ShipIt

Coding agents · LLM observability · design systems · security · macOS

We build the parts of the toolchain that usually get skipped: the agent that actually opens the pull request, the runtime underneath it, the trace that explains what an AI system did and what it cost, and the security pass that runs before the model and costs nothing.

Everything below is open source. Most of it runs on your keys and your infrastructure.

One runtime, many surfaces. The same agent core powers the GitHub teammate, the terminal, and the cloud — with observability wired through everything.

flowchart LR
    subgraph surfaces["Surfaces"]
        forge["🔨 Forge<br/>GitHub App · Action · CLI"]
        cli["⌨️ ShipIt CLI<br/>terminal agent"]
        cloud["☁️ ShipIt Cloud<br/>remote control plane"]
    end

    subgraph core["Core"]
        runtime["🤖 shipit-agent<br/>Python agent runtime<br/>tools · MCP · RAG · memory · permissions"]
    end

    subgraph quality["Quality & Security"]
        qa["🧪 ShipitQA<br/>autonomous browser testing"]
        sec["🛡️ Sentinel<br/>autonomous AI pentesting"]
    end

    obs["📊 Watchtower<br/>tracing · cost · PII masking"]

    forge --> runtime
    cli --> runtime
    cloud --> runtime
    runtime --> qa
    runtime --> sec
    runtime -.traces.-> obs
    forge -.traces.-> obs
    qa -.evidence.-> obs

pip install shipit-agent

Observability for LLM applications. One coherent record of what an AI system did: which prompt ran, what it cost, which tenant it belonged to, which tools it called, what it retrieved — and why it chose what it chose.

PII is masked before anything is persisted or leaves the process, because masking at display time is theatre once the raw value is on somebody else's infrastructure. Events fan out to Langfuse for analysis and to your own database as the system of record, with failures isolated per destination.

Python 3.11+ MIT · zero required dependencies · framework-agnostic

pip install shipit-watcher

Project	Version	Stars	Language	Last commit
forge
shipit_agent
shipit-watchtower
shipit-ui-design	—
ShipIt_Palette	—
snappilot
keyvault	—

The public repos are the foundation. On top of them, currently in private development and closed beta:

Bring your own key. Nothing here proxies your traffic through us or marks up a provider's price. You hold the credentials, you see the bill.

Deterministic where it can be. A model is good at judging whether something is reachable and worth worrying about. It is unreliable at checking four thousand lines the same way twice. So the checks that can be deterministic are, they run first, and they cost nothing.

Not crying wolf is the hard part. A scanner people stop reading is worse than no scanner. A trace nobody trusts is worse than no trace. Most of the work in these tools is in what they don't say.

Own your data. Traces land in your database. Usage lands in your SQLite file. Dashboards are account- or token-gated by default, never public.

ShipIt is built by Rahul Raj — a full-stack & AI engineer in Warsaw, Poland, who has shipped production systems across edtech, automotive, insurance, and health, and now spends his time on one question: how far can autonomous agents go before a human needs to step in — and how do you make that boundary safe, observable, and cheap?

Every tool here comes from running agents in production and hitting the same walls: agents that talk instead of committing, traces nobody can read, security scans nobody trusts, and bills nobody can attribute. ShipIt is the toolchain that came out of fixing those.