ForgeSynapse
ClawLite
The personal AI assistant that finally respects the word “personal”
It runs entirely on your own machine and you talk to it from Telegram, so there is no new app to learn. Open source under Apache 2.0, no subscription, no telemetry, and a one-click installer that never asks you to open a terminal.
- v0.5.0
- Apache-2.0
- Python 3.11+
- Windows installer
- Local-first
The problem
“Personal” is doing a lot of work in that sentence
Most assistants that call themselves personal route every conversation through someone else’s servers. Your questions, your documents and your context become training data, retention policy and jurisdiction — three things you do not control and cannot audit.
The alternatives that genuinely respect privacy have a different problem: they are projects, not products. They expect you to install a runtime, run terminal commands and configure safeguards yourself. For anyone who is not a developer, that is not friction. It is a wall.
ClawLite is built on the opposite assumption: you should not have to understand security in order to be secure, and you should not have to open a terminal in order to own your own data.
What it is
A product, not a project to assemble
Not a cloud assistant
Inference runs locally through Ollama, with zero outbound traffic. Nothing leaves the machine unless you explicitly allow it.
Not a developer toolkit
One-click Windows installer. No terminal, no administrator rights, no dependency tree to resolve. It runs entirely in user space.
Not a subscription
Open source under Apache 2.0. There is no paid tier, no usage meter and no account. What you install is the whole thing.
A local-first assistant
It searches the web when it needs to, remembers context across conversations, and tells you plainly whenever it steps outside your machine.
How it is built
Security by default, not by configuration
- 01
Local-first inference
Runs on Ollama on your own hardware. If Ollama is unavailable and you have configured a cloud key, it can fall back — and it always says so. No silent transfers.
- 02
Human approval required
Any high-impact action — sending an email, running a command, modifying a file — needs explicit approval. This is the default and cannot be silently disabled.
- 03
Sandboxed execution
Code runs in an unprivileged container with no host access, no network egress unless permitted, and a read-only filesystem outside its work directory.
- 04
Prompt injection defence
External content — web pages, emails, documents — is treated as data, never as instructions. The boundary is architectural, not a best-effort filter.
- 05
Memory that fits on a laptop
A three-layer memory — working, episodic and semantic — over a local database, with an ONNX embedding engine of roughly 300 MB instead of the usual 2 GB.
Getting started
From download to first message
- 01
Download
Take the installer from the latest release on GitHub. No account and no administrator rights required.
- 02
Run the wizard
A setup page opens in your browser and asks for two keys: a Telegram bot token from @BotFather, and a free Tavily key for web search.
- 03
It starts itself
Once setup finishes, ClawLite runs in the background. There is nothing else to configure.
- 04
Talk to it
Open Telegram and write. It searches when it needs to, remembers what you told it, and asks before doing anything consequential.
Manual install
Or do it the developer way
The installer is the recommended route, but nothing is hidden behind it. On any operating system with Python 3.11 or later and Ollama running, the manual path is three commands.
git clone https://github.com/forgesynapseltd/ClawLite.git
cd ClawLite
pip install -r requirements.txtRequirements: Python 3.11+, Ollama, a Telegram account and a free Tavily key.
What it runs on
- OllamaLocal inference
- TelegramInterface
- TavilyReal-time web search
- SQLiteLocal memory
- DockerExecution sandbox
Who it is for
For people who want the assistant, not the assembly
People who are not developers
The whole product is designed on the assumption that the user should never need to open a terminal. That is the clearest difference between ClawLite and everything else in this space.
Anyone under a duty of confidentiality
Lawyers, clinicians, accountants and consultants handling material that cannot be pasted into someone else’s cloud.
Sovereignty-minded organisations
Teams that need to state, with evidence, where their data went — and to know that the answer is “nowhere”.
Modest hardware
A 300 MB embedding engine instead of 2 GB means local memory works on an ordinary laptop, without dedicated GPU infrastructure.
Limitations
Where we are behind, honestly
ClawLite is newer and smaller than the dominant project in this space. Those gaps are real and material, and pretending otherwise would undermine the advantages that are real.
- The community ecosystem is smaller. The leading project has thousands of integrations already proven in production. If you need one specific niche integration today, that project may be the more practical choice.
- It is not yet battle-tested at enterprise volume. The isolation and governance model is solid by design, but design is not the same as a production track record at scale.
- It has no foundation behind it. The leading project carries institutional weight that takes years to build, and we do not have the equivalent yet.
- It needs Ollama running on your machine, and the quality of the answers depends on the local model you choose. Local-first is a trade: you gain privacy and give up the largest frontier models.
- The cloud fallback is a real cloud call. It is optional, it is off unless you configure a key, and it always announces itself — but if it triggers, that request leaves your machine.
Download it and check the claims yourself
The source is public and the installer is a single file. Nothing on this page has to be taken on trust — that is rather the point of the product.