Feature map
Marshal does three jobs for your team:
- Work alongside an agent. Open a space, pair with Claude Code in your browser, and stop when you are done. Your files wait for you.
- Delegate. Hand a task to a run and get a result back, or let a deputy pick up the same task every morning or on every new GitHub event.
- Govern. Decide what each space can reach, keep credentials away from agents, and see who did what in the audit log.
Each feature below has a line on when to reach for it. Runs, deputies and webhooks are available when your administrator enables them. To see the features work together, see what teams build with Marshal.
Work in a space
| Feature | Use it when |
|---|---|
| Spaces and storage | You want a private environment you can create, connect to, stop, resume and fork, with your files saved between uses. |
| Compute sizes | A build or test suite needs more CPU, memory or disk, even in a space that is already running. |
| Docker in spaces | The agent builds images, runs containers and brings up compose stacks, inside the space’s network policy. |
| Billed compute cost | You want to see what a space or a run costs in CPU, memory and disk, next to its LLM cost, priced from your AWS or Google Cloud bill. |
| Memory | You are tired of repeating how you work. Write it once, and agents also keep notes on what they learn. |
| Retention and archive | You want old spaces out of the way, with a set time to restore them before cleanup. |
Let agents work without you
| Feature | Use it when |
|---|---|
| Start and track runs | You want to start a run from the web, the CLI or a deputy, and see every run in one list. |
| The run page | You want to follow a run live, answer its questions, message it, take over, cancel it or continue it. |
| Timeouts and parking | You want to cap how long a run works. A run that waits for your answer uses no compute. |
| Deputies | The same task comes up again and again. A deputy keeps the mission, its environment and its memory. |
| Deputy contacts | A deputy’s questions should reach the right people, or land in a Slack channel your team watches. |
| Deputy playbooks | You want a ready-made deputy. Read it, copy it, or set it up in a single step. |
| Orchestration | A space, run or deputy should start child runs and follow them, with limits you set and no personal token in the container. |
| Schedules and triggers | Work should start on a timer or when an outside event arrives, with nobody pressing a button. |
Connect Marshal to other systems
| Feature | Use it when |
|---|---|
| Integrations | Your agents need GitHub, Slack, Atlassian and other services. |
| Webhooks | Your own systems should post signed events that start deputy missions. |
| Connect an AI agent (MCP) | An agent outside Marshal, such as Claude Code on your laptop, should start runs and answer their questions as you. |
Know and govern what happened
| Feature | Use it when |
|---|---|
| Audit log | You need to answer “who did this” or “what did this agent do”. |
| Security Center | A package version turned out to be vulnerable. Block it, fix the manifests and stop the spaces that hold it. |
| Output redaction | An agent printed a token. It is hidden from the output, the transcript and Slack, and your admins know which run did it. |
| Pull request impact | You want to show the pull requests your agents open and merge, and how your team’s pace changed after adopting Marshal. |
| Alerts | You want to hear about a space left running for days or a spike in blocked traffic, in the app or by email. |
| Admin dashboards | You want one view of adoption, cost, network traffic, credentials and cluster capacity across the organization. |
| Licensing | You want to see who counts as a licensed user, how your deputies use the concurrency limit, and ask for more capacity. |
| Costs | You want to know what your org spends on LLM and on compute, who drives it, where your storage goes, and which disks you can reclaim. |
Other places to look
- Concepts explains the ideas behind these features.
- Getting started walks you through your first steps.
- How-to guides cover one task at a time.
- Help and fixes starts from what you see on screen.
- CLI reference lists every command.