Engineering status of running repid-engine so prompt/response text stays on your box. This is a skeleton: it states exactly what boots today and what is still hosted-only. No marketing claims.
- Boot coupling.
src/config.tsthrows at import whenSUPABASE_URL/ a service key is absent, so a clone with no hosted Supabase cannot even start. - Prompt egress. The HAL quorum POSTs prompt text to cloud LLM APIs
(
api.groq.com,api.deepseek.com,api.anthropic.com, …).
This change adds three env-guarded seams that address both, all default-OFF (hosted behavior is byte-for-byte unchanged when the flags are unset).
| Env | Default | Effect |
|---|---|---|
LOCAL_MODE |
off | Boot without a hosted Supabase. A missing SUPABASE_URL/key is backfilled with the loopback default http://127.0.0.1:54321 instead of throwing. Announced loudly at boot. |
ONLY_ATTESTATIONS_LEAVE |
off | Data-locality boundary. Prompt/response text may not egress to a non-local host; proofs / EAS anchors / chain RPC still may. Enforced at the HAL provider call choke point (src/hal/cross-llm-client.ts → assertPromptEgressAllowed). |
ZKP_SERVICE_URL / PLONKY3_PROVER_URL under LOCAL_MODE |
unset = no prover | A proof request carries the raw agent id and score, so in LOCAL_MODE the prover is called only at a loopback URL (localhost, 127.0.0.1, [::1]). The hosted pin is never a default; any other host is refused before a byte is sent. With no loopback prover, score events record no proof, and /prove-repid and /batch/prove answer 503 with status PROVER_UNAVAILABLE (no HMAC stand-in). src/config/prover.ts → localProverGate. |
LOCAL_LLM_BASE_URL (or OPENAI_BASE_URL) |
unset | OpenAI-compatible base URL for the quorum. Redirects every openai-compat provider to <base>/chat/completions — e.g. an Ollama (http://localhost:11434/v1) or vLLM box. anthropic-native is never redirected (wrong wire shape). |
"Local" for the boundary = loopback, RFC1918/private LAN, link-local, and the
host.docker.internal / ollama / vllm service conventions
(src/selfhost/egress-guard.ts).
- The data-local core process —
src/selfhost.ts, entrydist/selfhost.js. Boots with no hosted Supabase, binds0.0.0.0:$PORT, serves:GET /health— liveness + whether a hosted DB is configured.GET /selfhost/status— the boundary it is running under, plus an explicitboots_todayvsstub_for_selfhostsplit.
- Config loads without hosted Supabase under
LOCAL_MODE(was a hard throw). - The egress boundary — with
ONLY_ATTESTATIONS_LEAVE=true, a prompt bound for a cloud LLM is refused before it is sent; a local target and all commitments (attestation/chain) pass. - The LLM base-URL override — the quorum's openai-compat providers retarget to a local model with one env var.
- The RepID score path runs against a LOCAL persistent store. In
LOCAL_MODE,src/db.tsresolvesdbto a supabase-js-shaped adapter over a local store (src/selfhost/local-store.ts: node:sqlite file, or a JSON file if no driver loads) instead of the hosted supabase-js client. The realupdateRepId()— fetch agent, decay, ecosystem-need weight, redemption, delta, audit-row insert, score write-back, supply-rate upsert, badge check — completes end-to-end on the local box. Proven intests/selfhost-local-store.test.ts: a synthetic agent'sCODE_CONTRIBUTIONevent writes arepid_score_eventsrow locally, movescurrent_repid, persists durably (a fresh adapter on the same file re-reads it), and aglobal.fetchtripwire (wired through the boundary'sisLocalHost()) is never called — zero data egress.
Verified boot (LOCAL_MODE, no SUPABASE_URL, boundary on, local model base):
[LOCAL_MODE] enabled — data-local boot. supabase=LOOPBACK http://127.0.0.1:54321 (no hosted DB) · ONLY_ATTESTATIONS_LEAVE=true · LOCAL_LLM_BASE_URL=http://localhost:11434/v1
[selfhost] repid-engine v1.0.0 data-local boot on :39217
local_mode=true hosted_db=false
only_attestations_leave=true prompt_egress_to_cloud=blocked
llm_quorum_target=local local_base_url=http://localhost:11434/v1
Covered by tests under tests/selfhost-*.test.ts (offline), including
tests/selfhost-local-store.test.ts which drives the real score path locally.
- The full
src/index.tsentrypoint and its ~20 background workers/crons (proof-drain, receipt-indexer, eas-anchor, score-monitor, HAL production loggers, …). These read/write hosted-Supabase tables beyond the four the score path uses, and several also do on-chain / provider I/O. The local store adapter coversrepid_agents,repid_score_events,repid_ecosystem_supplyandrepid_badges— enough forupdateRepId(), not enough for the workers. Runningdist/index.jsin LOCAL_MODE would still hit unbacked tables, so the self-host entry (src/selfhost.ts) deliberately does not start them. Making the workers local is the next slice (more tables + the worker loops themselves). - The local store is JS-side filtered, not a query planner: it loads a table's rows and filters in memory. Correct and durable at single-operator scale; it is not built for large-table analytical queries.
- The embedding call in the HAL agreement step also egresses text
(
api.openai.com/v1/embeddings). The guard classifiesembeddingas content-bearing, but the embedding client is not yet wired throughassertPromptEgressAllowed; under the boundary the quorum falls back to local Jaccard similarity, so it degrades rather than leaks — but the explicit guard at that call site is still TODO.
Local process:
npm run build
LOCAL_MODE=true ONLY_ATTESTATIONS_LEAVE=true \
LOCAL_LLM_BASE_URL=http://localhost:11434/v1 \
node dist/selfhost.js
curl localhost:3000/selfhost/statusContainer (Dockerfile.selfhost bakes LOCAL_MODE/ONLY_ATTESTATIONS_LEAVE on):
DOCKER_BUILDKIT=1 docker build -f Dockerfile.selfhost \
--secret id=gh_token,env=GH_TOKEN -t repid-engine:selfhost .
docker run --rm -p 3000:3000 \
-e LOCAL_LLM_BASE_URL=http://host.docker.internal:11434/v1 \
repid-engine:selfhost(The gh_token build secret is only needed while @hyperdag/proof-verifier
resolves from a private GitHub repo.)