Connect AlertForge to an AI agent

AlertForge's agent surface is one Streamable HTTP MCP endpoint: https://alertforge.ai/api/mcp. Unauthenticated calls reach one read-only info tool. For account tools, Claude, ChatGPT, Hermes Agent, OpenClaw, and other OAuth-aware clients discover AlertForge's secure sign-in automatically from the endpoint, then open the login and approval screen. Public tool discovery is not proof of account access: the agent must successfully call get_balance after approval. Remote Hermes installations use AlertForge's hosted approval link, so customers never paste a localhost callback, open a VPS port, or create an SSH tunnel. No password or key belongs in the conversation.

Paste the request below into your agent. It handles discovery and every setup action its app permits. The customer keeps control of the security boundary: only they sign in and approve access.

If an agent cannot use OAuth, create a revocable key at /dashboard/agents. Add the endpoint above with the request header Authorization: Bearer <key>. Put the key only in the client's connection settings — never in a public conversation or document. AlertForge shows client-specific instructions immediately after creating the key.

One tool needs no account connection. The rest act on your account and require OAuth or a fallback access key — grouped below by what they build.

Public — no auth

Animated overlay packs — account connection required

Stream Alerts & transparent videos — account connection required

Every new overlay pack starts with a Style Sheet, not a scene. AlertForge's builder always asks the streamer the same 3 questions before generating anything — channel name, how they'd like the logo or character, and how they'd like the theme and colors — and the agent surface asks a real user those same 3 questions rather than inventing an answer. plan_design_sheet prices that Style Sheet from the user's real answers, create_design_sheet spends the plan and starts rendering it, and get_pack_status reports back the finished image for the agent to show the user. approve_design_sheet then records that the user — never the agent on its own judgment — actually approved it.

Only once that approval is recorded can scenes be planned: plan_overlay_pack prices scenes against the approved project and returns the exact credit cost before anything is created — call it without an approved project and it refuses, returning the same 3 questions again instead of guessing a theme. Scenes now include four sheet-pack kinds — Twitch panels, activity-feed labels, subscriber badges, and channel emotes — each rendered as one themed sheet and cropped into individual tiles. create_overlay_pack spends that plan and attaches the scenes to the project right away. get_pack_status advances the pack and reports progress per scene — an agent calls it in a loop until every scene is done. From there, either export_pack_zip for a downloadable ZIP, or hand the editor URL back to the user to finish in the browser.

Stream Alerts follow the same price-first, two-step pattern: call list_alert_types to see what's available and what it costs, then plan_alert_pack with one alert, a handful, or the full set — a single alert is fully supported, not just a whole pack. create_alert_pack spends the plan and starts rendering (an active subscription is required at this step). get_alert_pack_status is a plain read-only check — nothing advances automatically, the render happens on AlertForge's own schedule, so just call it again after a minute or two. Every alert needs a character; generate_alert_character can create one from a description (polled with get_alert_character_status), or you can upload your own in the dashboard and pass its path in.

A standalone transparent video (no character, no alert wiring) works the same way, one step simpler: plan_transparent_video, create_transparent_video, then get_render_status to poll it.

Generation spends credits from your AlertForge account — the same balance the web builder draws from. Every plan tool (plan_design_sheet, plan_overlay_pack, plan_alert_pack, plan_transparent_video) quotes the exact cost up front, and the matching create tool won't spend anything your balance can't cover. generate_alert_character also spends credits, at the price list_alert_types reports; every other tool is free to call. See /pricing for what credits cost and how each tier allots them. If auto top-up is enabled on the account and a create tool runs short, the 402 response comes back with a real Stripe Checkout link instead of just an error — pay once and retry the same plan.

Every account tool requires an active paid subscription: Starter, Pro, Max, or Ultra. A credit top-up alone does not unlock account tools. If the subscription lapses, the next account call stops. Revoking a connected app or fallback key at /dashboard/agents takes effect immediately. Each connection is rate-limited and capped on how many new Style Sheets, packs, alerts, and videos it can start per day. These limits exist to catch a runaway agent loop, not to constrain normal use.

export_pack_zip gives you the flat agent pack: each scene's WebM/PNG plus a manifest.json, zipped with a signed download link. The full themed folder download — the same canvas-composited branding and scene/stinger/alert layout the web app's own "Download pack" button produces — isn't part of the agent pack. Every tool result includes the editor URL; open it to get that from the web app.

• /docs/apiAPI & agent access — the unauthenticated MCP tools and machine-readable discovery documents
• /docs/webhooksAlertForge webhooks — current webhook availability and the supported status-polling path
• /.well-known/agent-skills/index.jsonAgent Skills index — step-by-step skills for building a pack and reading pricing
• /llms.txtllms.txt — whole-site fact sheet for AI agents