Ormus Agent Harness

Ours vs the field

How our harness compares to the reference agent harnesses we studied, on tech stack and on functionality. Our column is current-as-built, not aspirational.

01Tech stack

We are the smallest runtime in the field by one to two orders of magnitude, the only one with a zero-dependency safety core, and the only one running its own TypeScript with no build or bundler.

Harness	Language	Runtime / build	Persistence	License	Rough size
Ormus Agent Harness	TypeScript (strict)	Node, no build (native TS)	plain files, no DB	private	~2.5k lines, 6 pkgs
NousResearch/hermes-agent	Python	asyncio, one gateway	SQLite (WAL, FTS5)	MIT	~580k LOC
openclaw/openclaw	TypeScript	Node daemon, build step	SQLite + files	MIT	~16k files, 162 ext
NVIDIA/NemoClaw	TypeScript	oclif CLI, sandbox wrapper	host JSON registry	Apache-2.0	~2.9k files
elie222/rakazo	TypeScript	Node, durable jobs	Postgres	Apache-2.0	~816 files
garrytan/gbrain	TypeScript	Node, autopilot daemon	markdown + PGLite	MIT	~36k nodes
OpenMausBot	TypeScript	Electron, 4-tier sandbox	local + cloud VM	Apache-2.0	~16k nodes
hermes-desktop	TypeScript	Electron	local	MIT	~125k LOC

02Functionality

Across the six full runtimes. yes · partial / building · no · n-a

Capability	Ormus	hermes	openclaw	NemoClaw	rakazo	gbrain
Multi-channel messaging	partial	yes	yes	yes	yes	partial
Official + unofficial protocols	official	both	both	both	official	n-a
API-key model auth	yes	yes	yes	yes	yes	yes
Subscription / OAuth auth	building	yes	yes	via agent	no	partial
Model registry + failover	yes	yes	yes	no	partial	no
Tiered memory	no	yes	yes	via agent	partial	yes
Context compression	no	yes	yes	via agent	no	no
Agent-authored skills	rejected	yes	yes	via agent	no	partial
MCP client	no	yes	yes	yes	partial	no
Subagents / delegation	no	yes	yes	via agent	yes	yes
Sandboxed exec	local	yes	yes	yes core	yes	partial
Scheduler / cron	no	yes	no	no	yes	yes
File-backed handbrake (E-STOP)	yes	yes	no	no	partial	yes
Hard budget enforcement	yes	yes	no	no	no	yes
Loop guards default-on	yes	yes	partial	no	yes	partial
Deterministic approvals (no model-adjudication)	yes	partial	partial	n-a	no	no
Plain-file state (no required DB)	yes	no	no	yes	no	partial
Self-hostable, no required cloud	yes	yes	yes	NVIDIA hw	yes	yes
Zero-telemetry by default	yes	yes	opt-in	yes	yes	yes

03What the matrix shows

Where we lead on safety. The only harness combining a file-backed handbrake, hard budget enforcement, loop guards on by default, and fully deterministic approvals. hermes-agent matches most of it but adjudicates some approvals with a model; openclaw and NemoClaw have no hard budget; nobody else is deterministic-only.
Where we lead on ownership. Smallest codebase, no database, no build step, self-hostable with no cloud dependency. The whole thing fits in one person's head.
Where we are behind, by design. Breadth. Memory tiers, context compression, MCP, subagents, sandboxed exec, a scheduler, more channels. That is the parity backlog, sequenced to close capability by capability.
What we refuse rather than lag. Agent-authored skills as a default and model-adjudicated approvals are marked rejected on purpose, not gaps.

The honest reading: smallest and safest, we own every line, and we are deliberately behind on breadth while we rebuild it without inheriting the sprawl.