LiveKit Agents

# Fix list: LiveKit Agents

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/livekit-agents, the October 2026 research run, assessed 17 October 2026. Grade BB, 72.5 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on LiveKit Agents: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Security & auth, 62 out of 100, up to 6.7 more on the total

Why it scored 62: Framework reading. Agents authenticate to a LiveKit server with an API key and secret that sign JWT access tokens with room grants and a lifetime. No analytics calls were found in the source. Uploads to LiveKit Cloud happen only when the agent connects to LiveKit Cloud media servers and the project enables Agent Observability, which the terms call a feature the customer elects to enable. `record` then defaults to the project setting, with `record=False` or per-category opt-outs, 30-day retention and PII redaction that is off by default. The default turn detector streams user audio to LiveKit Inference when deployed to LiveKit Cloud or in `dev` mode with credentials, which the docs state (20 of 30). No built-in approval step or read-only mode. `disallow_interruptions()`, `on_duplicate="confirm"`, MCP `allowed_tools`, prebuilt tasks with `require_confirmation` and the docs' advice to gate critical actions in code. Hosted agents run in sandboxed per-instance environments per the security page (10 of 20). Callers' speech and MCP results reach the model, and a search of llms-full.txt found no prompt-injection guidance (3 of 15). OpenTelemetry traces and metrics per session, session reports, and Agent Insights on LiveKit Cloud (14 of 15). security.txt with a contact and PGP key but no Expires field, a disclosure policy with safe harbour and a hall of fame, a SOC 2 Type II report for LiveKit Cloud under NDA, and no OSV advisories for livekit-agents or @livekit/agents on 17 October. No SECURITY.md in the repository and no paid bug bounty found. CI pins its actions by commit (15 of 20).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-security):

- 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option.
- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.
- 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10.
- 0 to 15, audit logs or per-call visibility for the operator.
- 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public.

Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing.

## 2. Reliability, 67 out of 100, up to 6.6 more on the total

Why it scored 67: Graded as a framework with the local-software lines, since LiveKit Cloud is optional and not graded. Official packages on PyPI (Python 3.10 to 3.14 declared) and npm (Node.js 20 or later per the docs) (20). Public workflows run on every push to main and on pull requests. Unit tests on Linux with Python 3.12 over 256 test files, ruff, and mypy on Python 3.10 and 3.13, with a small blockguard suite on Linux, macOS and Windows. GitHub's pages and API weren't readable from our session, so a passing main is not confirmed (15 of 25). 705 open issues and pull requests on 13 October per the repos.ecosyste.ms mirror. Of the 100 most recent issues (31 August to 16 October 2026), 68 were open, 41 of those with no comment, and 16 open ones carried the bug label. In the past year 167 of 369 issues were closed, in 65 days on average (10 of 25). No CHANGELOG file. Release notes are on GitHub, which we couldn't read, and the repository's release-notes script asks for Breaking changes and Security sections. Migration guides exist for 0.x to 1.0, AGENTS.md makes patch bumps the default, and 2.0.0b1 reached PyPI on 12 October with no 2.0 migration guide found (7 of 15). 1.8.8, stable since 1.0 (15).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability):

Hosted APIs, MCP servers, models and platforms.

- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).
- 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so.
- 15, rate limits documented with numbers.
- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.
- 10, an SLA published for any paid tier.
- 10, the surface agents use is generally available, not beta or preview.

Local packages, SDKs, frameworks and stdio MCP servers.

- 20, installs from an official package with supported runtimes stated.
- 25, a public CI and test suite, passing on the default branch.
- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).
- 15, semver discipline and breaking changes called out in a changelog.
- 15, version 1.0 or later, or declared stable.

Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors.

## 3. Payments & pricing, 60 out of 100, up to 5 more on the total

Why it scored 60: No x402, MPP or L402 (0 of 40). Read with the rule for open-source software that has a paid hosted option. LiveKit Cloud publishes per-minute prices for agent sessions and recordings and per-unit prices for every LiveKit Inference model without a login (20). The Apache-2.0 packages install with no card, and LiveKit Cloud's Build plan is $0 with no card (20). The framework and a LiveKit server can both run self-hosted with no account or sign-up (20).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments):

The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/).

- 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which.
- 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for "contact sales" or prices behind a login.
- 20, a free tier or trial that doesn't need a card.
- 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API).

Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied.

Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol.

## 4. Agent ergonomics, 80 out of 100, up to 3.3 more on the total

Why it scored 80: Framework reading. An MCP server is about ten lines with `mcp.MCPToolset`, with `allowed_tools`, `filter_tools()` and a `tool_result_resolver` to trim results before they reach the model, but Python only (20 of 25). Tools can be added, removed or replaced per agent and per state, flags such as `IGNORE_ON_ENTER` hide a tool at the greeting, and handoffs swap the active toolset (16 of 20). `ToolError` returns a message the model can act on, invalid arguments come back as a `ToolError` naming the tool, and by default a session closes only after more than three unrecoverable STT, LLM or TTS errors of one kind (17 of 20). Interrupted tools finish in the background and keep their result in history, `disallow_interruptions()` protects tools that write, and `on_duplicate` can reject, replace or confirm a second call. Duplicates are matched by tool name only, and there are no idempotency keys (14 of 20). `AgentSession` supplies VAD and a turn detector by default and LiveKit Inference takes model strings with one key, in Python and Node.js, but every agent needs a LiveKit server and three environment variables (13 of 15).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics):

- 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries).
- 20, pagination, filtering and output-size controls.
- 20, actionable, documented error responses, codes and messages an agent can recover from.
- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.
- 15, sensible defaults, few required parameters, and official SDKs in at least two languages.

Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs.

## 5. Schema & documentation, 81 out of 100, up to 3.1 more on the total

Why it scored 81: Library reading. A typed Python package (`py.typed`, mypy in CI) with a Python and Node.js API reference on docs.livekit.io. The llms.txt link to the Python reference's Markdown returns 404, while the HTML reference loads (22 of 25). llms.txt with section indexes, `.md` on every page, llms-full.txt and a docs MCP server (10). Pages say when to use a realtime model or a pipeline, when to choose `v1` or `v1-mini`, when to disallow interruptions and how to gate critical actions, and mark MCP as Python only (16 of 20). Function tools take their schema from typed signatures and docstrings, raw JSON Schema is accepted, and argument validation failures go back to the model as `ToolError` (13 of 15). 20 example folders in the repository, 39 recipes, and an events and error-handling reference (13 of 15). No changelog in the repository or docs, release notes only on GitHub, and the quickstart's `agent.py` still ends with `agents.cli.run_app(server)`, which 1.8.8 marks as deprecated (7 of 15).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema):

APIs and MCP servers.

- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).
- 10, llms.txt or Markdown docs served for agents.
- 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference.
- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.
- 0 to 15, examples and documented error responses.
- 15, versioning and a public changelog.

Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference.

## 6. Transparency & trust, 82 out of 100, up to 1.6 more on the total

Made of editorial 68, provenance 95.

Why it scored 82: Apache-2.0 for the framework and its 79 plugin packages. The turn-detector models are under LiveKit's proprietary model licence, and the core package now requires `livekit-local-inference`, whose metadata reads "Apache-2.0 AND LicenseRef-LiveKit-Model" with no source repository named. The README calls the project "Fully open-source" and names the model licence in its licence section, and `MODEL_LICENSE` itself calls LiveKit Agents a proprietary framework (20 of 30). The terms, the observability docs and the pricing page agree on 30-day observability retention, zero data retention for LiveKit Inference and no training on customer data, and a DPA and sub-processor list are published. The privacy policy (15 October 2026) excludes data LiveKit processes for customers. The pricing page lists LiveKit Inference models from Fish Audio, Gradium and Speechmatics, which the sub-processor list of 10 September 2026 does not name (22 of 30). Deprecations are flagged in code and docs (`mcp_servers`, `cli.run_app`, the text turn-detector plugin), mostly without dates. One warning names 30 April 2026 for removing `api_version` and is still in 1.8.8, and 2.0.0b1 has no migration notes we could find (9 of 20). What reaches LiveKit is documented with opt-outs (`record`, `TurnDetector(version="v1-mini")`), and the sub-processor list names each entity and country (17 of 20).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency):

- 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms.
- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).
- 0 to 20, a deprecation policy or notices with dates.
- 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted).

The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two.

Provenance checks not met in full (half of this category, computed from checked facts):

- Terms of service: read, states 6 of the 7 things a reader expects, and has 1 clause that costs points (7.1 of 10)
- Privacy policy: read, states 6 of the 8 things a reader expects (8.5 of 10)

## 7. Maintenance & community, 85 out of 100, up to 1.3 more on the total

Why it scored 85: `livekit-agents` 1.8.8 on PyPI on 15 October 2026 (30). 17 PyPI releases in the 90 days to 17 October, from 1.6.7 on 25 July to 1.8.8, and 14 npm releases of `@livekit/agents` (20). 563 commits by 173 authors since 19 July. Per the issues.ecosyste.ms mirror, past-year pull requests were closed in 16 days on average (570 of 1,085), past-year issues in 65 days (167 of 369), and 41 of the 68 open issues among the latest 100 had no comment. CODEOWNERS and an auto-assign workflow are in place (13 of 25). Python 1.8.8 and Node.js 1.9.3 were both released on 15 October (15). Renovate, a uv lockfile that holds new dependencies back seven days, pinned actions and CI on every push, with the pass state unconfirmed (7 of 10).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance):

- 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older.
- 20, at least three releases or dated changelog entries in the last 90 days.
- 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15.
- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).
- 10, package health, current dependencies and CI.

Models are read for deprecation notice periods and model churn rather than release counts.

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- Judgement call on category. Listed in voice-agents, as the brief suggested, because its main job is running a spoken conversation (pipeline or speech-to-speech), which puts it in head-to-heads with Pipecat and the hosted voice platforms. It is also an agent framework with MCP and handoffs, so `alsoIn: ["frameworks"]` could be added at landing; we didn't add it.
- Graded on the Python package `livekit-agents`, with `@livekit/agents` for Node.js as the second SDK. LiveKit Cloud and LiveKit Inference are described in `details` with their published prices and are not graded, except that the Payments lines read the paid option's pricing, as the checklist says for open-source software with a hosted offering.
- The format's only link between related listings is the computed same-company list, which groups listings by `provenance.domain`. This listing uses livekit.com, as livekit-telephony does, so the two link to each other. LiveKit Telephony is also named in `notable` and the Telephony detail.
- firstReleased is 2023-11-14, the first upload of `livekit-agents` (0.0.1) per the packages.ecosyste.ms mirror of PyPI. The repository agrees. version.py was set to 0.0.1 on 20 October 2023 and bumped to 0.0.2 in "Cut releases for agents, plugins-core, and plugins-vad" on 15 November 2023. The earliest git tag is `livekit-agents@0.8.0`, on 24 July 2024.
- No deduction. The README calls the project "Fully open-source" while the core package requires `livekit-local-inference`, whose metadata adds a LiveKit model licence. The README names the model licence in its own licence section, so the claim is qualified, and transparency carries the point instead.
- `MODEL_LICENSE` defines LiveKit Agents as "the proprietary LiveKit software framework", while the framework's LICENSE, NOTICE and package metadata are Apache-2.0. Recorded as an inconsistency in LiveKit's documents, not a change of licence.
- unchecked: the source of `livekit-local-inference`. Its metadata names no repository, and the package itself was not downloaded.
- Read in source, not tested. The default turn detector picks `v1` in `dev` mode whenever `LIVEKIT_API_KEY` and `LIVEKIT_API_SECRET` are set, and sends them to agent-gateway.livekit.cloud unless `LIVEKIT_INFERENCE_URL` is set. The docs describe this as `dev` mode with LiveKit Cloud credentials. What happens with a self-hosted server's keys was not established.
- unchecked: whether Agent Observability is on for new LiveKit Cloud projects. The docs say it must be enabled in project settings and the terms call it a feature the customer elects to enable, but the dashboard wasn't opened.
- unchecked: whether Fish Audio, Gradium and Speechmatics models on LiveKit Inference run through a host already on the sub-processor list (Baseten, Modal and others are named) or need entries of their own.
- unchecked: GitHub releases (the only release notes), whether CI passes on main, and reply times on issues. WebFetch refused every page in this session, and github.com and its API are blocked to our shell. Stars and issue figures come from the repos.ecosyste.ms and issues.ecosyste.ms mirrors, synced on 13 and 16 October 2026.
- WebFetch refused every URL in this session, so first-hand pages were read with curl after checking robots.txt on each host (docs.livekit.io, livekit.com, status.livekit.io, huggingface.co, pypi.org, the ecosyste.ms hosts; registry.npmjs.org and pypistats.org serve none). status.livekit.io disallows /api/, so its history page was read. PyPI's project page answered with a JavaScript client challenge and was not read.
- Lapse, corrected. PyPI's JSON API (pypi.org/pypi/livekit-agents/json), which pypi.org's robots.txt disallows, was fetched in the same command that read the robots.txt. The response was deleted and nothing in these files cites it. Release dates come from PyPI's RSS feed and the packages.ecosyste.ms mirror instead.
- We did not install or run the framework. Behaviour described here comes from reading the source at commit f548149 and the docs as rendered on 17 October 2026.

## Weaknesses

- The turn-detector models are under a proprietary licence that forbids use outside LiveKit Agents and training other models on their output
- No CHANGELOG file, release notes only on GitHub, and 2.0.0b1 on PyPI since 12 October 2026 with no migration guide found
- 705 open issues and pull requests per a GitHub mirror, and 41 of the 68 open issues among the latest 100 had no comment
- No built-in approval step for tool calls, and no prompt-injection guidance found in the docs
- In `dev` mode with LiveKit Cloud credentials, the default turn detector streams user audio to LiveKit Inference

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Set `LIVEKIT_URL`, `LIVEKIT_API_KEY` and `LIVEKIT_API_SECRET` before starting. Every agent joins a LiveKit room, on LiveKit Cloud or a self-hosted server
- Pass `inference.TurnDetector(version="v1-mini")` to keep turn detection local, and read `MODEL_LICENSE` before using LiveKit's turn models anywhere but LiveKit Agents
- Pass `record=False`, or a dict per category, to `session.start()` if audio, transcripts, traces and logs must not reach LiveKit Cloud
- Call `context.disallow_interruptions()` first in tools that write, and set `on_duplicate="reject"` or `"confirm"`. Duplicates are matched by tool name only, not arguments
- Wrap MCP servers in `mcp.MCPToolset` with `allowed_tools` (Python only). The `mcp_servers` parameter is deprecated

## What the review panel asked for

- a 2.0 migration guide
- dated deprecation windows
- Update the quickstart
- Fix the llms.txt reference link
- Idempotency keys for tools
- Documented limit-hit behaviour
- tool approval primitive
- argument-aware duplicate matching

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.