Sign-in
Sulcus uses Firebase Authentication for Google, GitHub and email sign-in, requires a verified email address, and verifies identity on the server for every API request.
How Sulcus controls access, runs hosted code, handles API keys and limits what it collects, and where the current limits are. What applies depends on where your agent runs.
Every project, run and live stream is checked against your workspace.
Each hosted run gets its own container with fixed CPU, memory and time.
Supplied when a hosted run starts and not stored with it.
Integrations record what ran, not prompt or response text. Hosted logs are the exception.
Sulcus uses Firebase Authentication for Google, GitHub and email sign-in, requires a verified email address, and verifies identity on the server for every API request.
Projects and runs belong to a workspace. Reading a project, starting or stopping a run, deciding an approval and opening a live event stream all require workspace membership, checked on the server. Unknown and unauthorized IDs return the same not-found response.
There is no team invitation or member management in the app, and no configurable per-run permissions. Workspace members currently have the same run permissions.
Hosted runs
Each hosted run executes in its own container as an unprivileged user, with fixed CPU, memory, process and time limits, a read-only root filesystem, dropped Linux capabilities and a temporary working directory that is removed with the run.
Sulcus’s production configuration requires hosted repository runs to use the gVisor container runtime with fixed DNS resolvers, and checks the runner before accepting runs. Runs can reach the public internet so your app can call model providers. Private network ranges, cloud metadata addresses and other runs aren’t reachable.
The API accepts a validated public HTTPS repository, Git reference, Python entrypoint and framework. It doesn’t accept arbitrary container images, commands, host paths, flags or network names.
These layers reduce what a run can reach. They aren’t a guarantee of isolation, and Sulcus makes no certification claim about them. Installing dependencies runs third-party build code, and your own code can send data wherever the network allows.
Secrets in hosted runs
You add API keys and other secrets when you start a run, and your app reads them as environment variables. They aren’t saved with the project, the run record or its events.
Runtime events, application output and failure diagnostics are redacted for the values you supply and for common credential patterns. Redaction can’t catch everything, and it can’t stop your own code from sending a secret somewhere else.
Sulcus doesn’t offer saved project secrets or rotation. Enter API keys again for each run. Credentials Sulcus itself issues or needs for a connection, such as a paired computer’s credential, a Google ADK ingestion token or a Microsoft Copilot Studio client secret, are handled separately and described below.
Data collected from hosted runs
Event history is kept for the most recent 5,000 events per run.
Application output is the exception. If your code prints a prompt, a model response or personal data, that text is stored with the run. Redaction only targets credentials, so avoid printing sensitive data.
On your computer
By default the Sulcus command-line tool sends activity metadata: session, turn and tool identifiers, agent and tool names, model name, status, error type, timing, exit codes, change counts and token counts. It also sends the project folder’s name, a hash of its path and its Git remote address without credentials, so runs can be grouped by project.
Prompts, model responses, command output, file contents, file diffs and your environment aren’t sent. Unrecognized fields are dropped on your computer before upload, and the server filters events again before storing them. Command text and file paths are sent only with an approval request, at the decision-context level below.
A computer that is new to Sulcus starts at decision context: when an agent asks for approval, the request carries the command (up to 2,000 characters), the working directory, up to 10 affected file paths and, for a network request, the hostname. It’s kept on the approval, shown to the workspace members who can decide it, and not added to the run’s events. Common credential shapes are removed on a best-effort basis, so a command or path can still be sensitive. At the private level, used by computers set up before sharing levels existed and by sulcus login --private, an approval shows only which kind of action is waiting. You change the level on the computer with sulcus device privacy; Sulcus can’t change it.
You link the tool to your account by entering a short code in the app. The credential is kept in your operating system’s credential store. sulcus logout revokes the computer on the server and deletes the stored credential.
Sulcus can pass your approve or deny decision to Claude Code or Codex for supported requests. It records the tool name, the decision and when it was made, plus the decision context above when the computer shares it. It never approves by default, and both agents keep their own permission and sandbox settings.
Stop at next action refuses Claude Code’s next tool call, or an idle session’s next prompt. Interrupt turn ends a Codex turn, and ends the Codex process if the turn hasn’t ended 5 seconds later. Sulcus sends the computer no command: the computer picks up a pending stop when it next reports. Stopping is best effort and fails open. It needs the computer online, the Sulcus companion or launcher running and, for Claude Code, current hooks. A stop that isn’t picked up within 120 seconds expires and the agent keeps running, and a Claude Code hook that times out or fails lets Claude carry on. Sulcus records who asked for the stop, when, and what ended the session. Cursor can’t be stopped from Sulcus.
Where this is available, a paired computer can accept requests from Sulcus to start Claude Code or Codex in a folder it has already reported. This is off unless you enable it on that computer, and it can ask for confirmation there each time. The task you type in Sulcus is relayed to the computer and kept only until the agent starts or the request ends. Claude Code started this way can be stopped with Stop process, which sends SIGINT, then SIGTERM, then SIGKILL; Codex with Interrupt turn.
Sulcus hooks in Cursor report finished tool calls and turns: each tool’s name, status and duration, a turn’s token counts and model, and the same project grouping as above. The prompt, commands and their output, file paths, file contents and edits are never sent. Cursor sends Sulcus no approval requests, so no decision context leaves the computer, and Sulcus can’t stop Cursor.
Deployed elsewhere
From ADK’s OpenTelemetry spans: agent, tool and model names; span, trace, tool-call and session IDs; start and end times; error status and type; input and output token counts.
Prompts, responses, tool arguments and results, span events, status messages and user IDs on ADK spans are discarded on arrival, before anything is stored. They may still be transmitted to Sulcus first. Setting ADK_CAPTURE_MESSAGE_CONTENT_IN_SPANS=false turns off ADK’s message-content span attributes at the source; identifiers, status and error information, and exception details may still be exported, depending on your ADK version and configuration.
Each connected project has a token your service sends with its telemetry. Sulcus shows it once and stores only a hash. You can replace or revoke it; revoking stops ingestion and keeps existing runs.
Sulcus queries your Azure Application Insights resource for agent, tool and reply records from your environment, and requests only identifiers, timestamps, duration, outcome, error type, and agent, tool and model names. Messages, tool inputs and results, and user identity aren’t requested, so they aren’t sent to Sulcus.
This connection needs a saved credential: the client secret of a Microsoft Entra app registration you create and give the Reader role on the Application Insights resource. Sulcus stores it encrypted, tied to that project, and never shows it again. Disconnecting deletes it. This is different from the API keys of a hosted run, which aren’t stored.
Email support@sulcus.dev with details and steps to reproduce.
This page describes how Sulcus works today. It isn’t a certification or a contractual guarantee.