curl -sSL https://openfreedom.it/installer/install.php | bashcurl -sSL https://openfreedom.it/installer/install-mac.php | bashcoming soonNew official benchmark run (7 real tasks, same machine, cleared caches, vanilla installs, same model deepseek-v4-flash): 7/7 completed vs 4/7 for OpenClaw, with OF 3.5× faster and 4.6× cheaper ($0.0255 vs $0.1164). On the creation tasks (T4-T6) OF always delivers the requested files, OpenClaw doesn't; on the repair task (T7) both pass 14/14 tests, with the master file never touched. Full details in the PDF report.
The new version introduces DYNAMIC THINKING: the reasoning level is chosen for each task by the gate complexity score — trivial → light, complex → deep. On the official benchmark (7 real tasks, same machine, cleared caches, vanilla installations): 7/7 completed vs 4/7 for OpenClaw, plus a further −33% cost and −12% time compared to the same version with fixed high reasoning. Details in the full PDF report.
Current architectures re-inject ~17KB of context on every tool call, make 3-6 calls for a 2KB answer, and "memory" is an archive that accumulates without understanding. The most powerful machines on Earth spend their time managing their own weight.
3-6 LLM calls per answer: turn, tool-call, retry, sub-agent. Each one re-injects the entire context.
3 tools = 3 × 17KB of wasted context. At 100 tools: ~2M tokens for a task that needs 100K.
Linear archives with no connections: the right context never arrives at the right moment.
In OPENFREEDOM the model does only two things: understand and answer. Everything else (reading, writing, searching, executing) is done by deterministic code, with backups and verification. Two calls, not dozens.
Understands the request and produces a JSON step plan. It executes nothing: it plans.
one LLM call with reasoning off or lowDeterministic Python: runs the steps, filters results, makes backups, verifies every write. Fail-closed by design.
zero LLMReads only the filtered data and answers in natural language. If the data is not enough, it says so — it does not invent.
one LLM call with reasoning off or lowMeasured in the field: ~4,200 input tokens per answer vs 12,750–25,500 of classic architectures. Save up to 70% on analysis, bug and coding tasks.
At the beginning it knows nothing, like a child. But it already has web, files, mail, code and automations. And it learns: every conversation makes it stronger. Growth is factorial, not linear.
Write what you need in plain language. OF builds it. One call, not a pipeline.
Your files, your mail, your web, your language. It works where you live.
Start a heavy analysis, close the laptop. It notifies you when it is done.
Memories connect like synapses: what matters weighs more, noise falls away.
Linux, macOS, Windows. Integrity check and signature. It runs on your machine.
Missing a component? OF generates it on the spot, but with hard limits: the core (engine, flow, gate, vault) stays untouchable, self-assembly never touches critical files.
Every memory is a node connected to the others like a synapse: what matters weighs more, noise falls by itself. Memory weighs with a probabilistic equation with a chain model: the things most important to you will be important to openfreedom too. This solution makes our memory much faster and lighter than a conventional flat memory.
Skills are installed signed and validated by the gate: never downloaded from the internet, never executed without control. Missing a component? OF generates it on the spot, with hard limits — the core stays untouchable.
It is our invention, in preview — even the name was coined for it. Zill watches how you work, recognises the flows you repeat and turns them into reusable patterns: the more you use it, the faster you become.
It is not a promise in the terms of service: it is architecture. Security lives in the code, not in the prompts.
Every token and secret lives encrypted on disk. No exceptions, no plaintext files.
Derived from you and your machine. Copy the files to another computer? Without the password you read nothing.
The key never leaves: it is never written, never logged, never sent. No cloud watching.
No keychain, keyring or external KMS. Isolation is security.
Per installation. No universal payload, no replicable attack.
Every action passes through deterministic code that checks perimeter, backup and result. The model never has the last word on permissions.
By architecture: the prompt can ask, never order. Permissions are decided by the deterministic gate, not by the text. A hostile prompt has no tools to call.
Not upcoming features: already implemented, with agent-guided setup and secrets always in the safe.
Already implemented, not upcoming: the channel connects with tokens encrypted in the vault. Dual setup: manual, or direct through the agent that guides you step by step, without touching configuration files.
Closed system of 5 contacts: they exchange messages and files and receive the agent's answers with simple forwards, like in a messaging app, all over Tailscale. The workflow with the agent is never interrupted.
Configure as many sites as you want, each with its own credentials, always stored encrypted in the vault. One agent, many worlds.
Same for email: multiple configured accounts, credentials encrypted in the vault, reading and sending without ever exposing anything.
We repeat it on purpose: no password ever enters any OpenFreedom process in clear text. It is encrypted before passing through any chat or log.
Data from the latest official benchmark (27 August 2026): 7 real tests, OF V.9.3 vs OpenClaw 2026.7.1-2, same machine (all-in-one PC), same model (deepseek-v4-flash), cleared caches and vanilla installations. OF was tested in two configurations: fixed F4 high and F4 dynamic (DYNAMIC THINKING).
| Metric | OpenClaw 2026.7.1-2 | OF V.9.3 · F4 high | OF V.9.3 · F4 dynamic | dynamic vs high | dynamic vs OC |
|---|---|---|---|---|---|
| Total time (7 tests) | 1079 s | 249 s | 219 s | -12% | -80% |
| Total cost | $0.1161 | $0.0240 | $0.0160 | -33% | -86% |
| LLM calls | 58 | 15 | 15 | 0% | -74% |
| Tests passed | 4/7 | 7/7 | 7/7 | — | — |
OF completes the suite in both configurations (7/7, OpenClaw stops at 4/7 on creation tasks). DYNAMIC THINKING further cuts OF's own costs: −33% cost and −12% time at equal quality. Details in the full PDF report.
7 real tests, OF V.9.3 vs OpenClaw 2026.7.1-2: same prompts, same files, same machine (all-in-one PC), same LLM model, cleared caches and vanilla installations. Accuracy verified on disk with a test suite, not on the word of the answer. OF tested with fixed F4 high and with F4 dynamic (reasoning level chosen per-task by the gate complexity score).
| Test | OpenClaw 2026.7.1-2 | OF V.9.3 · F4 high | OF V.9.3 · F4 dynamic |
|---|---|---|---|
| T1 · Knowledge | ✅ | ✅ | ✅ |
| T2 · Maths | ✅ | ✅ | ✅ |
| T3 · Finance | ✅ | ✅ | ✅ |
| T4 · Web app | ❌ | ✅ | ✅ |
| T5 · E-commerce | ❌ | ✅ | ✅ |
| T6 · CRM | ❌ | ✅ | ✅ |
| T7 · 100 KB repair | ✅ 14/14 | ✅ 14/14 | ✅ 14/14 |
| Total | 4/7 | 7/7 | 7/7 |
OpenClaw fails all 3 creation tasks (T4–T6): 4/7 total. OF is the only one closing 7/7, in both configurations. Dynamic picks the right level per task (low on trivial, medium/high on complex): same results, lower costs.
Rates and methodology aligned with official sources: DeepSeek API pricing · Artificial Analysis. References in the PDFs.
📄 01/09/2026 report (PDF · EN) 📄 01/09/2026 report (PDF · IT)
No subscriptions, no hidden fees, no mandatory cloud.
€0
The complete product, no feature limits: 200 tasks to evaluate everything calmly.
€49.90 VAT included lifetime
🎉 Launch promotion: pay once, it's yours forever. Updates included, zero fees.
€49.90 VAT included
🎉 Launch promotion
📱 Scan the QR for your lifetime license
After payment you will receive your activation key by email.
Disclaimer. The user who uses OpenFreedom is responsible, under the law, for any damage that their artificial intelligence may cause to third parties. OpenFreedom has full shell powers on the device where it is installed: this implies the possibility of data loss or malfunctions. The software is used at your own risk; the owner is not liable for direct or indirect damages arising from its use, except in cases of intent or gross negligence. Privacy notice