Secrets, code and dependencies.
Scans the workspace for secrets and masks each value before the model sees it. Checks code against rules mapped to CWEs. Inventories dependencies into a CycloneDX summary, and looks them up on OSV if you turn that on.
Security for AI. A little more careful.
A desktop terminal with a security workbench beside it. Real shells, 9-CLI convergence and attack-reachability maps on one side. On the other, a taint-verifying agent, Shannon-entropy Codebase DNA and host-enforced approval gates.
In Agent mode, Marlowe reads files, edits them and runs commands in the project
folder you choose. Anything that overwrites or deletes a file, or runs a command
other than pwd or a plain ls, stops and waits. You see
the line diff or the full command, then approve it once or deny it.
An approval covers one action in one run and cannot be reused. The app decides what needs approval. The model's own account of what it did is never taken as proof.
note.txt +0 −1
Engineering is for ordinary code work: 24 model rounds and 64 actions a run. The security playbooks run longer, 48 rounds and 128 actions, and add the tools below. The built-in tools need nothing else installed.
Scans the workspace for secrets and masks each value before the model sees it. Checks code against rules mapped to CWEs. Inventories dependencies into a CycloneDX summary, and looks them up on OSV if you turn that on.
Inspects HTTP headers, TLS, DNS and up to 32 TCP ports. The app rejects any target outside the scope you set, and any name that resolves to a private or cloud-metadata address you have not authorized.
Hashes and inspects files up to 10 MB. Pulls out IP addresses, domains, hashes, CVEs and ATT&CK IDs. Flags anomalies in JSONL, HTTP access, syslog and auth logs.
Raw scanner output buries real bugs inside dozens of style lints and static UI false positives. Marlowe gives you three live maps of your project: a Codebase DNA constellation plotting real Shannon byte entropy (0–8 bits/byte), a 9-CLI Convergence graph linking installed scanners to workspace and scope assets, and an Attack Reachability map that separates exploitable taint paths from severed false positives.
When Semgrep, Trivy, Gitleaks, OSV-Scanner, Nmap, Nuclei, TShark, Radare2 and
macOS otool are on your verified login PATH, Toolchain Doctor lights them
up automatically — and reconciles nodes back to clean green as soon as a fix lands.
Marlowe scans your loopback ports for running Ollama, LM Studio, Jan and vLLM servers, enriches local models with parameter size and vision tags, and flags small models that risk false positives on deep multi-step taint audits.
Prefer a frontier model? Switch in one click to Claude, OpenAI, Gemini, xAI Grok, DeepSeek, Mistral, Groq, Together, Fireworks or OpenRouter. Keys stay encrypted in your OS keychain and are never forwarded across endpoints.
pwd and a plain ls
Waits. You see the full command, its working directory and its time limit.
Approved commands are not sandboxed. They run with your own permissions and can reach files outside the workspace, or the network. Review the command shown before you approve it.
Marlowe is built on Electron, so the same app can be packaged for macOS, Windows and Linux. macOS is the one that has been exercised so far.
Illustrations with example values, drawn from the app's own screens. The window frames are generic.
Enter your email and GitHub handle to receive a Read-only invite to zero-theory-labs/marlowe (v0.4 · PolyForm Noncommercial 1.0.0).