,425).

Are the speed figures only vs Hermes?

Yes. The measured speed and memory figures are head-to-head vs Hermes Agent, on the same Mac Studios. Hermes, OpenClaw, Grok Bot, Muse, and Dots are peers in the category; only the Hermes comparison is measured.

Does the agent team run in-house?

Yes. Cambriyn replaces agent platforms such as Hermes Agent and Factory: your agent team runs in-house, on hardware you own. The core of Cambriyn is open source, so your engineers can inspect every line.

The clients

Two thin clients share the API: Mission Control, the desktop app, and the cambriyn terminal client. The list opens any card in an inspector with its state, owner, goal, and note.

Cambriyn work list with one card open in the inspector: its state, owner, goal, and note.

Ops

Deploys, health checks, and alerts, with every agent’s config in one place.

Every layer, tuned for one job

Cambriyn isn’t a wrapper around someone else’s model. We tuned the whole stack, from the engine kernels to Mission Control, for one job: running your company’s AI on hardware you own.

  1. Mission ControlThe Desk, the board, and a chat beside every screen, in one app.
  2. The harnessA Rust agent runtime: tools, guard, sandbox, scheduler, and failover.
  3. Skills and memoryA skill graph that loads only what a task needs, and memory the whole fleet shares.

Software

  1. Support modelsDecisions and memory on one base model, made specialists by runtime intervention kits.
  2. Brain modelsOur own fine-tunes, such as CyberFrost, beside the best open-weight brains.
  3. QuantizationEvery model compressed to fit the machine it runs on.

Models

  1. Engine kernelsInference kernels tuned for the silicon underneath.
  2. HardwareMac Studio, DGX Spark, Cambriyn’s own machine, or the machines you already own.

Silicon

The Cambriyn stack, from Mission Control at the top to the hardware at the bottom, grouped into software, models, and silicon. Cambriyn tunes every layer.