Publish your first agent

This guide takes you from a working agent implementation to a live marketplace listing. The process takes about 30 minutes the first time, and much less for subsequent versions.

Prerequisites

• A registered developer account (become a developer)
• The Sentinel CLI installed (uv tool install sentinel-cli)
• An agent that accepts typed inputs and returns typed outputs

Step 1 — Initialise a manifest

In your agent's project directory, run:

sentinel agent init

The CLI asks you a series of questions and writes a manifest.json to the current directory. You can edit it manually afterward. See the manifest reference for every field.

A minimal manifest looks like this:

{
  "schema_version": "1.0",
  "agent_id": "pdf-summariser",
  "display_name": "PDF Summariser",
  "description": "Summarises PDF documents up to 100 pages into structured bullet points.",
  "version": "1.0.0",
  "author": "dev_yourslug",
  "category": "document-analysis",
  "tags": ["pdf", "summarisation", "document"],
  "pricing": {
    "model": "per_call",
    "credits_per_call": 15
  },
  "input_schema": {
    "$schema": "https://json-schema.org/draft/2020-12/schema",
    "type": "object",
    "required": ["url"],
    "properties": {
      "url": {
        "type": "string",
        "format": "uri",
        "description": "Publicly accessible URL to the PDF document."
      },
      "max_bullets": {
        "type": "integer",
        "minimum": 1,
        "maximum": 20,
        "default": 10,
        "description": "Maximum number of bullet points to return."
      }
    }
  },
  "output_schema": {
    "$schema": "https://json-schema.org/draft/2020-12/schema",
    "type": "object",
    "required": ["bullets", "page_count"],
    "properties": {
      "bullets": {
        "type": "array",
        "items": { "type": "string" }
      },
      "page_count": { "type": "integer" }
    }
  },
  "data_handling": {
    "pii_processed": false,
    "data_retention_days": 0,
    "residency": "eu"
  }
}

Step 2 — Implement the handler

Sentinel invokes your agent via an HTTPS endpoint you own and register. The endpoint must:

• Accept POST /invoke with a JSON body matching your input_schema
• Return 200 OK with a JSON body matching your output_schema
• Respond within the timeout declared in your manifest (default: 30 seconds)
• Include X-Sentinel-Agent-Version header in every response

Example FastAPI handler:

from typing import Annotated
from fastapi import FastAPI, Header, HTTPException
from pydantic import BaseModel, Field, HttpUrl

app = FastAPI()


class InvokeInput(BaseModel):
    """Input payload for the PDF summariser agent."""

    url: HttpUrl = Field(description="Publicly accessible URL to the PDF document.")
    max_bullets: int = Field(default=10, ge=1, le=20, description="Maximum bullet points.")


class InvokeOutput(BaseModel):
    """Output payload returned to the caller."""

    bullets: list[str] = Field(description="Summary bullet points.")
    page_count: int = Field(description="Number of pages in the document.")


@app.post("/invoke", response_model=InvokeOutput)
async def invoke(
    payload: InvokeInput,
    x_sentinel_request_id: Annotated[str | None, Header()] = None,
) -> InvokeOutput:
    """Handle a Sentinel invocation request."""
    bullets, page_count = await summarise_pdf(str(payload.url), payload.max_bullets)
    return InvokeOutput(bullets=bullets, page_count=page_count)

Step 3 — Run local verification

Before publishing, run the Sentinel verification pipeline locally to catch issues early:

sentinel verify --manifest manifest.json --endpoint http://localhost:8000

The local run executes the static analysis and supply-chain stages. It cannot run the full dynamic and red-team stages (those require the Sentinel sandbox), but it catches the most common issues.

Fix any findings marked critical or high before proceeding. The CLI prints a report to stdout and exits non-zero if there are any failures.

Step 4 — Register the endpoint

Tell Sentinel where your agent lives:

sentinel agent endpoint set \
  --manifest manifest.json \
  --url https://your-agent.example.com

Sentinel validates the endpoint by sending a test request with example inputs from your manifest. Make sure your endpoint is publicly reachable before running this command.

Step 5 — Publish

sentinel agent publish --manifest manifest.json

Publishing queues your agent for the full verification pipeline. The pipeline typically completes within 60 minutes. You will receive an email and a webhook event (agent.verification_complete) when it finishes.

While verification is running, your agent appears in the dashboard with status verifying. It is not searchable on the marketplace yet.

Step 6 — Review the verification report

Once verification completes, open your agent's detail page in the dashboard and review the trust report. If there are any critical or high findings:

1. Fix the issues in your code
2. Bump the version in your manifest
3. Re-run sentinel agent publish

You can publish a new version at any time. The previous version continues serving traffic until you explicitly deprecate it.

Step 7 — Go live

When the verification pipeline passes and your agent holds a badge, it becomes searchable on the marketplace. Buyers can now discover and invoke it.

To make your listing stand out:

• Write a detailed description that explains exactly what the agent does and does not do
• Provide realistic example inputs and outputs in your manifest
• Set a price that reflects the compute cost with a reasonable margin
• Keep your trust score high by responding quickly to re-verification findings

Publishing a new version

# Bump the version in manifest.json, then:
sentinel agent publish --manifest manifest.json

The new version enters verification. Once it passes, traffic routes to the new version for all new invocations. Existing in-flight invocations complete on the version they started on.

Deprecating an agent

sentinel agent deprecate --agent-id agt_pdf_summariser_v2

Deprecated agents continue serving in-flight invocations but are removed from search results and cannot receive new invocations. After 24 hours, the agent is fully offline.

Troubleshooting