Meta Llama

Bring Meta's first-party Llama API into your flows. Agents run chat completions with tool calling and structured JSON output, ask single-prompt questions, screen content with Llama Guard moderation, and discover which models a key can use.

4 actions API key available

Meta Llama website ↗ Platform Documentation ↗ Capability data verified 2026-07-25

An inbound support message arrives in the shared mailbox

Moderate Content screens the customer's text with Llama Guard before it reaches the model

Ask Llama drafts a reply under a system prompt that holds tone and policy constraints

Moderate Content screens the drafted reply against the hazard taxonomy

The flow routes any flagged draft, plus anything touching refunds or account changes, to review

A support lead approves or edits the held draft before it goes anywhere

The approved reply sends through the team's mailbox

What This Integration Enables

This connector is Meta's first-party Llama API, and its most operationally interesting piece is not the text generation. It is that safety screening ships as a first-class action. Moderate Content runs Llama Guard against any text and returns specific hazard categories from a published taxonomy: S5 Defamation, S6 Specialized Advice, S7 Privacy, S10 Hate, and the rest. That turns content safety from a policy document into a workflow step, one an agent runs on the way in and on the way out. Running against Meta's first-party deployment also keeps the account relationship simple: FlowRunner is BYOK, so the Llama API key, its limits, and its bill stay yours, while the flow layer adds what an unattended process needs, retries, routing, and a record of who approved what. - Draft, summarize, classify, and translate with Ask Llama, no messages array required - Run multi-turn conversations with Chat Completion, including tool calling and multimodal content with Llama 4 Maverick and Scout - Enforce structured output with JSON-schema response formats so extraction feeds systems, not parsers - Screen user input and generated output with Llama Guard before either crosses a boundary - Discover which models a key can reach with List Models before a flow depends on one

Without FlowRunner

Model output ships unscreened Whatever the model generates goes wherever the workflow points it, and nobody checks the direction

Moderation is a homemade regex Input screening is a keyword list someone wrote in an afternoon and nobody maintains

Extraction output is freeform text The model answers in prose, and a brittle parser downstream guesses at the fields

With FlowRunner

Both directions pass through Llama Guard User input is screened before the model sees it, and generated output before a customer does

Flags carry a taxonomy, not a hunch Llama Guard returns specific hazard categories, so routing rules act on S6 differently than S10

Structured output is enforced JSON-schema response formats make Chat Completion return fields, not prose, for downstream steps

Use Case Scenarios

The support inbox with a screen on both doors

Inbound messages are screened with Moderate Content before any model call, so prompt-injection attempts and abusive content get routed to a human queue instead of into an agent's context. Clean messages get a reply drafted by Ask Llama, the draft is screened again on the way out, and only unflagged drafts about routine topics send automatically through [Gmail](/integrations/gmail-service). Everything else waits for a person. The flags are category codes, not guesses, so the routing rules are auditable.

Documents in, fields out, no parser in between

Vendor documents arrive as attachments. Chat Completion runs with a JSON-schema response format that names the fields the finance team needs: vendor, amount, currency, due date, early-payment terms. The response is structured by contract, not by hope, so the flow writes rows directly to [Airtable](/integrations/airtable) and [Google Sheets](/integrations/google-sheets) without a fragile text parser. Documents where the model reports low certainty on an amount get held for a person instead of written as fact.

A channel assistant that picks its own tools, within limits

On a [Slack](/integrations/slack) message, the flow passes the thread to Chat Completion with tool definitions for the team's internal lookups: order status, account standing, shipment tracking. The model decides which tool fits the question, FlowRunner executes the call, and the answer posts back to the thread. List Models runs in the flow's preflight, confirming the key still reaches the intended Llama 4 model before the assistant takes questions, so a plan change fails at deploy time rather than mid-conversation.

Human-in-Loop Highlight

The operation that cannot be recalled here is the send: an Ask Llama draft leaving through the company's own mailbox under the company's name. Llama Guard narrows the risk but also defines its own limit, and the taxonomy says where that limit is. A reply that trips S6 Specialized Advice is not abusive content; it is the model drifting into financial, legal, or medical territory where a confident wrong answer creates liability. So the FlowRunner flow treats Guard's categories as routing, not as a verdict: S6 and S5 flags hold the draft and page a qualified reviewer with the customer's message, the draft, and the specific category that fired. The reviewer edits or approves, and the flow logs which categories fired and who cleared them. Screening at machine speed, judgment where the taxonomy itself says a person belongs.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via preferred channel

Human decides

Agent resumes with decision

Agent Capabilities

4 actions

Text Generation

Chat Completion Generates a chat completion with a Meta-hosted Llama model from a full messages array, with control over multi-turn context, system instructions, and multimodal content parts. Supports tool calling, structured JSON output via a response format object, stop sequences, and full sampling controls, returning the OpenAI-shaped response with choices, tool calls, finish reason, and token usage.
Ask Llama Sends a single prompt and returns just the assistant's text, with an optional system prompt for role, tone, and constraints. The fast path for summarizing, drafting, classifying, and translating inside a flow, returning the text with model, finish reason, and usage.

Safety

Moderate Content Checks text against Meta's Llama Guard safety models, returning whether content was flagged and which hazard categories matched across the fourteen-category taxonomy, from S1 Violent Crimes through S14 Code Interpreter Abuse. Run it on user input before the model, on model output before the user, or both.

Models

List Models Lists every Llama model the API key can access, including the Llama 4 Maverick and Scout multimodal models, Llama 3.3 text models, and the Llama Guard moderation models, with owners and creation timestamps. The preflight check before a flow depends on a specific deployment.

Frequently Asked Questions

What can FlowRunner do with Meta Llama?

FlowRunner agents can run Chat Completion, Ask Llama, and Moderate Content in Meta Llama, plus 1 more action.

Does connecting Meta Llama to FlowRunner require OAuth?

No. Meta Llama connects to FlowRunner with an API key, no OAuth flow required.

Can Meta Llama trigger a FlowRunner workflow automatically?

Meta Llama doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.

Start building with Meta Llama

$100 in credits. No card required. Connect in minutes.

Start Building Free Talk to us