# Deepgram L3 · Does the work Deepgram sells speech-to-text, text-to-speech and voice agents. **Can an AI agent use Deepgram? Yes, without signing up.** Our agent got into Deepgram on its own in 20 minutes and did 5 of its 9 core tasks. Category: Speech and voice AI · Connects by: API, CLI, SDK · Agent payments: None found · Driven: 1 Nov 2026 **How to connect** - MCP server URL: `https://api.dx.deepgram.com/kapa/mcp` (documented). Source: https://api.dx.deepgram.com/kapa/mcp - API base URL: `https://agent.deepgram.com` (documented). Source: https://developers.deepgram.com/openapi.json - CLI install: `brew install deepgram` (documented). Source: https://developers.deepgram.com/developer-tools/cli/installation - Docs page: `https://developers.deepgram.com/home` (tested in our drive). Source: https://developers.deepgram.com/home **Level 3: Does the work.** An agent gets in on its own and does half the core work or more. - Got in by: No key needed (20 minutes, no account) - Got as far as: Use: 5 of 9 core tasks (No payment made) - To reach L4: Do 80% of the core work (now 56%). 1 drive, awaiting confirmation [What the levels mean](https://agentdrives.com/method/levels?via=md) ## From finding it to leaving What an agent managed without a person, at each step. 1. **Find**: On its own. Docs it can act on: llms.txt and an OpenAPI. (373 doc pages read) 2. **Get in**: On its own. Got in without signing up. Email and password: stopped by a CAPTCHA. (20 minutes) 3. **Use**: Partly. Did: list every speech and voice model over the API without a key; transcribe sample calls and the agent's own streamed audio in the free playground. Not done: transcribe own audio files or call the API. (5 of 9 core tasks) 4. **Pay**: Not reached. Paying needs an account. 5. **Leave**: On its own. Nothing to cancel, made nothing to delete. ## Getting in 1 of 8 ways in worked without a person. The way that worked: **No key needed**, 20 minutes. Needs a browser. No inbox or person needed. - Agent on a server (API, MCP, CLI): No. Stops at the first API call: it needs a browser. - Assistant in its own browser (Like ChatGPT agent): No. Signing up needs a person: it asks for a bot check. - Signed agent (Known to Cloudflare): Yes. The sign-up page opens. | Way in | What happened | |---|---| | No key needed | Worked | | Email and password | Stopped by a CAPTCHA | | CLI code login | Stopped: The consent page needs an account | | Sign-in for apps (OAuth, MCP) | Stopped: The consent page needs an account | | Runs locally, no account | Not tried: Needs a person's account | | Continue with GitHub | Not tried: Not part of this drive | | Continue with Google | Not tried: Not part of this drive | | Continue with Microsoft | Not tried: Not part of this drive | [How agents get past bot checks](https://agentdrives.com/guides/let-signed-agents-past-your-bot-check.md?via=md) ## What an agent can do with Deepgram 5 of 9 core tasks done, and 12 of 19 others. - Reliability: core tasks that worked on all 3 runs: 76 of 100 - Handles mistakes: clear errors, retries without duplicates, honest limits: 70 of 100 - Surface tested: documented routes, tools and commands run and checked: 13 of 100 - Docs accuracy: documented operations that behaved as the docs say: 84 of 100 - Docs completeness: how fully the docs describe each operation: 51 of 100 Core tasks: | Task | Result | Done through | Tried, not done | Offered, not tried | |---|---|---|---|---| | Transcribe streaming audio in real time | Done | – | – | CLI | | Transcribe a pre-recorded audio or video file | Done | – | – | CLI | | Convert text to speech in a single request | Done | – | – | CLI | | Run a real-time voice agent conversation | Done | Web app | – | – | | Detect speaker turns and interruptions during a call | Done | Web app | – | – | | Stream text to speech over WebSocket | Blocked: The WebSocket needs a key. | – | API | – | | Dictate text into any Mac app | Blocked | – | Web app | – | | Record a meeting without a bot joining the call | Blocked | – | Web app | – | | Get a daily brief of calendar, email and messages | Blocked | – | Web app | – | 12 of 19 other tasks done: - Diarize speakers in a transcript: Done - Redact sensitive information from a transcript: Done - Prompt custom vocabulary with keyterms: Done - Summarize a transcript or block of text: Done - Analyze sentiment of a transcript or text: Done - Detect topics in a transcript or text: Done - Recognize speaker intent in a transcript or text: Done - Detect named entities in audio: Done - Automatically detect or switch spoken language mid-conversation: Done - Choose the speech or voice model: Done - Get a per-turn latency report from a voice agent: Done - List available speech and voice models: Done - Self-host Deepgram models in your own VPC or on-prem: Paid plan - Bring your own LLM or TTS provider to a voice agent: Blocked - Register a callback URL for async transcription results: Blocked - Ask the meeting transcript a question: Blocked - Learn a user's vocabulary and writing style over time: Blocked - Create an API key: Blocked - Invite a teammate to a project: Blocked - Read usage and billing balance: Blocked The plans they need: - Self-host Deepgram models: Enterprise Paid-plan tasks don't count toward the level. The drive used none (no free access: sign-up behind a visible reCAPTCHA); keyless playground only. What it can't do, and why: - Create an account: The sign-up shows a visible CAPTCHA checkbox. - Get an API key: Keys come only from a Console account. - Transcribe own audio files or call the API: The API needs a key; file upload needs sign-in. - Stream text to speech over WebSocket: The WebSocket needs a key. - Use Saga: It is a Mac and iOS app with no web or API route. - Pay: Buying happens in the Console, behind the CAPTCHA sign-up. **Surface audit:** 165 documented operations, 28 run, 22 passed, 4 not as documented, 137 blocked (sandbox or plan limits). ## How an agent connects API, OpenAPI, CLI, SDK and llms.txt found. Agent surfaces: API tested · MCP tested · CLI tested · SDK tested · Docs tested · Agent payments none found. Issues found: 30. - **MCP server**: 1 tested. A server for searching the docs only. (3 tools across its docs) - **API**: 3 tested. REST + WebSocket, 50 OpenAPI operations, 5 AsyncAPI channels. (55 operations across its docs) - **OpenAPI**: Yes. 3.1.0; responses list only 200 and 400. - **CLI**: 22 tested. deepctl 0.4.0 (dg), 23 commands, device-flow login. (52 commands across its docs) - **SDK**: 1 tested. Python deepgram-sdk 7.16.0, JS @deepgram/sdk 5.17.0, plus Go, .NET, Java, Rust. Packages: PyPI [deepgram-sdk](https://pypi.org/project/deepgram-sdk/), npm [@deepgram/sdk](https://www.npmjs.com/package/@deepgram/sdk), PyPI [deepctl](https://pypi.org/project/deepctl/). (48 methods across its docs) - **llms.txt**: Yes. https://deepgram.com/llms.txt - **Agent payments**: None found. The checkout wasn't reached: Pay as you go. - **Docs**: No gaps in what we read. Covers getting started, auth, error codes and code examples. 0 of 9 documented ways to do the core work worked when tried. https://developers.deepgram.com/home (373 of 660 pages read) [The standards](https://agentdrives.com/standard) ## Tips for agent builders - Read the docs as Markdown: append .md to any developers.deepgram.com page, or send Accept: text/markdown - Use https://developers.deepgram.com/openapi.json and asyncapi.json for exact request shapes - GET https://api.deepgram.com/v1/models lists models without a key - For CI, pass DEEPGRAM_API_KEY instead of `dg login`; `dg mcp` exposes the API to MCP clients with that key - Try features first in playground.deepgram.com: it shows the exact query parameters (e.g. summarize=v2, keyterm, language=multi) - Without a key, a browser agent can still use the playground's streaming and voice agent pages by giving the tab its own audio as the microphone - Check a model name against GET /v1/models first: a streaming session with an unknown model just fails to connect ## How Deepgram compares On our standard (27 criteria, each against every product we measured on it): - [Typed, tight input schemas](https://agentdrives.com/standard#mcp-06): **Ahead of most** 50 of 65 others lower, 15 the same. - [Docs an agent can navigate](https://agentdrives.com/standard#docs-11): **Ahead of most** 98 of 141 others lower, 43 the same. - [Spend an agent can see and cap](https://agentdrives.com/standard#pay-05): **Ahead of most** 85 of 118 others lower, 27 the same, 6 higher. - [The best way in is built for agents](https://agentdrives.com/standard#in-12): **Ahead of most** 98 of 147 others lower, 49 the same. - [Results sized for a context window](https://agentdrives.com/standard#mcp-09): **Ahead of most** 16 of 31 others lower, 15 the same. - [A tool list that fits](https://agentdrives.com/standard#mcp-11): **Ahead of most** 33 of 65 others lower, 32 the same. - [Truthful annotations](https://agentdrives.com/standard#mcp-07): **Level with most** 33 of 65 others the same, 32 lower. - [Sign-in for agents acting for a person](https://agentdrives.com/standard#in-08): **Level with most** 64 of 119 others the same, 55 lower. - [Versions an agent can pin](https://agentdrives.com/standard#api-07): **Level with most** 59 of 97 others the same, 38 lower. - [Answers sized for an agent](https://agentdrives.com/standard#api-04): **Level with most** 61 of 85 others the same, 24 lower. - [Prices an agent can read](https://agentdrives.com/standard#find-06): **Level with most** 80 of 105 others the same, 25 lower. - [Errors a program can read](https://agentdrives.com/standard#err-02): **In the middle** 40 of 104 others lower, 43 the same, 21 higher. - [Fast to the first call](https://agentdrives.com/standard#in-03): **Level with most** 71 of 85 others the same, 14 lower. - [Every way in works for an agent](https://agentdrives.com/standard#in-11): **In the middle** 56 of 145 others lower, 51 the same, 38 higher. - [Tool names that sort themselves](https://agentdrives.com/standard#mcp-03): **Level with most** 61 of 65 others the same, 4 lower. - [A path written for agents](https://agentdrives.com/standard#docs-10): **Level with most** 71 of 141 others the same, 39 lower, 31 higher. - [Real status codes](https://agentdrives.com/standard#err-01): **Level with most** 125 of 127 others the same, 2 lower. - [Errors say what to do next](https://agentdrives.com/standard#err-03): **Level with most** 56 of 103 others the same, 24 lower, 23 higher. - [WebMCP tools on the web app](https://agentdrives.com/standard#proto-02): **Level with most** 91 of 93 others the same, 2 higher. - [A request id on every answer](https://agentdrives.com/standard#err-07): **In the middle** 34 of 103 others lower, 29 the same, 40 higher. - [Agents pick the right tool](https://agentdrives.com/standard#mcp-05): **Level with most** 44 of 65 others the same, 1 lower, 20 higher. - [All validation problems at once](https://agentdrives.com/standard#err-04): **Level with most** 23 of 35 others the same, 12 higher. - [Pages fit one fetch](https://agentdrives.com/standard#docs-08): **Level with most** 87 of 141 others the same, 54 higher. - [One way to page through lists](https://agentdrives.com/standard#api-03): **Behind most** 53 of 87 others higher, 34 the same. - [Structured results](https://agentdrives.com/standard#mcp-08): **Behind most** 22 of 36 others higher, 14 the same. - [Export before leaving](https://agentdrives.com/standard#leave-03): **Behind most** 89 of 144 others higher, 55 the same. - [Descriptions that say when to use the tool](https://agentdrives.com/standard#mcp-04): **Behind most** 46 of 65 others higher, 19 the same. [How each criterion is graded](https://agentdrives.com/standard) ## For the team behind Deepgram **7 things to fix for agents.** Agent blockers: 3 found · Agent surfaces: 27 issues found. Safety review: 2 issues found. Our agent got through Use only in part, and didn't reach Pay. Prove Deepgram is yours, then see where each issue is, read your safety findings and ask for a free drive after a fix. How to fix each one comes with the Audit. - Find: Clear - Get in: Clear - Use: Partly - Pay: Not reached - Leave: Clear [Claim Deepgram, free](https://agentdrives.com/app/products?add=deepgram.com&ref=listing) · [What a drive covers](https://agentdrives.com/method?via=md) ## Where Deepgram fits - Category: [AI and agent infrastructure](https://agentdrives.com/categories/ai?via=md), [Speech and voice AI](https://agentdrives.com/categories/speech?via=md) - Wall: [Bot checks](https://agentdrives.com/access/get-in/bot-checks) (our agent got in another way) - Vendors: [Google reCAPTCHA](https://agentdrives.com/vendors/recaptcha) (on one way in) - Standards: [OpenAPI](https://agentdrives.com/standards/openapi) (its API, described), [llms.txt](https://agentdrives.com/standards/llms-txt) (its docs, listed for agents), [MPP](https://agentdrives.com/standards/mpp) (machine payments) - Steps: [Find](https://agentdrives.com/access/find), [Get in](https://agentdrives.com/access/get-in), [Use](https://agentdrives.com/access/use), [Pay](https://agentdrives.com/access/pay), [Leave](https://agentdrives.com/access/leave) Data: [JSON](https://agentdrives.com/api/tools/deepgram.json?via=md) · [Markdown](https://agentdrives.com/tools/deepgram.md?via=md) · [History](https://agentdrives.com/tools/deepgram/history?via=md) · [Atom](https://agentdrives.com/feeds/tools/deepgram.atom) · [JSON Feed](https://agentdrives.com/feeds/tools/deepgram.json) · licence [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) **Found an error?** [Tell us on this page](https://agentdrives.com/tools/deepgram?via=md#correct) or read [how corrections work](https://agentdrives.com/corrections?via=md).