SenHarness

An open-source, multi-tenant implementation of the Harness Engineering paradigm for enterprise AI agents.

English · 简体中文

---

What is SenHarness?

If you've ever watched an AI agent run for 30 minutes, forget its own plan, burn through your budget, and hallucinate a fix to a problem that never existed — you know why the model alone is not enough.

Harness Engineering is the discipline of designing everything outside the model: the file system, sandboxed shell, memory, retrieval, context governance, orchestration, guardrails, approvals. It's the agent's harness — the reins that keep it on track.

SenHarness is the Harness Engineering runtime for enterprises. One code base, two deployment modes:

• Self-hosted single-enterprise — one company, docker compose up, done
• Multi-tenant SaaS — same binary, many tenants, physical isolation

All under MIT license.

---

The Harness, in one line

SenHarness ships six layers around the model — context · tools · orchestration · memory · evaluation · constraints — so the agent stays on plan, on policy, and on budget.

---

What makes SenHarness different

1. Pluggable Agent Runtime. The AgentBackend protocol + gateway means any runtime — the bundled NativeBackend, OpenClaw remote workers, or a future agent framework — ships as a single adapter file.
2. Multi-tenant from day one. Every table is workspace_id-scoped. The same binary serves a single company on-prem or thousands of tenants in SaaS. No rewrite when you scale.
3. MIT license. Take it, ship it, fork it, build on it. No patent clause gotchas, no AGPL copyleft. The enterprise edition adds compliance packs and SLAs on top — the core stays free.

Under the hood: PostgreSQL + pgvector · Redis · FastAPI · SQLAlchemy 2 async · Next.js 15 + shadcn/ui · and the pydantic-ai ecosystem (harness · shields · backends · skills · subagents · todo · middleware · summarization).

---

Who is this for?

You are...	SenHarness gives you...
An employee at a company that deploys it	A personalized AI assistant that remembers your preferences, respects company rules, and asks for approval on sensitive operations
A company admin	Workspace setup in 10 minutes — invite members, configure LLM providers, wire up Slack/Feishu, set up audit trails
A developer / platform engineer	A Harness layer you can extend: plug new tools via MCP, write a custom Agent Runtime adapter, ship your own skill pack
An enterprise IT buyer	Multi-tenant RBAC · audit · HITL approval · vault with envelope encryption · pluggable keyring (env / file / passphrase / AWS KMS / GCP KMS / Azure KV / Vault)

---

Quick start

Prereqs: Docker 24+ · 4 GB free RAM · any LLM API key (OpenAI / Anthropic / DeepSeek / Ollama...)

git clone https://github.com/senweaver/SenHarness.git
cd SenHarness
cp .env.example .env
# Edit .env: set JWT_SECRET_KEY, DB_PASSWORD, REDIS_PASSWORD, plus at least one LLM key

docker compose up -d

# First-time setup (inside the backend container)
docker compose exec backend python -m cli.commands migrate
docker compose exec backend python -m cli.commands seed
docker compose exec backend python -m cli.commands create-admin

# Open http://localhost:3000

---

Core concepts

Concept	Description
Workspace	A tenancy / governance boundary. Hosts members, roles, departments, agents, knowledge, policies, audit. Use one workspace for a single company, multiple for a group or SaaS.
Department	Organizational tree inside a workspace. Any depth. Used for ownership, approval routing, role-based search.
Identity	A global account (email + password + optional OAuth + optional TOTP). One identity → membership in N workspaces with different roles.
Agent	The configurable unit an employee talks to. UI label configurable per workspace (assistant / digital employee / AI partner / ...).
Squad	A group of agents working together under a coordinating agent.
Skill Pack	Anthropic Agent Skills compatible SKILL.md — progressive loading.
Toolbox	Built-in + MCP + plugin tools, with per-agent binding.
Flow	Scheduled or triggered automation (cron · webhook · on_message · manual).
Channel	IM integrations (Slack · Feishu · Discord · generic webhook + more in V2).
Vault	Envelope-encrypted credential store, pluggable keyring.
Policy	Autonomy level (L1/L2/L3) · shields · budgets · tool ACL · approval workflows.
Agent Runtime	Pluggable execution backend.

---

Pluggable Agent Runtime

The AgentBackend protocol lives in backend/app/agents/kernels/base.py:

class AgentBackend(Protocol):
    backend_kind: str
    async def run(self, req: RunRequest) -> AsyncIterator[RunEvent]: ...
    async def cancel(self, run_id: uuid.UUID) -> None: ...
    def capabilities(self) -> BackendCapabilities: ...

Officially supported runtimes:

Runtime	Kind	Transport	Status
NativeBackend	native	In-process	✅ Stable
OpenClaw remote worker	openclaw	Gateway + long-poll over HTTPS	✅ Stable
Your custom runtime	your-backend	anything	Community-supplied

Built-in channel providers

Channels are the second pluggable layer — they bring IM platforms into SenHarness. Same registry pattern as Agent Runtimes.

Provider	Kind	Inbound auth	Outbound	Status
Slack	slack	v0 HMAC + 5-min replay window	chat.postMessage	✅
Feishu / Lark	feishu	verification_token (1.0/2.0)	Open API tenant token	✅
Discord	discord	Ed25519	Discord REST API	✅
DingTalk (钉钉)	dingtalk	HMAC-SHA256 + 60-sec replay window	custom-robot webhook	✅ V2
WeCom (企业微信)	wecom	SHA1 message-signature + AES-CBC payload	message/send REST	✅ V2
Generic webhook	webhook	shared inbound_token only	— (inbound-only)	✅
Your provider	your_kind	your call	your call	Community-supplied

---

Repository layout

SenHarness/
├─ backend/              FastAPI + SQLAlchemy + pydantic-ai
│  ├─ app/
│  │  ├─ agents/         Agent core (kernels / harness / tools / skills)
│  │  ├─ api/            REST + WebSocket routes
│  │  ├─ core/           config / security / middleware / rate limit / errors
│  │  ├─ db/             async engine / models / repositories
│  │  ├─ schemas/        Pydantic DTOs
│  │  ├─ security/       JWT / keyring / envelope encryption
│  │  ├─ services/       business services
│  │  └─ workflows/      flows / triggers / scheduler
│  ├─ tests/             unit / integration / e2e
│  └─ scripts/           ops / dev / probes
├─ frontend/             Next.js 15 · shadcn/ui · Tailwind 4
├─ docs/                 architecture · adapters · whitepapers · quickstart
└─ docker-compose*.yml   dev / prod / frontend-only compositions

---

Contributing

SenHarness welcomes contributions. See CONTRIBUTING.md for the development workflow, coding standards, and how to get your adapter listed in the official registry.

---

License

MIT © 2026 SenHarness Contributors

---