slng-ai/sdks

Monorepo for the Slng Unified API CLI and SDKs.

Layout

specs/         # vendored OpenAPI + AsyncAPI specs from gateway-specs
sdks/ts/       # TypeScript SDK (Stainless-generated HTTP + hand-written WS)
sdks/python/   # Python SDK (Stainless-generated HTTP + hand-written WS)
cli/           # Bun CLI (depends on sdks/ts via workspace)
scripts/       # spec sync + WS type generation

Quick start

bun install
bun run regen          # sync specs + regenerate WS message types
cd cli && bun run dev  # run the CLI from source

Stainless (HTTP SDKs)

See STAINLESS_SETUP.md for one-time onboarding.

CLI commands

Full reference in cli/README.md. The command trees:

Tree	Subcommands
tts / stt	synth and transcribe, one-shot or --stream
models / voices / whoami / login	catalogs and auth
agents	list get create update replace delete duplicate, calls {dispatch list get}, web-sessions create, push
tool	list get create update build run publish (create takes a Python file)
mcp	list get tools (read-only; attach servers in the dashboard)
secret	list get create (--secrets-file, --kind, --overwrite; values never readable back)
trunks	list get (--direction; read-only)
config	get set profiles use add remove reset

Every subcommand takes --json and --profile <name>. Names are matched exactly and case-sensitively. The shared-resource trees (tool, mcp, secret, trunks) hit routes that are mounted include_in_schema=False, so they are absent from specs/ and from the generated SDKs — they go through the raw-fetch helper in cli/src/lib/agents.ts.

CLI design

The CLI shells out to ffmpeg / sox / arecord for audio I/O — it never opens an audio device directly. Stream audio via pipes:

voiceai tts "hello" --model slng/deepgram/aura:2-en | ffplay -
arecord -f S16_LE -r 16000 | voiceai stt --stream --model deepgram/nova:3

Releasing the CLI

Push a tag from main. Nothing else is done by hand:

git switch main && git pull
git tag -a cli-v0.1.23 -m "cli 0.1.23"
git push origin cli-v0.1.23

The tag is the version. The workflow writes it into cli/package.json before it builds, so don't bump that file. The committed version stays 0.0.0-dev. If the AsyncAPI specs changed, run bun run regen and merge the diff before tagging.

A failed run can be re-run from the Actions page, or started by hand with the tag as input (workflow_dispatch).

The compile depends on one committed generated file, streaming/ts/messages.ts, refreshed by bun run regen. Models and voices are not bundled: the CLI reads them live from the public catalog (GET /v1/catalog/models). The CLI imports the HTTP SDK as voiceai-sdk from npm (published from this repo via Stainless), so the Stainless working copies in sdks/ aren't needed at release time.

The workflow then:

1. Builds voiceai-{darwin-arm64,darwin-x64,linux-arm64,linux-x64} via bun build --compile.
2. Attaches them to a GitHub Release at the tag.
3. Publishes voiceai-cli to npm with provenance, through npm trusted publishing (no token).
4. Commits Formula/voiceai.rb with the new version and SHAs straight to main of slng-ai/homebrew-tap (uses HOMEBREW_TAP_GITHUB_TOKEN).

Required, one-time, before the first tag:

• Create slng-ai/homebrew-tap on GitHub. Must be a public repo — Homebrew clones the tap over HTTPS without auth on every brew install. The repo can start empty; the workflow's tap step writes Formula/voiceai.rb on first release.
• Reserve voiceai on npmjs.com.

Required secret on slng-ai/sdks (a repo or org secret):

• HOMEBREW_TAP_GITHUB_TOKEN: a fine-grained token, owner slng-ai, only the repository slng-ai/homebrew-tap, with Contents: read and write. Unmute's release writes to the same tap, so one org secret can serve both repos.

The curl installer at scripts/install.sh resolves the latest cli-v* tag on the fly, so users always get the most recent release without us hardcoding versions.