engineering-mcp
Health Uyari
- License — License: NOASSERTION
- Description — Repository has a description
- Active repo — Last push 0 days ago
- Low visibility — Only 8 GitHub stars
Code Gecti
- Code scan — Scanned 12 files during light audit, no dangerous patterns found
Permissions Gecti
- Permissions — No dangerous permissions requested
Bu listing icin henuz AI raporu yok.
Engineering MCP is a repository-aware AI engineering control plane for orchestrating trusted coding workers through MCP, with lifecycle management, execution isolation, and verifiable task handoff.
Engineering MCP
Engineering MCP is an AI engineering control plane built on the Model Context Protocol (MCP).
It coordinates trusted coding workers such as Codex-based workers and Generic Harness integrations while keeping repository ownership, task lifecycle, and execution authority under explicit control.
Unlike a model router, Engineering MCP does not decide which model should solve a problem or expose provider credentials to tasks. Instead, it provides the engineering coordination layer:
- repository-bound task management;
- controlled worker delegation;
- lifecycle tracking and recovery;
- worker result validation;
- execution records;
- trusted worker profile selection;
- compatibility with multiple worker harnesses.
The system is designed around a simple principle:
Workers perform bounded execution. The control plane owns engineering state.
Engineering MCP keeps durable engineering truth in the repository, Git history, tests, and controlled ledger state rather than relying on temporary AI conversation context.
Architecture
MCP Host / OWNER
↓
Engineering MCP
lifecycle · ledger · repo safety
↓
Trusted Worker Runner
↓
Native Harness Adapter
↓
Harness-managed Provider / Model
Harness-Native Execution Principle
Engineering MCP standardizes:
WHAT must be done(task contract)- lifecycle ownership
- repository binding and Git verification
- final Engineering Worker Protocol result contract
Native Harnesses decide:
HOW to reason- tool use
- context management
- model/provider selection
- execution strategy
Engineering MCP is not a model router, universal model API, replacement Harness, or autonomous scheduler.
Quick start
npm install -g engineering-mcp-cli
engineering-mcp --help
engineering-mcp setup
engineering-mcp doctor
engineering-mcp profiles --worker-profiles /absolute/path/to/profiles.yaml
From inside a Git repository, start an OWNER server:
cd /path/to/repo
engineering-mcp --role owner
For local development from this repository:
npm ci
npm run typecheck
npm test
npm run build
node dist/cli.js --help
Verified capability matrix
Hosts
| Host | Support |
|---|---|
| Grok CLI | VERIFIED |
| Codex CLI | VERIFIED FOR MCP STARTUP / AUTO-REPO |
Grok CLI evidence includes real MCP OWNER use, cwd/subdirectory behavior, automatic repository binding, and real worker handoff. Codex CLI is verified for MCP startup and automatic repository discovery; do not treat broader Codex host scenarios as verified unless explicitly tested.
Worker Harnesses
| Worker Harness | Support |
|---|---|
| Codex CLI + GPT-5.6 Luna | VERIFIED |
| GenericCliAdapter | VERIFIED CORE |
| DSH + DeepSeek V4 Flash Max | VERIFIED / EXPERIMENTAL UPSTREAM CAVEAT |
DSH evidence is real Engineering MCP E2E against DSH 0.1.2-rc.1 through a trusted DSH bridge/wrapper. Engineering MCP integrates DSH as the Harness and does not claim direct DeepSeek API integration. See examples/harnesses/dsh/README.md for the verified architecture reference.
Community / unverified
| Harness | Label |
|---|---|
| Claude Code | COMMUNITY RECIPE — NOT MAINTAINER E2E VERIFIED |
| Qwen Code | COMMUNITY RECIPE — NOT MAINTAINER E2E VERIFIED |
| Kimi Code | COMMUNITY RECIPE — NOT MAINTAINER E2E VERIFIED |
| Other Harnesses | COMMUNITY RECIPE — NOT MAINTAINER E2E VERIFIED |
Automatic repository binding
Default public story:
engineering-mcp --role owner
Resolution precedence:
1. --repo
2. ENGINEERING_MCP_REPO
3. launch cwd → git rev-parse --show-toplevel
4. fail closed
Properties:
- a repo subdirectory resolves to the Git root;
- a linked worktree resolves to its own worktree root;
- outside Git fails with
REPOSITORY_NOT_FOUND; - binding is immutable after process startup;
- no ledger is opened on failed auto-discovery.
OWNER workflow
An OWNER process creates tasks, delegates to a Worker Harness, awaits the terminal result, and can recover or cancel a stuck task. The Worker Runner owns claim-time repository checks, process execution, result schema validation, and Git verification.
Core OWNER tools include:
create_taskget_tasklist_active_taskslist_worker_profilesdelegate_taskawait_delegationrecover_taskresume_taskcancel_taskclose_task
Worker Harnesses
Workers claim tasks through the trusted Worker Runner. Harnesses never own Engineering MCP lifecycle state. A Harness may be a native adapter (for example Codex CLI) or a declarative Generic CLI adapter.
Worker Profiles
Worker Profiles are trusted operator configuration that select a preconfigured Harness execution profile. They are loaded once at process startup and are immutable for the process lifetime.
OWNER
→ delegate_task(worker_profile = "my-generic-worker")
→ trusted Worker Profile registry
→ adapter + manifest + opaque Harness config
→ Worker Runner
→ native Harness
Use:
engineering-mcp --role owner --worker-profiles /absolute/path/to/profiles.yaml
OWNER can list trusted profiles with list_worker_profiles, and can select one with delegate_task(worker_profile = "...").
Without a profile file, the built-in codex-luna default preserves existing behavior. Worker Profiles do not allow task text to choose arbitrary models, providers, executables, or credentials. See Worker Profiles.
GenericCliAdapter
GenericCliAdapter executes a local Harness from a declarative engineering-cli-adapter/1 manifest. Manifests are shell-free: command, arguments, working directory, prompt transport, and result transport are explicit. Task text cannot select executables or argv.
Use:
engineering-mcp adapter validate manifest.yaml
engineering-mcp adapter probe manifest.yaml
adapter probe validates the manifest and then runs the equivalent of <command> --version with a bounded timeout and local execution only. It does not invoke a model or perform an authenticated smoke. adapter smoke, if exposed, is explicitly opt-in and may consume model quota.
Safety model
Engineering MCP is not an OS sandbox.
Trusted operator configuration can intentionally launch local Harness executables. External Harnesses run with the user's OS permissions. Engineering MCP control mechanisms include:
- persistent task lifecycle;
- one RUNNING slot;
- execution-instance ownership;
- writer-generation fencing;
- repository binding;
- HEAD verification;
- allowed/forbidden scope verification;
- explicit recovery;
- fail-closed worker protocol.
Worker testimony is not Git authority. The Runner independently verifies Git state.
Recovery
Recovery is explicit and OWNER-only. If a Worker Runner disappears or a dispatch becomes orphaned, recover_task terminates the active dispatch atomically and returns the task to a delegable state. cancel_task performs a similar termination for cancellation.
Support tiers
- VERIFIED — maintainer real authenticated E2E.
- VERIFIED CORE — core mechanism has real E2E evidence, but not every third-party Harness.
- COMMUNITY — recipe/config exists but maintainer has no authenticated E2E.
- EXPERIMENTAL — evidence exists but upstream interface is unstable/developer-preview.
- PLANNED — not implemented or not validated.
Limitations
- one RUNNING task slot;
- no autonomous scheduler;
- no parallel RUNNING;
- no auto commit/merge;
- no remote runners or cloud service;
- no provider router;
- no web UI;
- no heartbeat/lease subsystem yet.
Contribution / docs
See:
Yorumlar (0)
Yorum birakmak icin giris yap.
Yorum birakSonuc bulunamadi