Teach your robot by typing.

The LLM harness and runtime for real robots — the agent writes each skill once, episodic memory sharpens it with every attempt, and it runs on-device forever.

Join the waitlist Book a demo

12–20 ms

Control loop, on-device.
The cloud never drives.

Teach

robot.localconnected · offline okSTOP

plan · review before run0/5

01Locate red parts on trayact_tiny · 8 mspolicy

02Move right arm above partmove_to · 20 Hzprimitive

03Close gripper, confirm graspgripper · −42°primitive

04Carry to left bin, releasemove_to · clampedprimitive

05Verify, log the episodeepisodes.jsonlmemory

✓ saved · sort_red_parts · your robot knows 15 tasks

Type it. Your robot learns it.

Describe the task in plain English. The agent plans it step by step, shows you the plan before anything moves, and saves the skill on the robot — where it runs at control rate, with or without internet. Every run is logged to episodic memory, so your robot gets better at your tasks.

Teach

A chat window served from the robot — type the task, review the plan, run.

Remember

On-device episodic memory: every attempt logged, every failure a lesson the agent recalls.

Run local

Skills execute as deterministic code on the robot, online or not.

Own it

Your skills, your data, your API key. Open source, BYO everything.

Teach

wipe down the bench, then stack the cups

Two skills, then. I already know wipe_surface— I’ll reuse it and author stack_cups. Plan is ready for review.

Teach your robot…Teach

Type the task. Review the plan. Run.

BotCortex serves a chat window from the robot itself — no code editor, no API, no robot programmer. The agent writes the skill once, and you see every step it plans before anything moves.

See it live

Remember

memory/episodes.jsonlmemory

14:02failgrasp slipped — cup wall thinner than expected
14:04faillesson: approach 8° steeper, close to −38°
14:06okstack_cups v2 — 5/5 grasps held

recalled at teach time → +35% success (RoboInspector, ACM TIST 2026)

Every failure makes the next attempt smarter.

Every attempt is written to on-device episodic memory — what ran, what broke, what the lesson was — and recalled the next time you teach. Feeding failures back like this improves manipulation success by up to 35%.

Read the research

Run local

cloud round-trip, P993–5 s

BotCortex control loop, on-device12–20 ms

network unplugged · skills still runningSTOP

primitivepolicyvlahumanthe LLM plans; these run it

No cloud in the control loop.

A control loop needs 12–20 milliseconds; cloud round-trips can spike to seconds. Taught skills run as deterministic code on the robot — dry-run by default, joint limits clamped, STOP always on screen.

Book a demo

Own it

~/.openhorizon/config.toml

[brain]
model = "claude-opus-4-8"     # your key

[hands]
vla = "pi0"                 # or groot, none

[robot]
platform = "openarm_v1"
skills_dir = "~/.openhorizon/skills"

Your skills, your data, your API key.

Skills and failure logs live on your hardware and belong to you — not to a shared library in someone else's cloud. Bring your own model, your own VLA backend, your own robot — free on a single machine.

Follow on GitHub