GLC v1 — Gateway for LLMs and Channels

Session 12 update: this is the S11 deliverable. S12's deliverable is theschoolofai/glc_v2 — a containerised, Modal-deployed version of this gateway, designed as the pen-test attack target for the S12 assignment. Read v2's README to see what changed and how to participate in the pen-testing assignment.

GLC v1 is the Session 11 deliverable. It absorbs the V9 LLM gateway (text chat, vision, embeddings, cost ledger, provider routing) and adds the new channel and voice layer for the first time. The S9 agent runtime and the S10 Computer-Use skill point at this gateway unchanged.

Port: 8111 (V9 stays on 8109 for older student work).

Install

cd glc_v1
uv sync
uv run glc serve

The first boot creates ~/.glc/ with the audit log, pairing store, gateway db, channels.yaml, and a per-installation token. The token gates /v1/control/* and the channel WebSocket. Print it with uv run glc token.

Point your existing S9 / S10 client at it

export LLM_GATEWAY_V9_URL=http://localhost:8111

Every V9 route — /v1/chat, /v1/chat/batch, /v1/vision, /v1/embed, /v1/cost/by_agent, /v1/providers, /v1/capabilities, /v1/status, /v1/routers, /v1/calls, /v1/embedders — works identically against this port. The S9 Browser skill and S10 Computer-Use skill require no code changes.

New surfaces

Route	Purpose
POST /v1/transcribe	STT dispatcher → groq_whisper / whisper_cpp / gemini_live
POST /v1/speak	TTS dispatcher → kokoro / elevenlabs / cartesia / gemini_live / system_fallback
WS /v1/channels/{name}	Channel adapter control plane
POST /v1/control/kill	Out-of-band kill switch (loopback only)
POST /v1/control/pair	Issue a rotating six-digit pairing code
POST /v1/control/pair/confirm	Owner confirms a pairing code
GET /v1/control/presence	Channels registered, paired users, uptime

Daemonise

./daemon/install.sh                # macOS launchd, Linux systemd, Windows NSSM
./daemon/install.sh --uninstall    # remove the service
./daemon/install.sh --models       # fetch Kokoro + whisper.cpp base model

Your group's adapter

There are 22 group slots:

• 15 channel adapters under glc/channels/catalogue/<name>/.
• 7 voice providers under glc/voice/stt/providers/<name>/ and glc/voice/tts/providers/<name>/ (3 STT + 4 TTS). system_fallback ships fully implemented and is not a slot.

Group → slot assignments are fixed by the instructors and listed in GROUPS.md. There is no claim PR — the assignment table is the source of truth. If you think your group's row is wrong, raise it in the S11 chat (G<n> sub-channel) and a TA will correct it.

Workflow:

1. Find your group's row in GROUPS.md. Note the slot name and the Owned paths.
2. Read docs/ADAPTER_GUIDE.md.
3. Implement adapter.py (and schemas.py if needed) against the mock-API fake under tests/channels/mocks/ or tests/voice/{stt,tts}/mocks/.
4. Open your implementation PR. Include # Group: <name> and # Slot: <slot> markers in the PR description.
5. The adapter-pr.yml workflow runs three jobs against your PR:
   • boundary — fails if your diff touches files outside your Owned paths in GROUPS.md.
   • test-changed-slot — runs only the matching test_<slot>.py, plus ruff and mypy on your slot's dir.
   • scorecard — auto-comments a per-group rubric scorecard.
6. Branch protection requires a CODEOWNER review. The @theschoolofai team merges on green.

Pass CI

.github/workflows/ci.yml runs on every push and PR:

• ruff check . and ruff format --check .
• mypy glc tests
• pytest tests/ --cov=glc --cov-fail-under=80 -m "not requires_live_api"
• scripts/validate_envelope.py — envelope must not be re-shaped per adapter
• scripts/validate_policy.py — policy.yaml must parse and the five lecture defaults must load
• scripts/validate_claims.py — no duplicate group rows in GROUPS.md

Architecture and rationale

See docs/ARCHITECTURE.md. The six architectural moves from Session 11 §7 are the operational answer to the four documented OpenClaw incidents: ClawJacked, Summer Yue email deletion, Moltbook DB exposure, and the malicious-skills taxonomy.

What this scaffold does NOT do

• It does not ship any working channel adapter. Every adapter.py in glc/channels/catalogue/ raises NotImplementedError. That is the group assignment.
• It does not hold paid-API defaults. Free-tier sign-ups are linked in docs/VOICE_GUIDE.md.
• It does not pull LangChain, CrewAI, AutoGen, or Open Interpreter.

License

MIT — see LICENSE.