presenton Agent工作流

⚙️ Deployment Configurations

The tables below match the environment variables forwarded in this repository’s docker-compose.yml (production, production-gpu, development, and development-gpu). Put values in a .env file next to the compose file, or export them before docker compose up. The Electron app backend can read the same names when run outside Docker.

Other optional variables exist in code (for example advanced Mem0 paths, LiteParse runners, or FAST_API_INTERNAL_URL when Next.js and FastAPI are not same-origin); they are not wired in docker-compose.yml. Supported names are discoverable from servers/fastapi/utils/get_env.py and the Next.js server utilities under servers/nextjs/.

Air-gapped deployments

The Community gallery is enabled by default and reads presentation designs from the Presenton community service. Air-gapped deployments must disable it:

PRESENTON_COMMUNITY_ENABLED=false

For Docker Compose, add the value to the .env file next to docker-compose.yml before starting Presenton. For docker run, pass -e PRESENTON_COMMUNITY_ENABLED=false. This removes Community from the UI and prevents the backend from sending Community gallery or design-reference requests.

LLM and API keys

All advanced text-provider settings are optional. Use Reset advanced settings in Settings or onboarding to remove overrides and inherit application/provider defaults.

Ollama

Use when LLM is ollama:

Presentation memory (Mem0 OSS)

Mem0 uses local Qdrant + SQLite (OSS); memory is scoped per presentation.

By default the Docker runtime now points Mem0 at a local Ollama-compatible LLM endpoint, so it no longer needs an OpenAI key just to initialize. If you want to use OpenAI instead, set MEM0_LLM_BASE_URL/MEM0_LLM_API_KEY to your OpenAI-compatible endpoint and key. Docker images install the default spaCy model (en_core_web_sm) during build so Mem0 can start without extra setup on each run.

Document parsing (LiteParse)

Database

Image generation

These variables match docker-compose.yml. IMAGE_PROVIDER selects the backend (pexels, pixabay, gemini_flash, nanobanana_pro, gpt-image-2, gpt-image-1.5, comfyui, open_webui). Use OPENAI_API_KEY for OpenAI image modes and GOOGLE_API_KEY for Gemini image modes (same keys as the LLM section).

Existing IMAGE_PROVIDER=dall-e-3 configurations automatically migrate to gpt-image-2. Legacy standard quality maps to medium, and hd maps to high. Set GPT_IMAGE_2_QUALITY to choose a different quality.

The parallel image generation option applies everywhere images are generated: initial presentation generation, slide editing and regeneration, direct image requests, and assistant image tools.

Telemetry

Multi-user authentication

Presenton supports multiple accounts with a private workspace for each user. The first account becomes the primary administrator and can create, reset, or remove other accounts from Admin → Users.

Existing single-user installations are upgraded automatically: the current account becomes the primary administrator, while its presentations, templates, tasks, and other owned data stay attached to the same account.

Set up the primary administrator

On a new installation, open Presenton and follow the account setup screen. For an unattended Docker deployment, you can create the primary administrator on first boot with environment variables:

docker run -it --name presenton \
  -p 5001:80 \
  -e AUTH_USERNAME=admin \
  -e AUTH_PASSWORD=change-this-password \
  -v "./app_data:/app_data" \
  ghcr.io/presenton/presenton:latest

Usernames must contain at least 3 characters, and new passwords must contain at least 8 characters. Older six- or seven-character passwords remain valid after an upgrade.

Authentication environment variables

Presenton Cloud provider

Presenton Cloud is an optional, installation-wide generation provider. It is not an authentication method for the self-hosted instance. Create or sign in to the local administrator account first, then connect Presenton from provider onboarding.

Only the local administrator can connect, replace, or disconnect the provider. The browser displays a short device code and opens the hosted Presenton approval page. After approval, the delegated access and rotating refresh tokens are encrypted at rest and stored as one global provider credential; they are never returned to the browser or stored per local user.

Selecting Presenton as the text provider saves LLM=presenton. Presentation generation and document uploads then use the Presenton Cloud API and its token. Connecting the provider alone does not change generation: when another provider is selected, the existing local generation pipeline remains unchanged. Disconnecting revokes and removes the global credentials and deselects Presenton.

No OAuth client registration, client secret, or environment configuration is required. Official builds contain the cloud URL and first-party public device-flow client ID.

To rotate credentials from the environment:

docker stop presenton
docker rm presenton
docker run -it --name presenton \
  -p 5001:80 \
  -e AUTH_USERNAME=admin \
  -e AUTH_PASSWORD=new-secure-password \
  -e AUTH_OVERRIDE_FROM_ENV=true \
  -v "./app_data:/app_data" \
  ghcr.io/presenton/presenton:latest

For account recovery, use the same command with RESET_AUTH=true instead of AUTH_OVERRIDE_FROM_ENV=true. Both operations preserve the administrator's user ID and owned data, and invalidate existing browser sessions and API keys. Remove the one-time flag after the successful startup.

[!IMPORTANT] Do not remove authentication fields from app_data/userConfig.json to reset access. Presenton stores a hashed recovery copy of the primary administrator credentials and the session-signing secret there. Use the recovery variables above to preserve the database account and its ownership links.

To sign out, open Settings → Other → Sign out.

API and MCP authentication

Presenton exposes a client-neutral MCP 2025-11-25 Streamable HTTP endpoint at /mcp. When authentication is enabled, REST API and MCP clients use the same administrator-provisioned, user-scoped API key. Browser JWT cookies are not accepted as MCP credentials.

1. An administrator creates the target Presenton user, then calls POST /api/v1/admin/api-keys from an authenticated admin session with {"user_id":"...","label":"My API client","expiry_days":90}. The response contains the new key. Listing and revocation use GET /api/v1/admin/api-keys and POST /api/v1/admin/api-keys/{api_key_id}/revoke. An administrator can securely reveal a key again with GET /api/v1/admin/api-keys/{api_key_id}/token.

2. Configure any Streamable HTTP MCP client to send the generated sk-presenton-... key on every request:

{
  "servers": {
    "presenton": {
      "url": "http://localhost:5001/mcp",
      "type": "http",
      "headers": {
        "Authorization": "Bearer sk-presenton-0123456789abcdef.REPLACE_WITH_SECRET"
      }
    }
  },
  "inputs": []
}

Notes:

- This example uses VS Code's .vscode/mcp.json format. Use the equivalent Streamable HTTP and bearer-header configuration in Open WebUI or any other conforming MCP client. - API keys retain an Argon2 authentication hash and an encrypted, admin-only recoverable copy. They expire by default after 90 days and represent exactly the selected Presenton user. The same key authenticates general REST endpoints and MCP, but cannot sign in to the browser or change admin/provider settings. - The API key is exchanged inside the MCP process for a short-lived internal user session. The long-lived key is never forwarded to presentation APIs. - Revoking the key through the admin API takes effect immediately. - Standard, Smart Mode, and template generation tools return an async task ID. Poll get_job_status until completed or error; template listing is immediate. - MCP exposes only the polling Smart workflow (start_smart_presentation followed by get_job_status) so Standard and Smart generation use the same job model. - upload_template_assets accepts one PPTX and its replacement fonts in a single call, then returns the data needed by start_template_generation. - upload_files accepts reference documents and images as base64 and returns user-scoped paths that can be passed unchanged to Standard or Smart generation. - External REST API and MCP access are not available in the Electron desktop app (PRESENTON_ELECTRON=true). Its local HTTP servers are internal application services. Electron runs without account login by default and supplies a private credential generated at each launch to its own app and export requests; anonymous local clients cannot access those services. The MCP server remains disabled there. - Only asset subdirectories are served under /app_data; configuration, databases and logs are never web assets. Document preview readers accept .txt document files within the allowed asset/temp directories and reject export task JSON and private application files.

Note: LLM and image variables above are forwarded from docker-compose.yml when set in .env.

<br> <br>

Docker Run Examples by Provider

Same variables as compose; use -e instead of .env when running docker run directly.

- Using OpenAI <pre><code class="language-bash">docker run -it --name presenton -p 5001:80 -e LLM="openai" -e OPENAI_API_KEY="******" -e IMAGE_PROVIDER="gpt-image-2" -e CAN_CHANGE_KEYS="false" -v "./app_data:/app_data" ghcr.io/presenton/presenton:latest</code></pre>

- Using Google <pre><code class="language-bash">docker run -it --name presenton -p 5001:80 -e LLM="google" -e GOOGLE_API_KEY="******" -e IMAGE_PROVIDER="gemini_flash" -e CAN_CHANGE_KEYS="false" -v "./app_data:/app_data" ghcr.io/presenton/presenton:latest</code></pre>

- Using Vertex AI (API key mode) <pre><code class="language-bash">docker run -it --name presenton -p 5001:80 -e LLM="vertex" -e VERTEX_API_KEY="******" -e VERTEX_MODEL="gemini-2.5-flash" -e IMAGE_PROVIDER="gemini_flash" -e CAN_CHANGE_KEYS="false" -v "./app_data:/app_data" ghcr.io/presenton/presenton:latest</code></pre>

- Using Azure OpenAI <pre><code class="language-bash">docker run -it --name presenton -p 5001:80 -e LLM="azure" -e AZURE_OPENAI_API_KEY="****" -e AZURE_OPENAI_MODEL="gpt-4.1" -e AZURE_OPENAI_API_VERSION="2024-10-21" -e AZURE_OPENAI_ENDPOINT="https://YOUR-RESOURCE.openai.azure.com" -e IMAGE_PROVIDER="pexels" -e PEXELS_API_KEY="****" -e CAN_CHANGE_KEYS="false" -v "./app_data:/app_data" ghcr.io/presenton/presenton:latest</code></pre>

- Using Amazon Bedrock (on-demand model ID) — see docs/amazon-bedrock.md for inference profiles, IAM, and troubleshooting. <pre><code class="language-bash">docker run -it --name presenton -p 5001:80 -e LLM="bedrock" -e BEDROCK_REGION="us-east-1" -e BEDROCK_AWS_ACCESS_KEY_ID="****" -e BEDROCK_AWS_SECRET_ACCESS_KEY="**" -e BEDROCK_MODEL="us.anthropic.claude-3-5-haiku-20241022-v1:0" -e IMAGE_PROVIDER="pexels" -e PEXELS_API_KEY="****" -e CAN_CHANGE_KEYS="false" -v "./app_data:/app_data" ghcr.io/presenton/presenton:latest</code></pre>

- Using Amazon Bedrock (inference profile ARN, e.g. Claude Sonnet 4.6) <pre><code class="language-bash">docker run -it --name presenton -p 5001:80 -e LLM="bedrock" -e BEDROCK_REGION="us-east-1" -e BEDROCK_AWS_ACCESS_KEY_ID="****" -e BEDROCK_AWS_SECRET_ACCESS_KEY="**" -e BEDROCK_MODEL="arn:aws:bedrock:us-east-1:YOUR_ACCOUNT_ID:inference-profile/us.anthropic.claude-sonnet-4-6" -e IMAGE_PROVIDER="pexels" -e PEXELS_API_KEY="****" -e CAN_CHANGE_KEYS="false" -v "./app_data:/app_data" ghcr.io/presenton/presenton:latest</code></pre>

- Using Fireworks <pre><code class="language-bash">docker run -it --name presenton -p 5001:80 -e LLM="fireworks" -e FIREWORKS_API_KEY="****" -e FIREWORKS_MODEL="accounts/fireworks/models/llama-v3p1-8b-instruct" -e IMAGE_PROVIDER="pexels" -e PEXELS_API_KEY="****" -e CAN_CHANGE_KEYS="false" -v "./app_data:/app_data" ghcr.io/presenton/presenton:latest</code></pre>

- Using Together AI <pre><code class="language-bash">docker run -it --name presenton -p 5001:80 -e LLM="together" -e TOGETHER_API_KEY="****" -e TOGETHER_MODEL="openai/gpt-oss-20b" -e IMAGE_PROVIDER="pexels" -e PEXELS_API_KEY="****" -e CAN_CHANGE_KEYS="false" -v "./app_data:/app_data" ghcr.io/presenton/presenton:latest</code></pre>

- Using Ollama <pre><code class="language-bash">docker run -it --name presenton -p 5001:80 -e LLM="ollama" -e OLLAMA_MODEL="llama3.2:3b" -e IMAGE_PROVIDER="pexels" -e PEXELS_API_KEY="*******" -e CAN_CHANGE_KEYS="false" -v "./app_data:/app_data" ghcr.io/presenton/presenton:latest</code></pre>

- Using Anthropic <pre><code class="language-bash">docker run -it --name presenton -p 5001:80 -e LLM="anthropic" -e ANTHROPIC_API_KEY="****" -e IMAGE_PROVIDER="pexels" -e PEXELS_API_KEY="****" -e CAN_CHANGE_KEYS="false" -v "./app_data:/app_data" ghcr.io/presenton/presenton:latest</code></pre>

- Using LM Studio (local) <pre><code class="language-bash">docker run -it --name presenton -p 5001:80 -e LLM="lmstudio" -e LMSTUDIO_BASE_URL="http://host.docker.internal:1234" -e LMSTUDIO_MODEL="openai/gpt-oss-20b" -e IMAGE_PROVIDER="pexels" -e PEXELS_API_KEY="******" -e CAN_CHANGE_KEYS="false" -v "./app_data:/app_data" ghcr.io/presenton/presenton:latest</code></pre>

- Using OpenAI Compatible LLM API <pre><code class="language-bash">docker run -it -p 5001:80 -e CAN_CHANGE_KEYS="false" -e LLM="custom" -e CUSTOM_LLM_URL="http://***" -e CUSTOM_LLM_API_KEY="" -e CUSTOM_MODEL="llama3.2:3b" -e IMAGE_PROVIDER="pexels" -e PEXELS_API_KEY="*****" -v "./app_data:/app_data" ghcr.io/presenton/presenton:latest</code></pre>

- Running Presenton with GPU Support To use GPU acceleration with Ollama models, you need to install and configure the NVIDIA Container Toolkit. This allows Docker containers to access your NVIDIA GPU. Once the NVIDIA Container Toolkit is installed and configured, you can run Presenton with GPU support by adding the --gpus=all flag: <pre><code class="language-bash">docker run -it --name presenton --gpus=all -p 5001:80 -e LLM="ollama" -e OLLAMA_MODEL="llama3.2:3b" -e IMAGE_PROVIDER="pexels" -e PEXELS_API_KEY="*******" -e CAN_CHANGE_KEYS="false" -v "./app_data:/app_data" ghcr.io/presenton/presenton:latest</code></pre>

• Using an OpenAI-Compatible Image Provider

This routes all slide image requests through your OpenAI-compatible gateway (LiteLLM, Azure, vLLM, etc.) while keeping the text LLM configuration independent: <pre><code class="language-bash">docker run -it --name presenton -p 5001:80 -e IMAGE_PROVIDER="openai_compatible" -e OPENAI_COMPAT_IMAGE_BASE_URL="https://proxy.example.com/v1" -e OPENAI_COMPAT_IMAGE_API_KEY="******" -e OPENAI_COMPAT_IMAGE_MODEL="gpt-image-1" -v "./app_data:/app_data" ghcr.io/presenton/presenton:latest</code></pre>

#

✨ Generate Presentation via API

Generate Presentation

<p> <strong>Endpoint:</strong> <code>/api/v1/ppt/presentation/generate</code><br> <strong>Method:</strong> <code>POST</code><br> <strong>Content-Type:</strong> <code>application/json</code> </p>

<p> <strong>Authentication (API key):</strong><br> All <code>/api/v1/</code> routes except the public authentication endpoints require authentication. An administrator creates an access key under <strong>Admin → API keys</strong>. Send that <code>sk-presenton-...</code> key as <code>Authorization: Bearer YOUR_KEY</code>. API keys act as their owning user and cannot call browser-session-only administrator endpoints. </p>

Request Body

<tr> <td><code>content</code></td> <td>string</td> <td>Yes</td> <td>Main content used to generate the presentation. When using <code>slides_markdown</code> without an additional overall prompt, pass an empty string.</td> </tr>

<tr> <td><code>slides_markdown</code></td> <td>string[] | null</td> <td>No</td> <td>Provide one Markdown string per slide instead of asking Presenton to generate the slide outlines. The array length determines the number of slides.</td> </tr>

<tr> <td><code>instructions</code></td> <td>string | null</td> <td>No</td> <td>Additional generation instructions.</td> </tr>

<tr> <td><code>tone</code></td> <td>string</td> <td>No</td> <td> Text tone (default: <code>"default"</code>). Options: <code>default</code>, <code>casual</code>, <code>professional</code>, <code>funny</code>, <code>educational</code>, <code>sales_pitch</code> </td> </tr>

<tr> <td><code>verbosity</code></td> <td>string</td> <td>No</td> <td> Content density (default: <code>"standard"</code>). Options: <code>concise</code>, <code>standard</code>, <code>text-heavy</code> </td> </tr>

<tr> <td><code>web_search</code></td> <td>boolean</td> <td>No</td> <td>Enable web search grounding (default: <code>false</code>).</td> </tr>

<tr> <td><code>n_slides</code></td> <td>integer</td> <td>No</td> <td>Number of slides to generate. When omitted, the count is auto-detected. When using <code>slides_markdown</code>, its array length is used and <code>n_slides</code> can be omitted.</td> </tr>

<tr> <td><code>language</code></td> <td>string</td> <td>No</td> <td>Presentation language (default: <code>"English"</code>).</td> </tr>

<tr> <td><code>template</code></td> <td>string</td> <td>No</td> <td>Template name (default: <code>"general"</code>).</td> </tr>

<tr> <td><code>include_table_of_contents</code></td> <td>boolean</td> <td>No</td> <td>Include table of contents slide (default: <code>false</code>).</td> </tr>

<tr> <td><code>include_title_slide</code></td> <td>boolean</td> <td>No</td> <td>Include title slide (default: <code>true</code>).</td> </tr>

<tr> <td><code>files</code></td> <td>string[] | null</td> <td>No</td> <td> Files to use in generation. Upload first via <code>/api/v1/ppt/files/upload</code>. </td> </tr>

<tr> <td><code>export_as</code></td> <td>string</td> <td>No</td> <td> Export format (default: <code>"pptx"</code>). Options: <code>pptx</code>, <code>pdf</code> </td> </tr>

</tbody> </table>

Response

{
  "presentation_id": "string",
  "path": "string",
  "edit_path": "string"
}

Example (curl + API key)

curl \
  -X POST http://localhost:5001/api/v1/ppt/presentation/generate \
  -H "Authorization: Bearer sk-presenton-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
   "content": "Introduction to Machine Learning",
    "n_slides": 5,
    "language": "English",
    "template": "general",
    "export_as": "pptx"
  }'

Example: provide the content for every slide with slides_markdown

Each array item becomes one slide. Use ordinary Markdown headings, lists, emphasis, and tables to describe the audience-facing content. The <code>content</code> field is currently required by the API schema, so set it to an empty string when you do not want to provide an additional overall prompt.

curl \
  -X POST http://localhost:5001/api/v1/ppt/presentation/generate \
  -H "Authorization: Bearer sk-presenton-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "content": "",
    "slides_markdown": [
      "# Acme Q2 Review\n\nRevenue grew **18% year over year** while customer retention reached 94%.",
      "# What changed\n\n- Enterprise revenue increased 24%\n- Support response time fell to 42 minutes\n- Expansion revenue offset slower new-logo growth",
      "# Next quarter\n\n1. Launch the self-service onboarding flow\n2. Expand the enterprise pilot program\n3. Reduce infrastructure cost per account by 10%"
    ],
    "template": "general",
    "language": "English",
    "export_as": "pdf"
  }'

Example: provide numeric data for a chart or table layout

Markdown tables with numeric columns help Presenton select a compatible chart or table layout. You can use <code>instructions</code> for presentation-level visual direction without adding those directions to the visible slide text.

curl \
  -X POST http://localhost:5001/api/v1/ppt/presentation/generate \
  -H "Authorization: Bearer sk-presenton-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "content": "",
    "slides_markdown": [
      "# Regional revenue\n\n| Region | Q1 | Q2 |\n| --- | ---: | ---: |\n| North America | 2.4 | 2.9 |\n| Europe | 1.7 | 2.1 |\n| Asia Pacific | 1.1 | 1.6 |",
      "# Key takeaway\n\nAsia Pacific grew fastest at **45%**, while North America remained the largest market."
    ],
    "instructions": "Use a bar chart for slide 1 and an emphasis layout for slide 2.",
    "template": "general",
    "export_as": "pptx"
  }'

<blockquote> <strong>How <code>slides_markdown</code> works:</strong> It skips AI outline generation, but it is not a deterministic Markdown-to-slide renderer. Presenton still uses the configured text LLM to select suitable layouts and map each Markdown item into the template's structured fields, so text may be rephrased to fit. Provide at most 50 array items. Each item is limited to 100 words by default; longer items are truncated. Self-hosted deployments can change the per-slide limit with <code>MAX_OUTLINE_WORDS</code>. </blockquote>

Example Response

{
  "presentation_id": "d3000f96-096c-4768-b67b-e99aed029b57",
  "path": "/app_data/d3000f96-096c-4768-b67b-e99aed029b57/Introduction_to_Machine_Learning.pptx",
  "edit_path": "/presentation?id=d3000f96-096c-4768-b67b-e99aed029b57"
}

<blockquote> <strong>Note:</strong> Prepend your server’s root URL to <code>path</code> and <code>edit_path</code> to construct valid links. </blockquote>

Documentation & Tutorials

• Deploy Presenton
• Full API Documentation
• Generate Presentations via API in 5 Minutes
• Create Presentations from CSV using AI
• Create Data Reports Using AI

#