Friday Studio AI 런타임: 프롬프트, 기술 및 도구를 안정적인 구성으로 전환

AI agent orchestration platform. Workspaces run autonomous agents triggered by signals (HTTP, cron). The daemon manages workspace lifecycles — each workspace defines agents, signals, and jobs in a workspace.yml . Signals arrive (HTTP, CLI, cron), the daemon routes them to the workspace runtime, which spawns sessions where agents execute with MCP tool access. | Tool | Version | Why | |---|---|---| | Deno | 2.7.0+ | Runs the daemon, CLI, and TypeScript packages | | Go | 1.26+ | Builds the Go services under tools/ (pty-server, webhook-tunnel, friday-launcher) | | Node.js | 24+ | Needed for npx , Vite, and the web playground | | uv | 0.11+ | Provisions the managed Python that user agents run under (auto-installed by setup-dev-env.sh if absent) | | git | any recent | — | | Docker (optional) | any recent | Alternative path: run the full stack with docker compose up | You do not need Postgres for local development — the daemon uses SQLite by default. Postgres is only required when running the link credential service in production. git clone https://github.com/friday-platform/friday-studio cd friday-studio deno install # install JS/TS deps (also runs husky via prepare hook) cp .env.example .env # open .env and set ANTHROPIC_API_KEY (or another provider key) The example file documents every variable the daemon reads. The minimum to run a real agent is one LLM provider key. deno task atlas daemon start --detached Verify it's up: curl -sf http://localhost:8080/health && echo " daemon ok" deno task atlas daemon status Send a prompt through the CLI — the daemon routes it to the bundled chat workspace and returns a chat id you can follow up on. deno task atlas prompt "Write a haiku about TypeScript" deno task atlas chat # list recent chats deno task atlas chat --human # readable transcript To run one of the bundled example workspaces (HTTP-triggered Claude Code agent): deno task atlas workspace add ./examples/claude-code-smoke curl -X POST http://localhost:8080/webhooks/run-code \ -H "Content-Type: application/json" \ -d '{"prompt": "explain this repo in two sentences"}' Browse examples/ for more — pr-review-github , jira-bugfix-labeled , claude-code-with-skills , and others — each is a self-contained workspace.yml you can copy and edit. deno task atlas daemon stop docker compose up This runs the daemon, web UI (Studio), credential service, PTY server, and webhook tunnel together. Ports default to 1xxxx to avoid host collisions — see docker-compose.yml . For interactive development with the Svelte UI, hot-reload daemon, and webhook tunnel running side by side: deno task dev:playground Open http://localhost:5200. Working in-tree (running the daemon out of the repo, not via the desktop installer) needs a one-time setup so the daemon's user-agent spawn path finds a managed Python with the SDK installed: bash scripts/setup-dev-env.sh Idempotent. Safe to re-run after every friday-agent-sdk version bump. What it does: - Verifies uv is on PATH (installs from astral.sh if not). - Writes the env vars the daemon needs ( FRIDAY_UV_PATH ,UV_PYTHON_INSTALL_DIR ,UV_CACHE_DIR ,FRIDAY_AGENT_SDK_VERSION ) into the daemon's envfile (~/.atlas/.env or~/.friday/local/.env , matching whichever home holds your live state). - Pre-warms the uv cache — Python 3.12 + the pinned friday-agent-sdk wheel — so the first user-agent spawn doesn't pay the download as cold-start latency. - Uninstalls any stale editable friday-agent-sdk installs from system Pythons that would otherwise shadow the uv-managed copy. Restart your daemon after running it so the new env vars take effect. The desktop installer handles the same setup automatically; this script is for in-tree development only. # Daemon lifecycle deno task atlas daemon start --detached # start (background) deno task atlas:dev daemon start # start with auto-restart (session-aware) deno task atlas daemon status # health check deno task atlas daemon stop # stop # Interact deno task atlas prompt "your prompt" # send a one-shot prompt deno task atlas chat # list recent chats deno task atlas chat --human # show transcript # Develop deno task typecheck # deno check + svelte-check deno task lint # deno lint + biome check --write deno task fmt # biome format --write deno task test $file # run a vitest file # Go services go test -race ./... golangci-lint run apps/ atlasd/ # Daemon — HTTP API, workspace lifecycle (TS/Deno) atlas-cli/ # CLI entry point — `deno task atlas` (TS/Deno) link/ # Credential / OAuth service (TS/Deno) ledger/ # Resource & activity storage service (TS/Deno) studio-installer/ # Friday Studio launcher app — tray, daemon # supervisor, autostart (Tauri/Rust) packages/ # @atlas/* libraries — core, agent-sdk, config, # fsm-engine, llm, logger, mcp, memory, skills, # storage, workspace, signals, … examples/ # Self-contained workspace.yml examples # (claude-code-smoke, pr-review-github, # jira-bugfix-labeled, …) tools/ agent-playground/ # Web client (SvelteKit) — dev tool and the # production UI bundled in Friday Studio evals/ # Agent eval harness CLI friday-launcher/ # System tray launcher + daemon supervisor (Go) pty-server/ # WebSocket → PTY shell bridge (Go) webhook-tunnel/ # Cloudflare tunnel → daemon webhook forwarder (Go) Config: friday.yml (platform-wide) · workspace.yml (per-workspace) · CONTRIBUTING.md (dev guidelines + code style) A workspace agent declared type: "user" is a Python file the daemon spawns as a subprocess. The agent code calls capabilities (LLM, HTTP, MCP tools, streaming) over NATS through the friday-agent-sdk package — no provider keys, no MCP plumbing in your code. Minimal agent: from friday_agent_sdk import AgentContext, agent, ok @agent(id="hello", version="1.0.0", description="Reverses input.") def execute(prompt: str, ctx: AgentContext): return ok({"reversed": prompt[::-1]}) if __name__ == "__main__": from friday_agent_sdk import run run() Register and run: curl -X POST http://localhost:8080/api/agents/register \ -H "Content-Type: application/json" \ -d '{"entrypoint":"/abs/path/to/agent.py"}' curl -X POST "http://localhost:8080/api/agents/hello/run?workspaceId=user" \ -H "Content-Type: application/json" \ -d '{"input":"hello"}' The daemon spawns uv run --python 3.12 --with friday-agent-sdk== agent.py under the hood — the version is pinned in tools/friday-launcher/paths.go (bundledAgentSDKVersion ) and threaded through to the daemon via the launcher (or setup-dev-env.sh for in-tree work). Bumping the SDK is a deliberate change in three pin sites: that constant, the same value in the Dockerfile , and BUNDLED_AGENT_SDK_VERSION in apps/studio-installer/src-tauri/src/commands/prewarm_agent_sdk.rs . Rule of thumb: use type: "user" only when each call's decision is mechanical (regex / schema / fixed routing). For any LLM-judgment work (classifying, summarizing, choosing among options), use type: "llm" with MCP tools — that's faster to author and easier to maintain. Reference: - SDK: friday-platform/agent-sdk · authoring guide vendored atpackages/system/skills/writing-friday-python-agents - Memory access from agents: narrative is the only supported strategy today — see packages/system/skills/writing-to-memory - Spawn resolution: apps/atlasd/src/agent-spawn.ts — three-tier fallback (uv-run →FRIDAY_AGENT_PYTHON → barepython3 ) Claude Code skills support a structured planning-to-execution pipeline: - Design — /brainstorming refines a rough idea into a design doc via Socratic questioning. Outputs todocs/plans/ . - Iterate — /improving-plans critiques the design, surfaces gaps, and outputs an improved version. Keep iterating until the questions become trivial. - Tasks — /creating-tasks converts the validated design into tracked work items with dependency graphs. - Execute — /implementing-tasks spawns parallel agents. Each claims a task, implements, commits, moves on. - Polish — /polishing runs a self-review team (lint, slop, tests, design) before PR. - Ship — /opening-pr creates a PR with summary and test plan. Post-ship: /reviewing-code for reviewing others' PRs. docs/ARCHITECTURE.md — components and data flowCONTRIBUTING.md — how to send patches, code style, hard rules (CLA required)SECURITY.md — vulnerability disclosure Friday is source-available under the Business Source License 1.1. You can read, modify, and self-host the code under the Additional Use Grant, which permits free production use for personal use, organizations under 5 people, and businesses with under $1M ARR. Production use outside those bounds, or that competes with Tempest Labs' offering, requires a commercial license. Each released version converts automatically to Apache-2.0 one year after that version is first distributed (so the current version's Change Date is 2027-04-30; later versions get their own date). For commercial-license inquiries: [email protected]. Third-party components retain their original licenses — see THIRD_PARTY_LICENSES.md and NOTICE (includes MPL-2.0 §3.2 source-availability notice and license elections for dual/tri-licensed deps).