A Deepgram MCP that sends Token, not Bearer, and takes audio as a URL
Paste one API key and Deepgram speech to text, text intelligence, the live model catalog, and the whole management plane of projects, keys, members, usage, and billing become workflow steps for visual automations and AI agents. The two nodes that spend money say so in their own descriptions, and the other 24 are free management calls.
What the Deepgram MCP connector does
The Deepgram connector for Mattering Automations is 26 nodes curated from the 50 operations across 36 paths in Deepgram's REST API: pre-recorded speech to text from a media URL, text intelligence over any text, the model catalog, and the management plane of projects, API keys, members and their scopes, invites, request history, usage, billing, credit balances, and purchases. It is a curated surface rather than a wrapper around everything, and the parts left out are named plainly further down.
Every node speaks MCP, the open protocol for connecting AI assistants to tools, so Claude, ChatGPT, and any MCP-capable assistant can use them, alongside Matter, the AI assistant built into the platform. An assistant can transcribe a recording, summarize what was said, report how many audio hours a project burned last week, or list who still has an API key on a project, without anyone opening the Deepgram console.
Because Mattering is a work platform rather than middleware, voice work lands somewhere useful by default: a transcript can become the body of a task in Mattering Productivity, a summary and its sentiment can become a row in a Mattering database, and a scheduled balance check can message a channel before the credit runs out, all on one canvas next to the rest of the work.
Three things are worth knowing before you wire anything, because each of them is a place people lose an afternoon. The authorization header scheme word is Token, not Bearer, and a Bearer header returns 401 with a perfectly valid key. Audio reaches Deepgram as a URL that Deepgram fetches itself, never as an upload, which is a deliberate design decision rather than a missing feature. And transcription is metered pay-per-use billed by audio duration, so a longer file costs proportionally more; no rate is quoted anywhere here on purpose, because rates change and a stale number would be worse than none.
One connection, three surfaces
- Visual automations. Drag Deepgram nodes onto the canvas and wire them to triggers, conditions, and other apps.
- AI agents, yours or ours. These nodes speak MCP, the open protocol, so Claude, ChatGPT, and any MCP-capable assistant can use them, and Matter, the agent built into the platform, uses them natively.
- The whole platform. Combine with native tasks, projects, and other connectors, no middleware in between.
Every Deepgram node, ready to drop on the canvas
26 nodes, generated from the connector itself so this list always matches what ships. Expand a category to browse.
Read and query nodes 16
Deepgram: Get Billing Breakdown
Dollar spend for a project over a date range, optionally grouped by API key, deployment, line item, or tag. Where the usage breakdown node reports hours and tokens, this one reports money. Free read.
Deepgram: Get Member Scopes
Read the permission scopes a member holds on a project. Run it before changing scopes so you know the exact strings this project uses. Free read.
Deepgram: Get Model
Fetch metadata for one model by its uuid: canonical name, architecture, supported languages, version, and whether it supports batch or streaming. Free read.
Deepgram: Get Project
Fetch one Deepgram project by id: its name and whether it has opted out of the Model Improvement Program. Free read.
Deepgram: Get Project Balances
Read the outstanding credit balances on a Deepgram project, each with an amount and its units (such as USD). Since every transcription draws this down, this is the node for a low-balance alert automation. Free read.
Deepgram: Get Project Key
Fetch one API key record by id: its comment, scopes, tags, expiration date, and owning member. The secret key value is never returned, only shown once at creation time in the Deepgram console. Free read.
Deepgram: Get Request
Fetch one request record by id: which path was called, which API key made it, when, what status code came back, and whether a callback was used. This is the request LOG, not a transcript store, so do not rely on it to retrieve the output of a callback-based transcription. Free read.
Deepgram: Get Usage Breakdown
Usage totals for a project over a date range: audio hours, agent hours, tokens in and out, text-to-speech characters, and request counts, optionally grouped by API key, endpoint, model, method, tag, or deployment. This is the node to wire into a scheduled cost-watch automation. Free read.
Deepgram: List Models
List every Deepgram model available to the account, split into speech-to-text models (data.stt) and text-to-speech voices (data.tts) with their languages, versions, and uuids. Use it to populate the 'model' field on the transcription node with a value that is live today rather than one hard-coded months ago. Free read.
Deepgram: List Project Invites
List the outstanding invitations on a Deepgram project, each with the invitee's email and the scope they were invited with. Free read.
Deepgram: List Project Keys
List the API keys on a project for auditing: each entry carries the key id, its comment, its scopes, its creation date, and the member it belongs to. Deepgram does NOT return the secret key value here, so this is safe to run in an automation. Free read.
Deepgram: List Project Members
List the members of a Deepgram project with their member ids, email addresses, names, and scopes. Free read.
Deepgram: List Projects
List every Deepgram project the API key can see. Takes no configuration and costs nothing, so it doubles as the credential health check. The response carries only project ids and names, never key material. Most other management nodes need a project_id from here.
Deepgram: List Purchases
List the credit purchases on a project: the original amount of each order, its units, when it was created, and when it expires. Pair it with the balances node to see how much of each purchase is left. Free read.
Deepgram: List Requests
List the API requests a Deepgram project has served, with filters for date range, endpoint, method, status, and API key. The audit trail behind a billing question or a failing integration. Free read.
Deepgram: List Usage Fields
List which models, features, tags, and processing methods the project actually used in a date range. Run it before the usage breakdown node to learn the valid filter values instead of guessing them. Free read.
Action nodes 10
Deepgram: Analyze Text
COSTS MONEY ON EVERY RUN. Deepgram text intelligence is metered pay-per-use, billed by the tokens in the text you submit and the output produced, so a long document costs proportionally more than a short one. No rate is quoted here on purpose because rates change: see https://deepgram.com/pricing for the current metering. Run Deepgram text intelligence over plain text: summarization, topic detection, intent recognition, and sentiment scoring, in any combination. Feed it a transcript from the Deepgram: Transcribe Audio from URL node, an email body, a support ticket, or any other text. Supply either 'text' (the text itself) or 'url' (a URL Deepgram fetches the text from), never both.
Deepgram: Create Project Invite
Invite someone to a Deepgram project by email at a given scope. Sends them a real invitation email. Reversible with the Deepgram: Delete Project Invite node while it is still pending. Free.
Deepgram: Delete Project
PERMANENTLY DELETE a Deepgram project. This is not reversible: the project, its API keys, and its member assignments go with it, and any automation still using a key from that project stops working. Use only with a project id you are certain about.
Deepgram: Delete Project Invite
Revoke a pending invitation by email address. Classed as a reversible write rather than destructive because nothing is lost: send the invite again to restore it. It has no effect on someone who has already joined, which is the Deepgram: Remove Project Member node.
Deepgram: Delete Project Key
PERMANENTLY REVOKE a Deepgram API key. Not reversible: the secret cannot be recovered, a replacement key is a different secret, and anything still authenticating with the deleted key starts failing immediately. Useful as the last step of a credential-rotation flow.
Deepgram: Leave Project
Remove the account behind this API key from the project. Destructive in practice: once the key's owner has left, the same credential may lose access to the project entirely and cannot re-add itself; another project member has to send a fresh invite.
Deepgram: Remove Project Member
REMOVE a person from a Deepgram project. Destructive: the member loses access at once, and getting them back means a fresh invite they have to accept. Their API keys on the project stop working.
Deepgram: Transcribe Audio from URL
COSTS MONEY ON EVERY RUN. Deepgram is metered pay-per-use and speech-to-text is billed by audio duration, so a one-hour recording costs roughly sixty times what a one-minute clip costs. No rate is quoted here on purpose because rates change: see https://deepgram.com/pricing for the current metering. New accounts get a free starting credit. Transcribe a pre-recorded audio or video file with Deepgram speech-to-text. Audio reaches Deepgram as a URL, never as an upload: the canvas cannot produce a binary request body, so this node sends the documented JSON body {"url": "..."} and Deepgram fetches the media itself. The plain transcript is at data.results.channels[0].alternatives[0].transcript, with per-word timings and confidences beside it, plus optional speaker labels, paragraphs, utterances, summaries, topics, intents, and sentiment depending on which options you turn on. Live streaming transcription is a WebSocket protocol and is not available as an automation node.
Deepgram: Update Member Scopes
Set a member's permission scope on a project. Reversible: read the previous value with the Deepgram: Get Member Scopes node first and write it back to undo. Free.
Deepgram: Update Project
Rename a Deepgram project. Reversible: set the old name back to undo it. Free.
Connecting Deepgram takes one credential
Deepgram issues API keys from its console at console.deepgram.com, under a project's API Keys tab. The secret is shown once, at creation, so copy it while it is on screen. Give the connector key the lowest scope that covers the nodes you plan to run: the management write nodes need an admin or owner scope, while transcription and text intelligence do not. The key travels in an Authorization header whose scheme word is Token, not Bearer, which Deepgram's reference documents verbatim and requires on all 50 of its operations.
In the Deepgram console (console.deepgram.com), open the project you want to automate, go to API Keys, create a key with the scope your automations need, and copy the secret before you navigate away.
In Mattering Automations, open Connectors, click Add Connector, choose Deepgram, and paste the key. The form asks for one value and nothing else: there is no header name to type and no scheme to choose, because the connector builds the Authorization header itself.
Click Test Connection. Mattering calls Deepgram's projects endpoint, which costs nothing and returns only project ids and names, with no key material anywhere in its response.
All 26 Deepgram nodes appear in the canvas palette. Run Deepgram: List Projects first, because most of the management nodes need a project id from what it returns.
Common uses for the Deepgram MCP
Transcription that bills by the minute
Deepgram: Transcribe Audio from URL is metered pay-per-use and billed by audio duration, so a one hour recording costs proportionally more than a one minute clip. No rate appears here on purpose; deepgram.com/pricing carries the current metering. Audio reaches Deepgram as a URL and Deepgram fetches the media itself, so a public HTTPS link works and so does a time-limited signed link from object storage, as long as it is still fetchable at the moment the node runs. The flat transcript sits at data.results.channels[0].alternatives[0].transcript, with per-word timings, confidences, and optional speaker labels beside it.
Meaning on top of the words
Deepgram: Analyze Text runs summarization, topic detection, intent recognition, and sentiment scoring in any combination. It is a separate endpoint from transcription and it takes either the text itself or a URL to fetch text from, never an audio file, so it works just as well on an email body or a support ticket as it does on a transcript from the transcription node. It is metered by the tokens in and out, so a long document costs more than a short one.
A model list that is live today
Deepgram: List Models returns every model available to the account, split into speech to text models and text to speech voices, with their languages, versions, and uuids, and Deepgram: Get Model fetches one by uuid. Use it to fill the transcription node's model field with a value that exists now rather than one copied from a tutorial written months ago. The voices are visible here too, which is the honest answer to where Deepgram's voice catalog lives in this connector: speech synthesis itself is not a node.
Spend you can watch from the canvas
Deepgram: Get Usage Breakdown reports audio hours, tokens, and request counts across a date range, Deepgram: Get Billing Breakdown reports the same period in money, Deepgram: Get Project Balances reads what credit is left, and Deepgram: List Purchases shows what was bought and when it expires. A low balance alert is a two node flow: read the balance on a schedule and message someone when it drops under your floor. Run Deepgram: List Usage Fields first to learn which models, features, and tags the project actually used, instead of guessing filter values.
Key and member audits that never expose a secret
Deepgram: List Project Keys returns each key's id, comment, scopes, creation date, and owning member, and Deepgram does not return the secret value there, which is what makes the node safe to run on a schedule. Deepgram: List Project Members, Deepgram: Get Member Scopes, and Deepgram: List Project Invites complete the picture of who can reach a project. Creating an API key is deliberately not a node here, because Deepgram's create response carries the full secret and that would write a live key into an automation execution record.
The request log behind a billing question
Deepgram: List Requests filters the requests a project has served by date range, endpoint, method, status, and API key, and Deepgram: Get Request opens one of them: which path was called, which key made it, when, what status code came back, and whether a callback was used. That is the audit trail for a surprising invoice or an integration that started failing. It is the request log rather than a transcript store, so it is not a way to retrieve the output of a callback-based run.
Example: A finished recording becomes a transcript, a summary, and a task, with the balance checked first
Built on the Automations canvas in a few minutes, or described in plain language to Matter and built for you.
Trigger: a webhook or a scheduled check delivers a URL to a finished recording. A signed, time-limited link from object storage is fine, as long as it is still fetchable at the moment the flow runs.
Deepgram: Get Project Balances reads the remaining credit, and a condition stops the flow before it spends anything if the balance has fallen below the floor you set. This step is free.
Deepgram: Transcribe Audio from URL sends the documented JSON body and Deepgram fetches the media itself. This is the step that costs money, metered by audio duration, so a long recording costs proportionally more. Leave the callback field blank so the transcript comes back inline; the node carries a long timeout so a long file can still finish in one call.
Deepgram: Analyze Text takes the flat transcript from data.results.channels[0].alternatives[0].transcript and returns a summary, topics, intents, and sentiment. It is metered by tokens, and it is the second and last step in this flow that spends anything.
The transcript, the summary, and the speaker labels land where the work is: a task in Mattering Productivity with the summary in the description, a row in a Mattering database, or a message to a channel.
Everything else here is free, and a weekly Deepgram: Get Usage Breakdown on the same project turns the flow's spend into a number the team can see without opening the Deepgram console.
Deepgram MCP questions, answered
What Deepgram operations are supported?
26 nodes curated from the 50 operations in Deepgram's REST API: pre-recorded transcription from a URL, text intelligence, the model catalog, projects, API keys, members and their scopes, invites, request history, usage, billing, credit balances, and purchases. Deliberately not shipped: the two speech synthesis endpoints, because both return the audio as a binary stream; the Voice Agent configuration plane, which is served from a different host and whose runtime is a WebSocket conversation; self-hosted distribution credentials; a superseded usage summary endpoint that Deepgram's own description points away from; and API key creation, because Deepgram's create response contains the full secret and shipping it would write a live key into an execution record. Four nodes are flagged destructive: Deepgram: Delete Project, Deepgram: Leave Project, Deepgram: Delete Project Key, and Deepgram: Remove Project Member.
Does this work with Claude, ChatGPT, or other AI assistants?
Yes. Mattering exposes these nodes over MCP, an open protocol for connecting AI assistants to tools, so any MCP-capable assistant, including Claude and ChatGPT, can connect to your workspace and use them. Matter, the assistant built into the platform, uses the same nodes natively. The 16 read nodes are free management calls and are the safest surface to hand an agent; the two metered nodes spend real money on every run, so gate those the way you would any spending step.
Why does a valid Deepgram API key return 401?
Almost always the scheme word. Deepgram's Authorization header takes Token, not Bearer, and its reference documentation states that verbatim and repeats it as a required header on every operation. Anyone pattern matching from other AI APIs writes Bearer and collects a 401 with a key that is perfectly good. Inside this connector you never type the header at all: the credential is a single secret value and the connector formats Authorization with the Token scheme itself, so the remaining 401 causes are a revoked key or whitespace pasted along with the secret. The scheme only becomes your problem if you also call Deepgram from a generic HTTP node.
Can I upload an audio file instead of giving a URL?
No, and that is a design decision rather than a gap. Deepgram's documented request body for pre-recorded transcription is JSON with a single required url, and Deepgram fetches the media itself. A raw binary upload exists in Deepgram's prose guides, but an automation canvas cannot produce a binary request body: node configuration is JSON, and the alternative of inlining the file as a base64 string would put megabytes into a config record. So the audio has to be reachable at a URL Deepgram can fetch when the node runs. In practice that is how most pipelines already move recordings: a public HTTPS link works, and so does a signed or temporary URL from object storage.
Does this connector do text to speech?
No, and it is worth saying so plainly rather than letting anyone find out mid-build. Both of Deepgram's speech endpoints return the synthesized audio as a binary stream, and the callback variant delivers those same bytes to a webhook, so every available shape was a bad one: base64 in an execution record, a node that spends money and returns only headers, or audio posted to something that stores JSON. Rather than ship a node that disappoints, the connector ships none. What it can honestly offer is visibility: Deepgram: List Models returns the text to speech voices alongside the speech to text models, with their languages and versions.
Can I transcribe a live stream?
Not through this connector. Deepgram's live transcription is a WebSocket protocol, and an automation node makes one request and returns one response; it cannot hold a bidirectional socket open across a run. That puts live streaming out of scope for an HTTP node, and the transcription node's own description says so. Pre-recorded transcription of a finished file is fully supported, and for most workflows the file is finished by the time an automation is the right tool anyway.
What does this cost, and can it spend money by accident?
Two of the 26 nodes are metered and the other 24 are free management calls. Deepgram is pay-per-use: speech to text is billed by audio duration, so an hour of audio costs proportionally more than a minute, and text intelligence is billed by the tokens in and out. No rate is quoted here or in any node description, on purpose, because rates change and per-model rates differ; deepgram.com/pricing is the source. Both metered nodes lead their descriptions with the warning, so a spending step is never mistaken for a free one, and the connector also ships the tools to watch the meter: usage, billing, balances, and purchases are all readable from the canvas.
What happens if I set a callback on the transcription node?
The node's output changes shape, and it is important to know that before wiring the flow. With a callback set, Deepgram answers immediately with only a request_id and posts the finished result to your URL later, so the transcript will not be in the node's output and the automation has to continue from a webhook trigger listening on that URL. There is no polling alternative: the request log read returns the record of the call, not the body of the result. Leaving the callback field blank gives the ordinary synchronous call that returns the transcript inline, which is the default and the recommended path; the node carries a long timeout so a long file can finish that way.
Put Deepgram to work inside your workspace
Connect once, then use every node in visual automations and AI agent workflows across the Mattering platform.
All product names, logos, and brands are property of their respective owners and are used for identification purposes only. Use of these names, logos, and brands does not imply endorsement or affiliation. Deepgram is a product of its respective owner; Mattering is an independent platform.