sphior-ai-governance-proxy

Zero-code AI governance recording. A self-hosted proxy that sits in front of your LLM provider (OpenAI / Anthropic-compatible). Point your app's base URL at it — it forwards every request unchanged and records the execution as tamper-evident, hash-only evidence to SPHIOR for ISO 42001 / EU AI Act / SOC 2.

• No code changes — just change your LLM base URL.
• Your provider keys never leave your infrastructure — they are forwarded upstream, never sent to SPHIOR.
• hash_only — only hashes and non-sensitive metadata (model, tokens, latency, trace id) are recorded.
• fail-safe — a recording failure never breaks your LLM call.

Records supporting evidence. Does not assert compliance, pass/fail, or an audit opinion.

Run

Get an ingestion token from your SPHIOR dashboard (Account → AI Governance → Ingestion tokens).

SPHIOR_AIGOV_TOKEN=<your-token> \
UPSTREAM_BASE_URL=https://api.openai.com \
npx -y sphior-ai-governance-proxy
# → listening on :8788 → https://api.openai.com

Then point your SDK/app at the proxy:

// OpenAI SDK
new OpenAI({ baseURL: "http://localhost:8788/v1", apiKey: process.env.OPENAI_API_KEY });

# or raw
curl http://localhost:8788/v1/chat/completions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -H "x-sphior-actor: support-agent" \
  -d '{"model":"gpt-4o-mini","messages":[{"role":"user","content":"hi"}]}'

Anthropic: set UPSTREAM_BASE_URL=https://api.anthropic.com and point at http://localhost:8788.

Coding agents

The agents your engineers run all day are LLM traffic too. Each one takes a base URL, so recording them needs no code and no plugin — point the base URL at the proxy and every call is recorded, whether or not the agent cooperates.

Codex CLI

# ~/.codex/config.toml
openai_base_url = "http://localhost:8788/v1"

Or as an explicit provider, if you already use one:

[model_providers.sphior]
base_url = "http://localhost:8788/v1"
wire_api = "responses"      # use "chat" for OpenAI-compatible gateways
env_key  = "OPENAI_API_KEY"

model_provider = "sphior"

Claude Code

SPHIOR_AIGOV_TOKEN=<your-token> UPSTREAM_BASE_URL=https://api.anthropic.com \
  npx -y sphior-ai-governance-proxy

ANTHROPIC_BASE_URL=http://localhost:8788 claude

Cursor / Windsurf / Cline

Set the custom OpenAI base URL in settings to http://localhost:8788/v1.

The proxy runs on your own machine or your own network. Provider API keys are forwarded upstream and never sent to SPHIOR, and prompts and outputs are never recorded — only hashes and metadata. Attribute per agent with SPHIOR_ACTOR=codex (or the x-sphior-actor header).

Attribution

Set the acting agent per request with the x-sphior-actor header (stripped before forwarding upstream), or a default with the SPHIOR_ACTOR env var.

Environment variables

| Var | Required | Default | |---|---|---| | SPHIOR_AIGOV_TOKEN | yes | — | | UPSTREAM_BASE_URL | no | https://api.openai.com | | PORT | no | 8788 | | SPHIOR_ACTOR | no | — (or x-sphior-actor header) | | SPHIOR_AIGOV_ENDPOINT | no | https://sphior.com/api/ai-governance/ingest | | SPHIOR_DEBUG | no | false |

What is recorded

Non-sensitive metadata only: model, tokens (from the response usage), latency, trace_id (the response id), actor, and a policy_decision of allow / error. Prompts and outputs are not sent to SPHIOR. Each event is canonicalized (RFC 8785 JCS), hashed (SHA-256), chained, and periodically anchored to an append-only, independently verifiable record you can check with the open sphior-verify tool — without trusting SPHIOR.

Other ways to record

Prefer explicit calls from your own code? Use the @sphior/evidence SDK (Node; Python and Go also available). From an AI agent, so it can record its own decisions — the layer a base-URL proxy cannot see? Use sphior-ai-governance-mcp.

License

Apache-2.0