Observe
See what the agent is doing, live.
- Workflow and agent steps as your framework runs them
- Model calls and the tokens each one reports
- Tool calls, with start, completion and failure
- Logs, categorized failures and a history of every run
Sulcus Cloud Early access
See your AI agent run. Stay in control when it matters.
Run your LangGraph, CrewAI or OpenAI Agents SDK app in Sulcus Cloud. Watch model calls, tool calls, token usage and failures as they happen — and set a token limit or stop the run at any time.
A live Sulcus run: 38,420 of a 50,000 token limit used, recent model and tool activity, and a Stop run action.
The problem
Once an agent is running, you need to see what it’s doing.
- It loops
It keeps calling the model or the same tool. Tokens climb; nothing moves forward.
- It stalls
A step hangs inside the workflow. From outside, it just looks busy.
- It fails halfway
Step four of seven throws, and piecing together steps one to three means digging through logs.
Inside Sulcus Cloud
One screen for the whole run.
Run Details updates live while the agent works, and stays available after it finishes.
Sulcus Cloud Run Details for a running LangGraph workflow: 14 AI calls, 9 tool calls, 1 minute 12 seconds of runtime, 38,420 of a 50,000 token limit used, no issues, a workflow tree of graph steps, a live activity feed, and a Stop run action.
- Workflow
- The steps your framework ran — graph nodes, crew tasks or agents.
- Activity
- Model calls, tool calls, errors and system events, in order.
- AI usage
- Reported tokens, measured against your per-run limit.
- Stop run
- Ends the run and its container from the dashboard.
Observe and control
More than a trace.
Most tracing tools record what an agent did. Sulcus Cloud also runs the agent, so it can stop the run and hold it to limits.
Control
Step in while it runs.
- Stop a Cloud run at any time
- Set a per-run token limit that blocks the next model call
- Fixed CPU, memory and time limits on every run
The token limit uses reported usage, so an in-flight request can finish first.
When it helps
Three moments Sulcus is built for.
Token usage keeps climbing.
See which calls are repeating as usage nears your limit. Stop the run, or let the limit end it.
It broke halfway through.
The workflow tree shows where the run got to and where time went. Logs and a categorized failure show what led up to it.
Yesterday’s run, today’s code.
The repository has moved on. The run still keeps the exact commit and configuration it started with.
Framework support
Bring the agent app you already have.
- No Sulcus imports in your code
- Sulcus instruments the framework’s entry points automatically
- Your framework still defines the agents and the workflow
LGCAOA
LangGraph
Graph and node steps, model calls and tool activity, with token usage and an optional per-run limit.
CrewAI
Crew, agent and task activity, plus model and tool events where CrewAI exposes them.
OpenAI Agents SDK
Agent lifecycle, handoffs, model and tool calls and guardrail events for SDK 0.22.x runs.
What you see varies by framework. Each integration page lists the details.
How it works
From repository to live run in three steps.
- 01
Connect the repository
Choose a public HTTPS Git repo, a branch, tag or commit, your Python entrypoint and the framework it uses.
- 02
Start the run
Add your API keys and an optional token limit. Sulcus starts a resource-limited container and instruments your framework.
- 03
Watch and step in
Follow the run live in Run Details, and stop it whenever you need to.
Developer Preview
Codex and Claude Code, visible in Sulcus Cloud.
Launch Codex or Claude Code locally through the Sulcus CLI and follow the session in the same Runs view. Prompts, command output and file diffs stay on your machine.
- Observation only: Sulcus Cloud can’t stop the local agent or approve or deny its actions.
- Early and changing. Not intended for production use.
Privacy
See how the agent behaves, not what it says.
Sulcus records
- Run status, timing and outcome
- Workflow steps and their structure
- Which model and tool calls ran, and how long they took
- Reported token counts
- Logs and failure details, with credentials redacted
- The resolved commit and run configuration
Sulcus doesn’t capture
- Your prompts
- Model outputs
- Tool inputs and outputs
- Runtime secrets. You supply them per run, and they aren’t stored.
Redaction of logs and errors is best effort.
Built for today
Where Sulcus fits right now.
A strong fit if you’re
- Building with LangGraph, CrewAI or the OpenAI Agents SDK
- Working in Python, with the code in a public Git repository
- Running trusted development workloads you want to watch
- Debugging agent behavior and keeping token usage in check
Current boundaries
- Public HTTPS repositories only, no private repos yet
- Python only
- Trusted code only: the container limits resources, but it isn’t a sandbox for hostile code
- Cloud runs have fixed platform CPU, memory and time limits
- Codex and Claude Code support is a Developer Preview
Status
- Available now
- Cloud Git runs for LangGraph, CrewAI, OpenAI Agents SDK
- Developer Preview
- Local Codex and Claude Code observation
FAQ
Common questions.
Do I need to change my code?
For LangGraph, CrewAI and OpenAI Agents SDK apps, no. Sulcus instruments the framework automatically when your entrypoint runs, so you don’t add Sulcus imports.
Where does my code run?
In a resource-limited Docker container operated by Sulcus. Sulcus clones your repository, resolves the exact commit, installs dependencies and runs your Python entrypoint there.
Can I use a private repository?
Not yet. Sulcus Cloud currently runs public HTTPS Git repositories only.
Does Sulcus store my prompts or model outputs?
No. The framework instrumentation deliberately doesn’t capture prompts, model outputs or tool inputs and outputs. It records which calls ran, how long they took and how many tokens they reported.
Can Sulcus stop my agent?
Runs in Sulcus Cloud, yes: stop a run at any time from Run Details. Local Codex and Claude Code sessions in the Developer Preview, no: Sulcus Cloud only observes those.
Is the token limit a hard cap?
It’s a guard, not a billing ceiling. Once reported usage reaches the limit, Sulcus blocks the next model call and ends the run. A request already in flight can finish and push usage slightly over.
Does Sulcus cost anything?
Sulcus is free to use during early access. Pricing will be introduced later. Model usage is billed by your AI provider, through the API keys you supply.