openSmith

KO | EN

gitlite — search

#ai-agent #list #ai-agents #sql #swift #proxy #kubernetes #json #anthropic #ai

(★ 9)

Local Priority, Native Terminal Personal Assistant Agent.

파일 탐색기

ZIP 다운로드

- pre-push
- - coding.py
- - README.md
- - coding.yaml
  - README.md
  - smith.yaml
- - code-review.yaml
  - requirements-research.yaml
  - tdd-development.yaml
- - dangerous_commands.json
- - README.md
  - SOURCES.md
- - config.yaml
  - context.md
  - hooks.yaml
  - role.md
  - style.md
  - toolbox.md
  - workflow.md
- - explorer.yaml
  - implementer.yaml
  - reviewer.yaml
- - edit_file.py
  - get_current_time.py
  - git_ops.py
  - glob_files.py
  - grep.py
  - list_dir.py
  - memory_ops.py
  - read_file.py
  - read_pdf.py
  - render_pdf_page.py
  - render_ui.py
  - shell.py
  - skill_load.py
  - skill_manage.py
  - sub_agent.py
  - todo.py
  - web_crawl.py
  - web_fetch.py
  - web_search.py
  - write_file.py
- output_style.md
- README.md
- __init__.py
- config.py
- database.py
- hash_chain.py
- paths.py
- pyproject.toml
- uv.lock
- yaml_utils.py
- - 0001-approval-gated-host-capabilities.md
- - directory-access-limitation.md
- - 10-系统架构.md
  - 11-全链路白盒地图.md
  - 12-术语表.md
- - 03-Common-基础设施.md
  - 04a-ReAct-Loop-设计.md
  - 04b-LLM模块设计.md
  - 04c-Context-上下文系统.md
  - 04e-Execution-运行生命周期与管线.md
  - 04f-Tool-工具系统.md
  - 04g-Safety-安全与审批.md
  - 04i-MCP-外部工具协议.md
  - 04j-Observability-可观测性.md
  - 05-Engine-记忆系统.md
  - 07-Server-平台后端.md
  - 08-Shell-终端前端.md
  - README.md
  - wiki-01-总览.md
  - wiki-03-架构总览.md
  - wiki-04-Engine-核心执行.md
  - wiki-05-记忆系统.md
  - wiki-06-安全与安全边界.md
  - wiki-08-Agents-内容层.md
  - wiki-README.md
- - agent-memory-learning-upgrade.md
- - 01-产品定位.md
  - 02-快速上手.md
- - 40-Common.md
  - 41-Engine.md
  - 42-Agents.md
  - 43-Server.md
  - 44-Shell.md
- - 50-开发规范.md
  - 51-路线图.md
  - 52-竞品对比.md
- - 02-架构与借鉴.md
  - 03-与知识库的拆分与集成.md
  - 05-QoderWake技术调研.md
  - 06-Agent文件与技能系统设计参考.md
  - 07-QoderWake完整产品逆向.md
  - 08-QoderWake前端架构分析.md
  - 12-OpenHanako调研与借鉴.md
  - README.md
- - 2026-07-11-streaming-lifecycle-design.md
  - 2026-07-13-agent-memory-quality-research.md
  - 2026-07-15-claude-code-subagents-research.md
  - 2026-07-16-cli-component-rendering.md
  - 2026-07-17-pi-architecture-research.md
  - 2026-07-22-cli-table-rendering.md
  - 2026-07-22-shell-component-adoption.md
  - 2026-07-29-hermes-agent-memory-research.md
  - 2026-07-30-coding-skillchain-orchestration.md
  - 2026-08-07-pi-agent-design-review.md
  - loop-engineering.md
  - mainstream-agent-design-research.md
  - README.md
- - 20-Agent-Loop.md
  - 21-记忆系统.md
  - 22-上下文治理.md
  - 23-工具与安全.md
  - 24-子Agent委派.md
  - 25-LLM集成.md
  - 26-MCP集成.md
  - 27-可观测性.md
  - 28-身份与路由.md
  - 29-Skill技能系统.md
  - 30-Sandbox执行环境.md
- 00-文档阅读指南与表达规范.md
- check-links.py
- README.md
- - __init__.py
  - assembler.py
  - budget.py
  - compression.py
  - fitting.py
  - summary.py
- - __init__.py
  - assets.py
  - events.py
  - evidence.py
  - run_signature.py
  - runtime_control.py
- - __init__.py
  - catalog.py
- - __init__.py
  - client.py
  - config_fields.py
  - contracts.py
  - events.py
  - factory.py
  - model_config.py
  - observability.py
  - port.py
  - replay.py
  - usage.py
- - __init__.py
  - client.py
  - config.py
  - session_pool.py
- - __init__.py
  - _changeset.py
  - _files.py
  - _guards.py
  - _review.py
  - _snapshot.py
  - compile.py
  - dream.py
  - history.py
  - maintenance.py
  - MEMORY_POLICY.md
  - policy.py
  - store.py
  - user_learner.py
- - __init__.py
  - diagnosis.py
  - health.py
  - incidents.py
  - index.py
  - projections.py
  - proposals.py
  - reader.py
  - recorder.py
  - runtime.py
  - summary_store.py
  - trace_store.py
- - __init__.py
  - approval.py
  - eval_guard.py
  - fact_gate.py
  - risk.py
  - tool_guard.py
  - tool_policy.py
- - __init__.py
  - host.py
  - macos_seatbelt.py
- - __init__.py
  - executor.py
  - loader.py
  - registry.py
  - settings.py
  - store.py
- - conftest.py
  - test_full_audit_fixes.py
- - __init__.py
  - interface.py
  - ledger.py
  - registry.py
  - snapshot.py
  - truncation.py
- __init__.py
- AUDIT-LONGHORIZON-2026-08-29.md
- pyproject.toml
- uv.lock
- - __init__.py
  - main.py
- - agent-product-research.md
- - conftest.py
  - test_agent_profile_service.py
  - test_agent_router.py
  - test_agent_service.py
  - test_auth.py
  - test_auto_task_manual_stop.py
  - test_auto_task_repo.py
  - test_auto_task_service.py
  - test_common_infrastructure.py
  - test_config_loader.py
  - test_config_service.py
  - test_cron.py
  - test_e2e_smoke.py
  - test_generation_stats.py
  - test_health_nonce.py
  - test_llm_recording.py
  - test_mcp_service.py
  - test_observability_service.py
  - test_profile_files.py
  - test_project_instruction_service.py
  - test_run_service.py
  - test_scheduler.py
  - test_schema_identity.py
  - test_session_history_cutoff.py
  - test_session_schema.py
  - test_session_service.py
  - test_skill_service.py
  - test_token_stats_service.py
- pyproject.toml
- uv.lock
- - smith.js
- - run-tests.mjs
- - activity.test.ts
  - activity.ts
  - api.test.ts
  - api.ts
  - approval.test.ts
  - approval.ts
  - auth.test.ts
  - auth.ts
  - bridge.test.ts
  - bridge.ts
  - code-block.test.ts
  - code-block.tsx
  - commands.test.ts
  - commands.ts
  - composer-chunk.test.tsx
  - composer.test.ts
  - composer.tsx
  - conversation.test.ts
  - conversation.ts
  - dev-server.test.ts
  - dev-server.ts
  - diff-block.test.ts
  - diff-block.tsx
  - diff-parse.ts
  - history.test.ts
  - history.ts
  - hooks.test.ts
  - hooks.ts
  - hud.test.ts
  - hud.tsx
  - index.tsx
  - ink-static-cache.test.tsx
  - input.test.ts
  - input.ts
  - list-navigation.test.ts
  - list-navigation.ts
  - markdown-layout.test.ts
  - markdown-layout.ts
  - markdown-segments.test.ts
  - markdown-segments.ts
  - markdown-table.test.ts
  - markdown-table.tsx
  - model-picker.test.ts
  - model-picker.ts
  - multi-select-list.tsx
  - output.test.ts
  - output.ts
  - package-contract.test.ts
  - panel-components.test.tsx
  - panel-container.tsx
  - presentation.ts
  - queue.ts
  - run-panel.tsx
  - sanitize.test.ts
  - sanitize.ts
  - setup.test.ts
  - setup.ts
  - skill-mention.test.ts
  - skill-mention.ts
  - skill-presentation.ts
  - smith-ui-schema.test.ts
  - smith-ui-schema.ts
  - smith-ui.test.tsx
  - smith-ui.tsx
  - store.test.ts
  - store.ts
  - streaming-markdown.test.ts
  - streaming-markdown.ts
  - tabbed-panel.tsx
  - term.ts
  - text-layout.test.ts
  - text-layout.ts
  - theme.test.ts
  - theme.ts
  - token-panel.tsx
  - token-stats.test.ts
  - token-stats.ts
  - transcript-state.test.ts
  - transcript-state.ts
  - transcript.test.tsx
  - transcript.tsx
  - window-size.test.tsx
  - window-size.ts
- .npmignore
- .npmrc
- biome.json
- package-lock.json
- package.json
- tsconfig.build.json
- tsconfig.json
.gitignore
AGENTS.md
CLAUDE.md
CONTEXT.md
README.md

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

> A local-first, terminal-native agent workbench.

![Python](https://img.shields.io/badge/Python-3.11+-3776AB?style=for-the-badge&logo=python&logoColor=white)
![TypeScript](https://img.shields.io/badge/TypeScript-3178C6?style=for-the-badge&logo=typescript&logoColor=white)
![FastAPI](https://img.shields.io/badge/FastAPI-009688?style=for-the-badge&logo=fastapi&logoColor=white)
![SQLite](https://img.shields.io/badge/SQLite-003B57?style=for-the-badge&logo=sqlite&logoColor=white)
![React](https://img.shields.io/badge/React-61DAFB?style=for-the-badge&logo=react&logoColor=black)
![Anthropic](https://img.shields.io/badge/Claude-191919?style=for-the-badge&logo=anthropic&logoColor=white)
![OpenAI](https://img.shields.io/badge/OpenAI-412991?style=for-the-badge&logo=openai&logoColor=white)
![License](https://img.shields.io/badge/License-MIT-green?style=for-the-badge)

Smith is a single, always-on agent that runs on your machine. It keeps
conversation context, accumulates memory across sessions, and switches workflows
through skills. One agent, one conversation — no sub-agents, no routing layer to
reason about.

Everything runs locally: the shell, the API server, the execution engine, and
the SQLite database. The only outbound traffic goes to the LLM provider you
configure.

---

## Why One Agent

Multi-agent frameworks spend their complexity budget on orchestration —
delegation, hand-off protocols, shared scratchpads, and the failure modes each of
those introduces. openSmith spends it on the single agent instead:

| Concern | How openSmith handles it |
|---|---|
| Different task types | **Skills** — a task-specific workflow loaded into the prompt, not a separate agent |
| Multi-step work | **Skill chains** — declared pipelines with gates between stages |
| Long-lived context | **Memory** — compiled from evidence, not appended transcripts |
| Dangerous operations | **Tool guard + fact gate** — a non-bypassable boundary plus a challenge layer |
| Correctness of intermediate steps | **Gates** — a stage must produce its contract before the next one runs |

---

## What It Does

- **Interactive terminal** — a single rich Ink shell, launched with `smith` from any project directory
- **Skill-based workflows** — debug, plan, review, or reply directly, chosen per task
- **Skill chains** — requirements research, TDD development, and code review run as gated pipelines
- **Real tools** — file I/O, shell, Git, web search, and MCP servers, each behind permission levels
- **Persistent memory** — sessions, agent memory, and project context survive restarts
- **Multi-provider LLM** — OpenAI-compatible and Anthropic, routed by use case (interactive / gate / background)
- **Auditable by construction** — a tamper-evident hash chain over the audit log, plus structured execution events

---

## Quick Start

### Prerequisites

- Python 3.11+
- [uv](https://docs.astral.sh/uv/)
- Node.js 22+

### Install

```bash
# Backend
cd server && uv sync

# Terminal shell — build, then link `smith` onto your PATH
cd ../shell && npm ci && npm run build && npm link
```

If an older release left a `smith` from `uv tool install`, remove it first — it
points at the deleted Python CLI and shadows the shell entry point:

```bash
uv tool uninstall agent-smith-server
```

`npm link` symlinks the global `smith` to this working copy rather than copying
it, so `smith` works from any directory and finds `server/` through its own
install path. The trade-off: it breaks if the repository moves, and it picks up
source edits only after `npm run build`. Verify with:

```bash
ls -l "$(which smith)"   # → .../lib/node_modules/smith-shell/bin/smith.js
```

### Configure

Set your LLM provider through environment variables:

```bash
export AGENTSMITH_LLM_PROVIDER=openai          # openai / anthropic
export AGENTSMITH_LLM_API_KEY="sk-..."
export AGENTSMITH_LLM_BASE_URL="https://api.openai.com/v1"
export AGENTSMITH_LLM_MODEL="your-model"
```

Or write `~/.agent-smith/config.yaml`:

```yaml
llm:
  provider: openai
  api_key: sk-...
  base_url: https://api.openai.com/v1
  model: your-model
```

### Run

```bash
# From any project directory — the backend starts automatically
smith
```

### Context Files

`~/.agent-smith/SMITH.md` is your user-wide instruction file: it applies to every
run. A repository's own `.smith/SMITH.md` holds rules that belong to that project
only. Both are read while a run is assembled, so edits take effect on the next
request without restarting the backend.

Inside the shell, `/reload` starts a fresh session while keeping the previous one
in history, so the next task begins with the current context.

---

## Architecture

```mermaid
graph LR
    U[User] --> S[Shell · Ink / React]
    S -->|HTTP + SSE| V[Server · FastAPI]
    V --> E[Engine · execution runtime]
    E --> A[Agents · identity / skills / tools]
    E --> C[Common · paths / YAML / SQLite]
    E --> L[LLM provider]
    E --> M[MCP servers]
    L --> R[Execution events]
    M --> R
    R --> V
    V --> S
    S --> O[Response]
```

Dependencies flow one way — `server → engine → common`. The engine never imports
FastAPI, and `agents/` imports nothing from the other layers: it is loaded at
runtime, so its contract is file shape, not Python types.

| Layer | Directory | Responsibility |
|---|---|---|
| Infrastructure | `common/` | Paths, YAML config, SQLite connection, audit hash chain. No business logic. |
| Execution | `engine/` | Agent framework: LLM, ReAct loop, pipelines, memory, skills, tools, safety, observability. No platform knowledge. |
| Content | `agents/` | Smith's identity seed, pipelines, gates, conditions, tools, skills, hooks. Pure content. |
| Platform | `server/` | FastAPI app: orchestration, session and agent lifecycle, HTTP endpoints. |
| Terminal UI | `shell/` | Ink shell. Talks to the server over local HTTP, auto-starts the backend. |

---

## Repository Layout

```
openSmith/
├── agents/                     # Content layer — loaded at runtime, never imported
│   ├── conditions/             #   Predicates deciding whether a pipeline node runs
│   ├── gates/                  #   Stage contracts: what a node must produce to advance
│   ├── identities/             #   Identity and route declarations (smith.yaml, coding.yaml)
│   ├── pipelines/              #   Skill-chain definitions
│   ├── safety/                 #   Declarative safety rules
│   ├── skills/                 #   Built-in skills, one directory per SKILL.md
│   ├── smith/                  #   Smith's identity seed: role, style, workflow, hooks
│   ├── tools/                  #   Tool providers (TOOL_META + execute)
│   ├── output_style.md         #   Response style layer injected into the prompt
│   └── README.md
│
├── common/                     # Infrastructure layer — zero business logic
│   ├── config.py               #   Configuration loading
│   ├── database.py             #   SQLite connection management
│   ├── hash_chain.py           #   Tamper-evident audit log chain
│   ├── paths.py                #   Single source of truth for the data root (~/.agent-smith)
│   ├── yaml_utils.py           #   Safe YAML read/write
│   └── pyproject.toml
│
├── engine/                     # Execution layer — the agent framework
│   ├── context/                #   Prompt assembly (trust-tagged layers)
│   ├── execution/              #   Run lifecycle and control flow
│   │   ├── hooks/              #     Pre/Post/Stop tool-lifecycle hook framework
│   │   ├── orchestration/      #     Run lifecycle, agent dispatch
│   │   ├── pipeline/           #     Pipeline and skill-chain execution
│   │   ├── react/              #     ReAct loop
│   │   └── routing/            #     Lexical intent routing
│   ├── identity/               #   Identity catalog and route resolution
│   ├── llm/                    #   Provider adapters, model configuration
│   ├── mcp/                    #   MCP client: external tool protocol
│   ├── memory/                 #   Evidence log, compiler, guards, policy
│   ├── observability/          #   Structured events, traces, metrics
│   ├── safety/                 #   Tool guard (non-bypassable) and fact gate
│   ├── sandbox/                #   Host execution isolation (Seatbelt on macOS)
│   ├── skill/                  #   Skill discovery and loading
│   ├── tool/                   #   Tool registry and dispatch
│   ├── tests/                  #   Engine test suite
│   └── pyproject.toml
│
├── server/                     # Platform layer — FastAPI
│   ├── app/
│   │   ├── infrastructure/     #     App wiring, startup, dependencies
│   │   ├── routers/            #     HTTP endpoints (thin: extract, call, return)
│   │   ├── schemas/            #     Request and response models
│   │   ├── services/           #     Orchestration: engine runtime, sessions, profiles
│   │   ├── utils/
│   │   └── main.py             #     ASGI entry point
│   ├── docs/
│   ├── tests/
│   └── pyproject.toml
│
├── shell/                      # Terminal UI — Ink / React
│   ├── bin/smith.js            #   CLI entry point
│   ├── src/                    #   Components, transcript rendering, API bridge
│   ├── scripts/                #   Test runner
│   ├── biome.json              #   Lint and format configuration
│   └── package.json
│
├── docs/                       # Documentation
│   ├── wiki/                   #   Primary design docs — 14 chapters, start here
│   ├── adr/                    #   Architecture decision records
│   ├── analysis/               #   Analyses and investigations
│   ├── capabilities/           #   Capability notes
│   ├── reference/              #   Reference material
│   ├── research/               #   Research notes
│   ├── superpowers/            #   Agent technique notes
│   └── 00–11 *.md              #   Earlier documentation set (not kept in sync)
│
├── .githooks/                  # Repository git hooks
├── AGENTS.md                   # Brief for coding agents working in this repo
├── CLAUDE.md                   # Working brief: architecture, boundaries, verification
└── CONTEXT.md                  # Short project context
```

---

## Documentation

**Start with [`docs/README.md`](docs/README.md)** — the reading map. Every topic
has exactly one authoritative document, all of it written against the source:

| Where | What |
| --- | --- |
| [`docs/guide/`](docs/guide) | What Smith is, and getting it running |
| [`docs/architecture/`](docs/architecture) | Layers, one request end to end, glossary |
| [`docs/subsystems/`](docs/subsystems) | How each subsystem runs — loop, memory, context, tools & safety, sub-agents, LLM, MCP, observability, routing, skills, sandbox |
| [`docs/layers/`](docs/layers) | What each code layer is responsible for |
| [`docs/project/`](docs/project) | Conventions, roadmap, external comparisons |

Superseded drafts live in [`docs/archive/`](docs/archive) and are **not current
fact** — each carries a banner naming the document that replaced it and why.
Writing and diagram conventions are in
[`docs/00`](docs/00-文档阅读指南与表达规范.md). When documentation and code
disagree, the code wins.

---

## Development

```bash
cd engine && uv run --extra test pytest tests   # Engine tests
cd server && uv run --extra dev  pytest tests   # Server tests
cd shell  && npm test && npm run check          # Shell tests + typecheck
```

Run the backend in the foreground when you need it directly:

```bash
cd server && uv run uvicorn app.main:app --port 8000
```

Note for non-macOS contributors: a number of engine tests exercise the macOS
Seatbelt sandbox and are skipped elsewhere. A Seatbelt test that *fails* rather
than skips on Linux is missing its platform marker.

---

## Contributing

New capability? Add a **skill**, not an agent. The architecture boundaries in
[`CLAUDE.md`](CLAUDE.md) are enforced by the import graph and by tests — keep
`engine/` free of HTTP concepts, and keep `agents/` importing nothing.

---

## License

MIT

openSmith

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Download README