npm package discovery and stats viewer.

Discover Tips

  • General search

    [free text search, go nuts!]

  • Package details

    pkg:[package-name]

  • User packages

    @[username]

Sponsor

Optimize Toolset

I’ve always been into building performant and accessible sites, but lately I’ve been taking it extremely seriously. So much so that I’ve been building a tool to help me optimize and monitor the sites that I build to make sure that I’m making an attempt to offer the best experience to those who visit them. If you’re into performant, accessible and SEO friendly sites, you might like it too! You can check it out at Optimize Toolset.

About

Hi, 👋, I’m Ryan Hefner  and I built this site for me, and you! The goal of this site was to provide an easy way for me to check the stats on my npm packages, both for prioritizing issues and updates, and to give me a little kick in the pants to keep up on stuff.

As I was building it, I realized that I was actually using the tool to build the tool, and figured I might as well put this out there and hopefully others will find it to be a fast and useful way to search and browse npm packages as I have.

If you’re interested in other things I’m working on, follow me on Twitter or check out the open source projects I’ve been publishing on GitHub.

I am also working on a Twitter bot for this site to tweet the most popular, newest, random packages from npm. Please follow that account now and it will start sending out packages soon–ish.

Open Software & Tools

This site wouldn’t be possible without the immense generosity and tireless efforts from the people who make contributions to the world and share their work via open source initiatives. Thank you 🙏

© 2026 – Pkg Stats / Ryan Hefner

@leiutils/mcp-transcribe

v0.11.0

Published

MCP server for LeiUtils: transcribe local audio/video files into a workspace matter, fetch and search the transcript, and record and discuss evidence from it.

Readme

@leiutils/mcp-transcribe

A local MCP server that lets an MCP client (e.g. Claude Code) hand a local audio/video file to LeiUtils, transcribe it into a workspace matter, and retrieve and search the transcript, and record evidence from it.

It runs as a stdio server that the client spawns via npx, reads the file off local disk, uploads it to LeiUtils in chunks, and exposes the job lifecycle as 13 MCP tools. It is thin transport + upload glue — the LeiUtils server remains authoritative for tenant isolation, billing, retention, and audit.

Tools

| Tool | Purpose | |------|---------| | list_matters(q?, cursor?) | List the workspace's matters: { matters: [{ id, name }], next_cursor }, one page per call. Choose a matter_id here before submitting. Optional q searches by name. | | list_transcriptions(matter_id?, status?, cursor?) | List the workspace's transcriptions, newest first: { transcriptions: [{ id, label, original_filename, status, stable_code, transcript_state, duration_seconds, created_at, matter, evidence_count }], next_cursor }, one page per call. Show label, else original_filename. evidence_count shows which transcriptions have evidence (null without the Evidence feature). | | submit_media_transcription(file_path, matter_id, language?, filename?) | Upload a local file and start a transcription. Returns { job_id, status, stable_code } once the upload is accepted. Transcription then runs asynchronously. | | get_transcription_status(job_id) | Return { status, stable_code }. Poll until stable_code is completed; processing_failed and rejected are final. | | fetch_transcript(job_id) | Return the transcript for a completed job; a not_ready error while it is still processing; a final job_failed / job_rejected error if it never will. | | search_transcript(job_id, q, limit?) | Search a completed transcript; returns ranked matches with segment anchors. Same not_ready / job_failed / job_rejected errors as fetch. | | search_matter(matter_id, q, limit?) | Search a matter's reviewed transcripts at once (the newest 200; searched_truncated when more exist): up to 25 ranked hits with job_id and segment anchors, the transcripts searched, and the unreviewed ones skipped (skipped_unreviewed_count). | | list_evidence(job_id, cursor?) | List a transcript's evidence, oldest first: titles, quotes, quote_state, attribution and marker_url, one page per call. | | create_evidence(job_id, seg_idx_start, seg_idx_end, title) | Create evidence from a range of a reviewed transcript. Returns a content-free acknowledgment with marker_url. | | get_evidence_discussion(evidence_id) | Return an evidence item's comments, oldest first, with one level of replies and their links. | | add_evidence_comment(evidence_id, body, parent_comment_id?, link_evidence_ids?) | Comment on evidence (or reply to a top-level comment), linking up to 10 other evidence items. | | play_evidence(evidence_id) | Play a saved evidence item's audio (exactly its stored range, at most 120 s) on your computer. Returns only { result_type, evidence_id, played, marker_url }; the audio is never returned to the model. On failure: an error code and the evidence's marker_url. | | stop_playback() | Stop the clip play_evidence is playing. Returns { result_type, stopped, evidence_id? }. Local only: no token, nothing sent to LeiUtils. |

Finding transcriptions and matters (0.6.0)

list_transcriptions reaches transcriptions you did not submit yourself. status is one of pending, queued, submitting, processing, reviewing, completed, failed, rejected (anything else is rejected before a request is made). transcript_state is reviewed or machine_unreviewed for a completed job and null otherwise. A response that does not match the documented shape fails closed with list_transcriptions_invalid_response.

marker_url (on create_evidence and list_evidence) opens the evidence in LeiUtils, with the player at its start; give it to the user. The connector accepts it only when it is exactly <origin of LEIUTILS_API_BASE_URL>/jobs/<job_id>?marker=<id>; any other value is an invalid response (on create_evidence, write_outcome_unknown).

list_matters accepts an optional q to search by name. Rules for q:

  • At most 100 characters, counted as Unicode code points (an emoji counts as one). A longer q is rejected with invalid_q before a request is made.
  • An empty q is ignored. Whitespace is sent as given; the server trims it.
  • A cursor is not bound to q: pass a cursor only together with the same q that produced it.

Evidence (0.7.0)

The evidence tools need a LeiUtils server with the integration evidence API, which exists only on servers running #887 PR-2: staging today, not production yet. They also need a workspace with the Evidence feature (Firm plan). Without the feature, every evidence call answers forbidden.

  • create_evidence works on reviewed transcripts only (not_reviewed otherwise). Take seg_idx_start / seg_idx_end from a search_transcript match whose transcript_state is reviewed, or from seg_idx in a reviewed fetch_transcript. The acknowledgment's seg_idx_* may be wider than requested.
  • add_evidence_comment replies only to a top-level comment; a reply to a reply is reply_depth_exceeded. If any of the link_evidence_ids is not found, nothing is written and the error is linked_evidence_not_found, whose evidence_ids names the ids that were not found.
  • Title and body limits, blank values and range order are checked by the server (validation_failed with fields, or range_invalid).
  • Evidence made through these tools shows in LeiUtils as made "via" the MCP client, using the name and version the client reports when it connects (printable ASCII only, at most 100 characters).
  • A response that does not match the documented shape fails closed: list_evidence_invalid_response, get_evidence_discussion_invalid_response, or write_outcome_unknown for a write.

Playing evidence (0.9.0)

play_evidence needs the audio:clip ability and a workspace with the Evidence feature (Firm plan); the evidence must be on a reviewed transcript. Each play is written to the workspace's audit log, and the Evidence board shows who a clip was played to.

  • The clip is a WAV file, saved to a private folder in your operating system's temporary folder (leiutils-clip-…/clip.wav) and played by the system's own player, then deleted as soon as the player has exited, also when it fails or you cancel. Neither the audio nor the file's path is ever sent to the model.
  • Players: macOS afplay; Windows PowerShell Media.SoundPlayer (the path is passed in an environment variable, never in the command); Linux paplay, or aplay if paplay is missing. No player found is no_player; a player that fails is failed. Nothing is run through a shell.
  • Fallback: on any failure the result carries the evidence's marker_url, which opens the transcript in LeiUtils on that evidence, where you can press Play.
  • Stopping (#1064): say "stop" and the agent calls stop_playback. Pressing stop or Esc in Claude Desktop ends the chat turn but does not cancel the running tool, so the clip would otherwise play to its end. One clip plays at a time: a new one stops the previous one.
  • A clip over 120 s is clip_too_long; more than 10 plays a minute on one token is rate_limited.

Writes, retries and unknown outcomes

  • Each create_evidence / add_evidence_comment call generates one Idempotency-Key and uses it for every attempt of that call. The key is never returned, logged or reused by another call.

  • A lost answer (no response, a timeout, a body that cannot be read) or a 500, 502, 503 or 504 is retried with the same key and the same body: at most 3 attempts, 0.5 s then 1 s apart, all within 2 minutes of the first, each attempt limited to 30 s. Nothing is sent after the 2 minutes, even if the machine slept in between.

  • Never retried: a malformed success, an unexpected status (such as a 3xx, 501 or 505), 409 idempotency_conflict, 410 resource_deleted and every other 4xx or 429.

  • replayed: true means an earlier attempt of the same call was saved; it is not a second item.

  • When the outcome cannot be known, the result is write_outcome_unknown with a reason:

    | reason | Meaning | |----------|---------| | invalid_response | The server answered success but not in the documented shape, or answered an unexpected status (such as a 3xx, 501 or 505). | | retries_exhausted | Every attempt was lost or failed with a retryable status. | | deadline_exceeded | The 2 minutes passed after a lost or failed attempt. | | retry_refused | A retry was refused (for example 401 or 404) after an earlier attempt may have been saved. | | cancelled | The call was cancelled after an attempt was sent and before a definite answer. |

    The item may already exist. Call list_evidence (for evidence) or get_evidence_discussion (for a comment) and look for it before calling the tool again: a new call is a new request and is never deduplicated against an earlier one, so it can create a duplicate.

  • Cancellation. When the MCP client cancels the call, or its own request timeout fires (many clients use 60 s, shorter than the 2-minute write limit), the attempt in flight is aborted, a pending backoff ends, and no further attempt is sent. If nothing had been sent yet, the result is write_cancelled: nothing was written. Otherwise the outcome is unknown (write_outcome_unknown with reason cancelled): aborting cannot undo a write the server already received. A cancelled call's result often never reaches the client, so after any cancellation or timeout of a write, reconcile with list_evidence / get_evidence_discussion before calling again. Cancelling a read aborts its request; reads are never retried.

Matters (required since 0.4.0)

Every transcription belongs to a matter, so submit_media_transcription requires a matter_id. Call list_matters first (one page per call; pass next_cursor back as cursor for the next page). An unknown matter_id returns the stable error matter_not_found and nothing is uploaded.

Resume. If the matter disappears while an upload finishes, submit_media_transcription returns matter_unavailable with { session_id, sha256 }. The file is already uploaded and the session is kept:

  1. list_matters and choose another matter.
  2. Call submit_media_transcription again with resume_session_id (the session_id), resume_sha256 (the sha256) and the new matter_id. file_path is not needed; a resume never reads the source file.

resume_session_id and resume_sha256 must be given together (resume_invalid otherwise).

Required token abilities:

| Capability | Ability | |------------|---------| | Upload (submit_media_transcription) | transcribe:submit | | List matters (list_matters) | transcribe:submit or transcribe:submit_url | | List transcriptions (list_transcriptions) | transcript:read | | Status / fetch (get_transcription_status, fetch_transcript) | transcript:read | | Search (search_transcript) | transcript:search | | Search a matter (search_matter) | transcript:search and transcript:read | | Read evidence (list_evidence, get_evidence_discussion) | transcript:read | | Write evidence (create_evidence, add_evidence_comment) | evidence:write | | Play an evidence clip (play_evidence) | audio:clip (opt-in; a viewer's token may hold it) |

The minimum end-to-end flow therefore needs transcribe:submit + transcript:read on the token; add transcript:search to enable search_transcript. A transcribe:submit-only token can start a transcription it cannot then observe. Abilities are chosen when you mint the token in LeiUtils account settings. evidence:write is needed on its own (transcript:read does not imply it), and a viewer cannot hold it.

IDs are case-sensitive (#1012)

Every id a tool returns (job_id, matter_id, evidence, comment and upload session ids) is a case-sensitive string, returned in either upper or lower case. Pass it back exactly as returned; a case-changed id is answered like an unknown one. Each tool that takes an id says so in its description.

Skill: evidence review (#924)

The package ships a Claude Code skill, skills/leiutils-evidence-review/SKILL.md, for the search → confirm → create loop:

  • it searches a matter (search_matter) or a transcript, and offers only reviewed passages;
  • it checks the transcript's existing evidence first (all pages), shows an identical item instead of duplicating it, and notes overlaps;
  • it proposes one candidate at a time and calls create_evidence only after an explicit yes to that candidate; "maybe", "all of them" or a new title are not consent. Comments are confirmed the same way, with their exact text;
  • after an unknown outcome it checks before saying anything was saved, and asks again before any new write;
  • it reports each item's marker_url (on an older server without it, it points you to the Evidence board).

Getting it (0.10.0 and later): you no longer copy anything by hand. The connector serves the skill over the connection your client already has, two ways:

  • As an MCP resource, skill://leiutils-evidence-review/SKILL.md, and the connector's startup instructions ask a client that can write files to install it at ~/.claude/skills/leiutils-evidence-review/SKILL.md and to keep it current. In Claude Code that means you should see a single file-write prompt the first time and nothing on later sessions. The write is the client's to make, under its own permission prompt — the connector never writes to your disk.
  • As an MCP prompt of the same name, for a client that cannot write files (Claude Desktop). Selecting it puts the same text into the conversation and leaves nothing on disk.

Either way you can also just ask for it: "use the LeiUtils evidence review skill".

Updating it. The skill carries its own version in frontmatter and the connector reports the version it ships. A client is asked to replace an older copy, and to copy the existing file to SKILL.md.bak-YYYY-MM-DD first — so a local edit of yours is never lost silently, though it is replaced. A copy at the shipped version, or newer, is left alone.

Needs: this connector 0.8.0 or later (0.9.0 to listen to clips with play_evidence, with a token that has audio:clip); a LeiUtils server with matter search (#885) and evidence links (#924 phase 2); the workspace's Evidence feature (Firm plan); a token with transcript:read, transcript:search and evidence:write. Without them the skill stops proposing evidence and says what's missing. list_matters is optional: it needs an upload ability (transcribe:submit or transcribe:submit_url), and without one the skill finds the matter through list_transcriptions (each row names its matter). If an older server returns no marker_url, the skill points you to the Evidence board instead of promising a link.

What it guarantees, and what it doesn't: the skill is an instruction to the agent. It makes asking first the default behaviour of a session that has it installed; it is not a server control. The server's own rules still apply (reviewed transcripts only, idempotent writes, audit, delete on the Evidence board). The quotes it shows stay in your Claude conversation history on your computer: the connector doesn't log transcript text, but it doesn't control the client's history.

Skill: transcribe a link (#1140)

The package also ships skills/leiutils-link-ingest/SKILL.md, for turning a public link (YouTube and other video/audio sites) into a transcription inside a matter:

  • it downloads the media on your computer with yt-dlp, converts it to 64 kbps mono audio with ffmpeg, and uploads that through submit_media_transcription. The link itself is never sent to LeiUtils or to anyone else;
  • it checks the link before putting it into any command. Only an ASCII allowlist of URL characters is accepted, and an accepted link is passed single-quoted; anything else is refused with the offending character named. & and $ are refused too, which for a YouTube link usually means sending it again without everything from the & onwards;
  • it works inside a temporary folder it creates and removes only that folder. Nothing is written to or deleted from your Downloads folder, in either shell;
  • it asks which matter, and whether you have the right to transcribe, in one message — before anything is downloaded.

Getting it: exactly as for the evidence-review skill above — the MCP resource skill://leiutils-link-ingest/SKILL.md, a prompt of the same name, and startup instructions asking a file-writing client to install it and keep it current. Nothing is copied by hand.

Needs: this connector 0.11.0 or later; yt-dlp and ffmpeg on your own machine (the skill says how to install them, and stops if they are missing); a token with an upload ability (transcribe:submit) and LEIUTILS_RIGHTS_ATTESTED=true. Only public links work: a private Drive / SharePoint / Teams / Zoom link needs your sign-in, and the skill says so rather than trying.

What it guarantees, and what it doesn't: the download runs on your machine under your own permissions, as yt-dlp always would — the connector gains no capability and runs nothing itself. The character allowlist is what protects you from a crafted link, and an allowlist is only as good as its character set; that residual risk is recorded on #1140.

Install / configure

Token issuance is gated server-side to enrolled orgs.

Add the server to your MCP client config. Example (.mcp.json / Claude Code MCP settings):

{
  "mcpServers": {
    "leiutils-transcribe": {
      "command": "npx",
      "args": ["-y", "@leiutils/mcp-transcribe"],
      "env": {
        "LEIUTILS_INTEGRATION_TOKEN": "${LEIUTILS_INTEGRATION_TOKEN}",
        "LEIUTILS_API_BASE_URL": "https://YOUR-LEIUTILS-INSTANCE",
        "LEIUTILS_RIGHTS_ATTESTED": "true"
      }
    }
  }
}

Environment

| Variable | Required | Default | Meaning | |----------|----------|---------|---------| | LEIUTILS_INTEGRATION_TOKEN | yes | — | Your org-pinned Claude Code Integration Token, minted in LeiUtils account settings. Revoke it there to instantly cut access. | | LEIUTILS_RIGHTS_ATTESTED | for submit_media_transcription | — | Must be exactly true. Setting it affirms you will only submit media you are authorized to transcribe. If unset, submit_media_transcription fails closed and no upload occurs. | | LEIUTILS_API_BASE_URL | yes | — | Base URL of the LeiUtils instance to connect to. There is no default: the server refuses to start without it. |

Handling the token securely:

  • Provide the token via the environment, never as a command-line argument.
  • Prefer your user-level MCP config. If you put the server in a project .mcp.json, keep that file untracked (git-ignored) or use an environment reference ("${LEIUTILS_INTEGRATION_TOKEN}") — never commit a literal token.

Security posture

  • The token and your local file path are never logged, echoed, returned in tool output, or included in error metadata. The token is sent only in the Authorization header. The server performs no request/response logging at all (enforced by a source-grep test).
  • Transcript content is returned only in successful fetch_transcript / search_transcript results. Evidence titles, quotes and comment bodies are returned only in successful list_evidence / get_evidence_discussion results; write acknowledgments and every error carry no title, body, Idempotency-Key or raw server text.
  • The client is never handed a storage location to write to: file bytes go to the LeiUtils API itself, so there is no signed upload URL to leak.

Uploads

  • The file is validated locally first (it must exist, be a regular file, be readable and be non-empty), then sent in sequential chunks through the LeiUtils API. The chunk size is set by the server when the upload session is opened. Each chunk carries its own SHA-256, and the final complete call carries the SHA-256 of the whole file.
  • Memory stays bounded whatever the file size: one chunk is read at a time, and the whole-file digest is computed by streaming the file from disk.
  • A chunk refused as chunk_in_flight (an earlier attempt at that index is still being settled) is retried up to five more times with a growing delay. Any other chunk failure ends the upload: the client asks the server to abort the session (best effort) and returns the stable error code. Submitting again starts a new upload from the first chunk.
  • The one resumable case is matter_unavailable (see Resume above): the bytes are already stored, so the retry only completes the session against another matter.
  • Uploads and reads set no timeout of their own and rely on the runtime's. The evidence writes are bounded (see Writes, retries and unknown outcomes).

Changes

0.11.0

  • New skill leiutils-link-ingest (#1140), served the same way as evidence-review: transcribe media from a public link into a matter, downloaded on your own computer. See Skill: transcribe a link.
  • Both shipped skills now appear in the startup instructions, each in its own section with its own version, and each as a resource and a prompt.
  • The instructions now cover an installed copy with no version key: treat it as older, back it up to SKILL.md.bak-YYYY-MM-DD, then replace it. They previously gave only "equal or higher" and "lower", so a copy installed before 0.10.0 — which is every copy that predates versioning — matched no case at all and the client was left to decide.
  • No change to any tool, its inputs, its abilities or its answers.
  • Published to latest. 0.10.0 went to next only and was superseded a day later without ever being promoted, so no release between 0.9.0 and this one is one you can have installed.

Upgrading from 0.9.0 or 0.10.0

  1. Get 0.11.0 (see Upgrading from 0.8.0, step 1, for the npx caching caveat).
  2. Restart your MCP client.
  3. Nothing else. Your token, environment and tool usage are unchanged. A copy of either skill in ~/.claude/skills/ that is older, or that carries no version, is copied to SKILL.md.bak-<today> and then replaced.

0.10.0

  • The leiutils-evidence-review skill is served over MCP instead of being copied by hand (#1138): as the resource skill://leiutils-evidence-review/SKILL.md and as a prompt of the same name. The connector's startup instructions ask a file-writing client to install it at ~/.claude/skills/leiutils-evidence-review/SKILL.md and to keep it current.
  • The skill carries its own version, so a stale copy is now detectable. An older copy is replaced after the existing file is copied to SKILL.md.bak-YYYY-MM-DD; a copy at the shipped version or newer is left alone.
  • The connector still writes nothing to your disk. Every write is the client's, under the client's own permission prompt.
  • No change to any tool, its inputs, its abilities or its answers.

Upgrading from 0.9.0

  1. Get 0.10.0 (see Upgrading from 0.8.0, step 1, for the npx caching caveat).
  2. Restart your MCP client.
  3. Nothing else. Your token, environment and tool usage are unchanged. If you already have the skill in ~/.claude/skills/, it carries no version, so it reads as older and will be replaced after a dated backup.

0.9.0

  • New tool play_evidence (#924 phase 3): plays a saved evidence item's audio on your computer. Needs a token with the new audio:clip ability and a server with the evidence clip endpoint. See Playing evidence.
  • New tool stop_playback (#1064): stops the clip that is playing. Say "stop"; some clients (Claude Desktop) don't cancel a running tool when you press stop.
  • The leiutils-evidence-review skill offers "Play it to check the speaker?" after each saved item, shows the saved quote (with a note if the transcript has changed since), and, only on your explicit yes, records the check as an evidence comment. Consent to play is never consent to post.

Upgrading from 0.8.0

  1. Get 0.9.0. With npx and no version, a new session normally runs the latest version, but npx can keep running a copy it cached earlier. If play_evidence is not in the tool list, pin the version in args: "@leiutils/[email protected]". With a global install, run npm i -g @leiutils/mcp-transcribe again; with a pinned version, change the pin.
  2. Restart your MCP client: clients load the tool list when they start.
  3. Token: play_evidence needs audio:clip, which tokens made before this release do not have and cannot gain. Create a new token in LeiUtils (Account → Integration Tokens) with audio:clip (plus the abilities you already use), put it in your client's environment, and revoke the old one. Every other tool keeps working with the token you have.
  4. Skill: copy leiutils-evidence-review into your skills folder again (see Skill: evidence review); 0.9.0 does not install it. 0.10.0 does.

Nothing else changed: the environment variables, the other tools' inputs, their abilities and their answers are as in 0.8.0.

0.8.0

  • New tool search_matter: search a matter's reviewed transcripts at once (#885), the newest 200 at most (searched_truncated says when more exist). Needs a token with transcript:search and transcript:read.
  • create_evidence and list_evidence return marker_url, a link that opens the evidence in LeiUtils (#924); list_transcriptions items carry evidence_count (#1013).
  • fetch_transcript and search_transcript end with job_failed / job_rejected on a dead job instead of not_ready (#1011).
  • add_evidence_comment refusals are distinct: linked_evidence_not_found and reply_depth_exceeded (#1014).
  • Every id-taking tool says ids are case-sensitive (#1012).
  • The package ships the leiutils-evidence-review skill (#924; see Skill: evidence review).
  • The new fields and codes need a server with milestone #3 (staging from 2026-10-29); against an older server search_matter fails (the endpoint does not exist) and the new fields are absent.

Upgrading from 0.7.0

  1. Get 0.8.0. With npx and no version (the configuration in Install / configure), a new session normally runs the latest version, but npx can keep running a copy it cached earlier. If search_matter is not in the tool list, pin the version in args: "@leiutils/[email protected]". With a global install, run npm i -g @leiutils/mcp-transcribe again; with a pinned version, change the pin.
  2. Restart your MCP client (Claude Code, Claude Desktop, …): clients load the tool list when they start.
  3. Token: search_matter needs transcript:search and transcript:read. A token's abilities are fixed when it is created, so if yours lacks one, create a new token in LeiUtils (Account → Integration Tokens), put it in your client's environment, and revoke the old one. The other tools keep working with the token you have.
  4. Skill (optional): copy leiutils-evidence-review into your skills folder (see Skill: evidence review). The package does not install it.

The environment variables, the existing tools' inputs and their required abilities are unchanged. Some error answers did change (see the list above): job_failed / job_rejected instead of not_ready for a dead job, and linked_evidence_not_found / reply_depth_exceeded for comment refusals. Code that matches on the old codes needs updating.

0.7.0

  • Four evidence tools: list_evidence, create_evidence, get_evidence_discussion and add_evidence_comment. They need a server with the integration evidence API and a workspace with the Evidence feature (Firm). Writes need the evidence:write ability.
  • Writes carry one Idempotency-Key per call and retry lost answers within a 2-minute limit; an unknown outcome is reported as write_outcome_unknown.
  • Cancelling a call (or a client request timeout) stops its writes: no attempt starts after it (write_cancelled before any send, else write_outcome_unknown / cancelled).
  • search_transcript's description says reviewed matches feed create_evidence.

0.6.0

  • New tool list_transcriptions: list the workspace's transcriptions, with optional matter_id and status filters.
  • list_matters gains an optional q (name search, at most 100 characters).

0.5.0

  • Breaking: the tools that transcribed media from a web address are removed. The server exposes five tools; local files are the only supported input.

0.4.0

  • Breaking: the submit tools require matter_id (the server now rejects uploads without one). 0.3.0 clients get 422 against current servers.
  • New tool list_matters.
  • resume_sha256 is normalised to lowercase (the server requires it).
  • Resume after matter_unavailable via resume_session_id + resume_sha256
    • a new matter_id, without re-uploading or re-reading the file.

Development

npm install
npm run build     # tsc -> dist/
npm test          # vitest
npm run dev       # tsx watch (stdio)

Verification — MCP inspector

LEIUTILS_INTEGRATION_TOKEN=... LEIUTILS_API_BASE_URL=https://YOUR-LEIUTILS-INSTANCE \
  LEIUTILS_RIGHTS_ATTESTED=true \
  npx @modelcontextprotocol/inspector node dist/server.js

Confirm 13 tools are listed (submit_media_transcription, get_transcription_status, fetch_transcript, search_transcript, list_matters, list_transcriptions, search_matter, list_evidence, create_evidence, get_evidence_discussion, add_evidence_comment, play_evidence, stop_playback) and that a submit → poll status → fetch_transcript loop works against a demo job.