Sulcus 开放早期访问，可观测并干预运行中的 AI 智能体

Sulcus Early access

Run supported agent frameworks in Sulcus, connect Claude Code and Codex from your computer, or observe Google ADK and Microsoft Copilot Studio where they already run. See what your agents do, and step in where Sulcus has control.

Hosted Python runsClaude Code and CodexGoogle ADK and Copilot Studio

The problem

Once an agent is running, you need to see what it’s doing.

1. It loops

   It keeps calling the model or the same tool. Tokens climb; nothing moves forward.

2. It stalls

   A step hangs inside the workflow. From outside, it just looks busy.

3. It’s about to act

   A coding agent wants to run a command or change a file, and you’d rather decide than find out afterwards.

Observe and control

More than a trace.

Tracing tools record what an agent did. Sulcus can also act, and how much depends on where the agent runs.

Hosted by Sulcus

Sulcus runs your Python agent from a public Git repository.

Works with

LangGraph, CrewAI, OpenAI Agents SDK, Google ADK

Sulcus sees

Live. Framework steps, model and tool calls, reported tokens, application logs and failures.

You control

Start, stop and run again. Optional per-run token limit. Fixed CPU, memory and time limits.

Setup

In the browser. No Sulcus imports in your code.

Set it up

On your computer

Claude Code and Codex keep running where you started them.

Works with

Claude Code, Codex (experimental)

Sulcus sees

Each session appears as a run: turns, tool activity, timing and outcome.

You control

Approve or deny supported approval requests. Ending observation doesn’t stop the agent.

Setup

The Sulcus command-line tool and a paired computer.

Set it up

Deployed elsewhere

The agent keeps running where it already runs. Sulcus works from its telemetry.

Works with

Google ADK, Microsoft Copilot Studio

Sulcus sees

After the fact, not live: agents, tool activity, timing and errors. What’s included differs by integration.

You control

None. Sulcus observes; your service stays in charge.

Setup

Google ADK exports traces to Sulcus. Copilot Studio is connected in the browser.

Set it up

Integrations

Bring the agent you already have.

• Your framework still defines the agents and the workflow
• Hosted runs need no Sulcus imports in your code
• What Sulcus sees varies by integration; each page lists it

All integrations

LG

LangGraph

Graph and node steps, model calls and tool activity where LangChain callbacks expose them, with reported token usage.

CA

CrewAI

Crew, agent and task activity, plus model and tool events where CrewAI exposes them.

Sulcus records

• Run status, timing and outcome
• Which agents, steps, model calls and tool calls ran
• Reported token counts
• For hosted runs, what your application prints, up to a limit
• Failure details, with known credentials redacted

The integrations don’t collect

• Prompt text
• Model responses
• Tool inputs and outputs
• The API keys you add to a hosted run, which aren’t stored

Anything a hosted app prints, including a model response, becomes part of the run history. Redaction is best effort, so avoid printing sensitive data.

FAQ

Common questions.

Do I need to change my code?

For hosted runs, no: Sulcus instruments LangGraph, CrewAI, the OpenAI Agents SDK or Google ADK when your entrypoint runs, so you don’t add Sulcus imports. Claude Code and Codex need the Sulcus command-line tool on your computer. A deployed Google ADK service needs three environment variables, and Microsoft Copilot Studio is connected in the browser with nothing added to the agent.

What can I do right after signing up?

Verify your email, then connect Claude Code or Codex, a deployed Google ADK service or a Microsoft Copilot Studio environment. Hosted repository runs may be limited to selected workspaces during early access; the app tells you whether your workspace can start one, and may offer a built-in offline example where that is available.

Can Sulcus stop my agent?

A run hosted by Sulcus, yes: Stop run ends the run and its container. An agent on your computer, no: you can approve or deny supported actions, but ending observation leaves Claude Code or Codex running. A service deployed elsewhere is observed only.

Does Sulcus store my prompts or model outputs?

The framework integrations don’t collect prompt text, model responses or tool inputs and outputs. Hosted runs do keep what your application prints to stdout and stderr, up to a limit, so anything your code prints, including a model response, appears in the run history. Avoid printing sensitive data.

Is the token limit a hard cap?

No. It’s a guard based on the token counts your framework reports. Once reported usage reaches the limit, Sulcus blocks the next model call or ends the run, depending on the framework. A request already in flight can finish and exceed the limit, so it isn’t a ceiling on what your provider bills.

Who pays for model usage?

You do, through your own provider. Where a run needs a model, you supply your own provider API key, and that usage may be billed by your provider. Sulcus shows token counts, not costs.

Get started

See your first run in Sulcus.

Create an account, then connect an agent: a public Git repository Sulcus runs for you, Claude Code or Codex on your computer, or a Google ADK service or Microsoft Copilot Studio environment you already run.

Early access. Hosted repository runs may be limited to selected workspaces; the app shows what yours can run.