ScopeGate Documentation
ScopeGate is an ephemeral credentials gateway for coding agents (Claude Code, Kimi Code, Cursor, OpenCode). The agent never holds secrets — it holds short-lived, minimum-scope capabilities. Secrets live in an encrypted local vault and are injected only at the outbound hop, just before the request leaves your machine.
Agent/CLI ──(MCP, zero secrets)──► ScopeGate ──(creds injected)──► GitHub / AWS / your MCPs
│
vault + policy engine + token minter + signed audit
Pages
- Quickstart — zero to first proxied tool call, agent-executable
- Security Model — the enforcement rules and why they hold
- Agent Protocol — the SKILL protocol agents follow
- CLI Reference — every command and flag
- Configuration —
scopegate.yaml,policies.yaml, env vars
The contract in one minute
- Capability ≠ credential. Agents request
github:call:create_pull_request
for 5 minutes; they never see the token that ultimately authorizes the call.
- Secrets enter out-of-band.
scopegate secret add <ref>from a human
terminal — hidden prompt or piped stdin, never chat, never argv.
- Write asymmetry. Agents propose policy changes
(policies.pending.yaml); only humans edit policies.yaml.
- Everything is audited. Append-only, hash-chained, Ed25519-signed JSONL;
inputs are hashed, never stored. scopegate audit verify detects tampering.
Status & roadmap
Implemented today: proxy + injection (bearer, env), OAuth2 refresh daemon with device-code re-auth, token minter (JWT / GitHub App / AWS STS / Google SA / Huly), policy engine with TTL grants, hard limits, rate limiting, PII redaction and human approvals, vault with AES-256-GCM (+ optional vaultd process isolation and OS keychain master keys), multi-harness init (Claude Code, Kimi Code, Cursor, OpenCode, .mcp.json), signed audit trail with verify/query/reindex, upstream registry with signed manifests, EdDSA attestation, honeytokens, and native bridges (Huly, Railway, Cloudflare, Google).
Also in this repo: ScopeGate Cloud — the optional multi-tenant management plane (scopegate cloud serve): landing page at /, product panel at /panel (fleet, approvals, capabilities, audit, versioned signed policies, billing), and the /v1 API with enroll, signed policy distribution, verified audit ingest, fleet revocation and the panel-driven approval loop. Local-first holds: gateways keep working with their local policy when the cloud is down.
Still future (see the roadmap EPICs under ../../docs/Implementacion/):
- Cloud hardening — Postgres store, SSO/SAML with per-team roles, Stripe
metering, on-prem packaging, SIEM export (EPIC-10 enterprise stories).
- Ecosystem growth — more registry upstreams and warm-pool fleets
(EPIC-12 follow-ups).