Install MCP + Skills

The Agent Plugin is the secondary sx surface: one project-scoped MCP server plus the sx-project operating skill. It works with Codex, Claude Code, and other clients that read standard MCP configuration and agent skills.

Before you connect

Install the sx customer package and authenticate once. If you have not done that, follow the Python SDK quickstart.

sx auth login
sx auth status

Connect a workspace

Run this in the workspace where you want the agent to operate:

sx project connect "$PROJECT_ID"

The command:

merges an sx server into .mcp.json without replacing other MCP servers;
writes the sx-project skill for supported agent clients;
stores only the non-secret project UUID in workspace configuration;
lets sx-mcp resolve the same operating-system credential as the CLI.

Reload the workspace or restart the agent client so it discovers the new server and skill.

Prove the connection

Paste this into the agent:

Use the sx project tools to resolve task <TASK_REF>. Do not mutate anything.
Return the exact task identity and explain which project is connected. If access is
refused, report the refusal and required grant without trying another project.

The agent should use experience.task and return either the canonical task or an explicit access refusal. It should not ask for an API key, base URL, request JSON, or repository access.

What the agent can use

Family	Representative tools
Experience	`experience.task`, `experience.search_segments`, `experience.selection.snapshot`, `experience.capture`
Embodiments	`embodiment.parts`, `embodiment.get`, `embodiment.validate`, `embodiment.readiness`
Worlds	`worlds.generate`, `worlds.evaluate`, `worlds.evaluate_subject`
Autonomy	`autonomy.train`, `autonomy.improve`, `autonomy.deploy`
Fleet	`fleet.research`, `fleet.rollback`, commissioning readiness tools
Operations	`operation.status`, `operation.updates`, `operation.result`, `operation.cancel`

Tool inputs are bounded schemas. The project and credential are injected by the connection; they are not tool parameters an agent can swap.

Set an authority boundary

Use four sentences:

Outcome: Make <RESULT> true.
Scope: Work only in project <PROJECT> on <TASK / DATASET / APPLICATION>.
Authority: You may <READ / SUBMIT BOUNDED WORK>; ask before <SPEND / PROMOTION / MOTION>.
Proof: Return exact resources, operation refs, terminal results, evidence, and refusals.

Level	Delegate	Keep human
Inspect	discovery, lineage tracing, compatibility checks, evidence review	whether to act
Operate	bounded capture, generation, training, or evaluation already authorized	changing intent or budget
Promote	readiness assembly and recommendation	deployment approval and rollback decisions
Physical	commissioning preparation and live evidence summaries	hardware admission, motion, and safety policy

Credentials and CI

Human agents use the OS credential created by sx auth login. Headless agents use a project-scoped SX_API_KEY injected by the runtime. Never paste either credential into a prompt, skill, MCP configuration, notebook, or source file.

Troubleshooting

Symptom	Fix
`sx` tools are missing	reload the workspace and verify the client reads `.mcp.json`
agent asks for a service URL	stop; the installed server owns hosted endpoints
project access is refused	grant that identity access to the intended project; do not switch silently
operation disappeared after disconnect	call `operation.status` with the saved reference
agent resubmits after losing output	reattach first; do not duplicate durable work

Next, use the agent playbooks for data curation, evaluation, improvement, and deployment readiness.