Evidence-driven self-evolution control plane for AI-agent products, with Loop Engineering, human approval, SaaS multi-tenancy, CI/CD delivery, and auditable release decisions.
EvoPilot helps teams operate AI-agent products as releasable software. It collects real evidence from runtime events, traces, evaluations, CI/CD, source changes, LLM calls, and user feedback; turns that evidence into reviewable evolution opportunities; waits for human approval; then drives code upgrades, delivery, and product-native GO / NO-GO release decisions.
It is not an agent runtime, a prompt playground, or a generic code generator. Agent runtimes do the work; EvoPilot governs whether the product is ready to evolve and release.
EvoPilot is marked GA Release V1.0 for its product control plane and Loop Engineering runtime.
As of the 2026-07-07 production validation, the EvoPilot SaaS multi-tenant deployment has reached GA stable Release status:
| Signal | Result |
|---|---|
| Production user E2E checks | 92 |
| PASS | 88 |
| FAIL | 0 |
| Low-severity WARN | 4 |
| Production health | UP |
| Readiness | READY |
| Postgres business store | READY |
| Real LLM path | GLM-5.1 credits and tokens recorded |
The authoritative release verdict is exposed by:
GET /api/v1/release/decisionsDetailed release evidence and deployment checklists live in docs/reference/release-package.md and docs/reference/production-user-e2e.md.
| Capability | What EvoPilot provides |
|---|---|
| Loop Engineering | Durable loop state, executor graphs, checkpoints, replay, watchdog recovery, worker leases, sandbox proof, and timeline audit. |
| GlobalGoal planning | A white-box goal layer above LoopRun that decomposes RC/GA objectives into ordered GoalTargets, binds each target to governed loops, and exposes progress, blockers, timeline, graph, evidence matrix, and final report. |
| Evidence ingestion | Runtime events, traces, logs, evaluations, release signals, APM-derived data, and user feedback. |
| Human approval | Reviewable proposals before high-risk evolution, source writeback, merge, or release actions. |
| Code upgrades | Bounded code-upgrader execution with allowed paths, validation commands, branch/commit evidence, and source closure. |
| CI/CD delivery | GitHub Actions and GitLab CI native project DevOps, deploy connectors, health gates, and pipeline evidence. |
| Release governance | Per-project release targets, evidence bundles, scenario matrices, risk registers, and GO / CONDITIONAL-GO / NO-GO decisions. |
| SaaS multi-tenancy | Platform admin, tenant admin, tenant user flows, workspace RBAC, tenant-aware evidence, quota foundations, and scoped secrets. |
| Dashboard | Chinese SaaS console for onboarding, projects, loops, approvals, release decisions, observability, tenants, users, and help manual workflows. |
Prerequisites:
- Node.js 22+
- npm
Run locally:
npm install
npm run build
npm run server:debugVerify the API:
http://127.0.0.1:19876/health
http://127.0.0.1:19876/ready
Run the standalone Dashboard from the sibling repository:
cd ../evopilot-dashboard
EVOPILOT_API_BASE_URL=http://127.0.0.1:19876 npm run devDebug mode is for local development and UI validation. Production mode is the default for real operation and requires authentication, real runtime boundaries, and configured LLM/source/CI/CD credentials.
The dashboard is the primary product surface for operators and tenant users, but it now lives in the standalone yeliang-wang/evopilot-dashboard repository. EvoPilot itself focuses on the API server, CLI, domain state, release governance, and execution control plane. The standalone Dashboard consumes EvoPilot APIs and includes the GlobalGoal Cockpit for RC/GA workflow visibility.
| Role | Main workflows |
|---|---|
| Platform admin | Create tenants, manage tenant admins, inspect SaaS readiness, observe cross-tenant release health. |
| Tenant admin | Manage workspace users, bind projects, configure credentials, run target loops, approve source/release actions. |
| Tenant user | Inspect projects, evidence, loop progress, approvals assigned to them, release decisions, and help manual steps. |
The full browser operation guide lives in the standalone Dashboard repository under docs/. EvoPilot keeps the API/CLI/control-plane guide in docs/guides/user-guide.md.
EvoPilot applies Loop Engineering to product evolution. For larger RC/GA objectives, the GlobalGoal layer sits above LoopRun and turns one global objective into multiple GoalTargets before each target is executed through the governed loop runtime:
GlobalGoal -> GoalTarget -> LoopRun -> Release Decision
|
v
Sandbox -> Context -> Harness -> Loop
| Layer | EvoPilot responsibility |
|---|---|
| Sandbox | Isolated execution boundaries, credential scope, protected paths, and deploy constraints. |
| Context | Durable loop state, evidence, artifacts, checkpoints, and replay context. |
| Harness | API control plane, RBAC, audit, worker leases, approvals, retries, and watchdog recovery. |
| Loop | Continue, stop, retry, approve, source-close, release, or route to humans. |
Architecture details:
- Continuous Evolution Control Plane
- Loop Runtime Architecture
- ProofOps Target Loop Mode
- Mainstream Loop Harness Alignment
Common commands:
npm run build
npm run check
npm run test:e2e:production
npm run release:soak:ga:activeProduction package commands:
npm run store:postgres:migrate
npm run store:postgres:backup
npm run store:postgres:restoreDocker:
docker build -t evopilot:1.0.0 .
docker compose up --buildDeployment details are in docs/operations/deployment.md.
Primary API surfaces include:
| Area | Examples |
|---|---|
| Health and readiness | GET /health, GET /ready |
| Auth and users | POST /api/v1/auth/login, GET /api/v1/users, POST /api/v1/users |
| Projects and evidence | GET /api/v1/projects, POST /api/v1/evidence/events |
| Project DevOps | POST /api/v1/projects/{projectId}/devops, POST /api/v1/projects/{projectId}/devops/preflight |
| Global goals | GET /api/v1/goals, POST /api/v1/goals, POST /api/v1/goals/{goalId}/plan, POST /api/v1/goals/{goalId}/advance, GET /api/v1/goals/{goalId}/snapshot |
| Loops | POST /api/v1/loops, POST /api/v1/loops/{loopId}/start, GET /api/v1/loops/{loopId}/timeline |
| Source closure | POST /api/v1/loops/{loopId}/source-closure/execute, POST /api/v1/loops/{loopId}/source-closure/review-decision |
| Release | POST /api/v1/release/evidence, GET /api/v1/release/decisions |
| SaaS | GET /api/v1/tenants, GET /api/v1/workspaces, GET /api/v1/saas/observability |
See docs/api/README.md and docs/api/openapi.json for the full API reference.
Start with the documentation index. The main entry points are:
| Reader | Start here |
|---|---|
| New user | Quick Start |
| AI agent or CLI automation | CLI and AI Agent Runbook |
| Dashboard integrator | Dashboard Integration |
| API integrator | API Reference and OpenAPI |
| Production operator | Operations |
| Architect | Architecture |
packages/core/ lifecycle, evidence, planning, review, delivery, release models
packages/server/ control-plane API and optional compatibility static host
packages/llm/ LLM gateway, routing, compression, metrics
packages/adapter-github/ GitHub adapter
packages/adapter-gitlab/ GitLab adapter
packages/adapter-local-git/ local Git adapter
docs/ user, API, deployment, architecture, testing, release docs
examples/ onboarding and integration examples
scripts/ E2E, LLM, Postgres store, release, and verification scripts
runtimes/ managed runtime images, locks, and supply-chain material
tests/ unit, smoke, functional, and E2E tests
The product Dashboard is maintained separately in yeliang-wang/evopilot-dashboard.
EvoPilot uses Apache License 2.0 and includes the standard public-repository governance entry points:
Governance verification:
npm run verify:oss-governanceEvoPilot is licensed under the Apache License, Version 2.0. See LICENSE and NOTICE.