arcmira, is both the command line and the TypeScript client. The PyPI package arcmira is the Python client. All three surfaces call the same /v1 API, read the key from ARCMIRA_API_KEY, and draw on the same credits.
All three are Apache-2.0.
Command line
arcmira resolve command that finds the id. Dated commands take --after (included) and --before (excluded). --json prints the raw response. Every --help ends with examples, and arcmira api calls any /v1 route with your key.
Exit codes: 1 an API error or a failed Premium transcript (the message names the code, any unlock link, and the price on a refused paid read), 2 a usage error, 4 a Premium transcript still transcribing. On 4, run the same command again later; it is never charged twice. After a failure, --retry starts one new attempt.
Connect your coding agents
npx arcmira setup finds Claude Code, Codex, Cursor, VS Code, Gemini CLI and Claude Desktop, adds the Arcmira MCP server to each, installs the skills, and turns updates on. --dry-run prints what would change. Sign in once per agent afterwards.
TypeScript
for await; each row costs credits, so stop when you have enough. Every method: reference.md.
Deno
Deno imports the same npm package. Save this assearch.ts:
Python
Premium transcripts
transcripts.get({ video_id, quality: "premium" }) in TypeScript and transcripts.get(video_id, quality="premium") in Python make the one Premium read. The result is ready, pending or failed; on pending, wait Retry-After and call it again. To try again after failed, pass retry: true (retry=True in Python) on that one call and drop it while you poll. A refused paid read raises PaymentRequiredError or ForbiddenError with the price in body.error.details.quote. See Transcripts.
Not ready for a paid plan? Tell us why and what you would pay for. Feedback is free to give and helps us build what you want.
Agent frameworks
These recipes give an agent one bounded tool: YouTube transcript search, up to five passages per call. SetARCMIRA_API_KEY and keep it in server-side code; the model supplies only query. Each search uses credits from your plan, then your on-demand budget, and follows your plan’s freshness window; the tool never reads a Premium transcript or changes monitors. Each example makes one request per call, with no automatic retries. When answering, keep partial, access, search_index and note, and cite each passage’s watch_url. An empty result is not evidence that a topic was never discussed. For keyless search of these docs instead, see Documentation MCP.
Haystack
Installhaystack-ai==3.2.0 and httpx==0.28.1. This tool calls the HTTP API.
tools=[tool] to your Haystack agent inside the HTTP context. An HTTP failure raises, and Haystack keeps the original error, with its body and doc_url, as the cause. Request, response and refusal behavior was tested with synthetic fixtures; production research and model inference were not.
LlamaIndex
Installarcmira==0.4.2 and llama-index-core==0.14.25. This tool uses the Python SDK.
tools=[search] to your LlamaIndex agent and keep the HTTP context open until it finishes. API failures raise the SDK’s typed errors with their codes and access details. The tool passed request, response and refusal checks with SDK 0.4.2 and LlamaIndex core 0.14.25; model inference was not tested. See FunctionTool for agent wiring.
Errors and versions
A non-2xx answer raises a typed error named for its status, such asPaymentRequiredError, whose body is the API’s error object. Switch on body.error.code; every code is on Errors.
The packages version separately from the API. A removed or renamed field is a major release. The CLI checks npm at most once a day and prints one line when a newer version is out.