Skip to content

thesongzhu/Friday

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

518 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

中文

Friday

A personal AI that learns your work, uses skills, and acts with approval.
Chat with it, give it goals, connect your tools, and let repeated work become auditable workflows.
Local-first · BYOK · Approval-first · Human-controlled · AGI-Like

Node >=22 MIT @thesongzhu/friday CI


What Is Friday?

Friday is a self-hosted personal AI and automation runtime.

It is meant to feel less like a blank chatbot and more like a private execution partner: you give a goal, Friday checks what it can do, uses configured capabilities, asks for the human-only pieces, executes, verifies the result, and records what it learned. Capability acquisition and self-upgrade flows are active work and should be treated as review-gated WIP rather than a fully autonomous promise.

Friday is not a magic fully autonomous system. It will not create accounts for you, bypass CAPTCHA, pay for services, take production-changing actions, or use credentials you have not provided. Its job is to do the work it can safely do, stop clearly when it needs you, and leave evidence behind.

The Product Loop

  1. You give a goal. Example: "Read these PDFs, compare them with the latest web results, and send a short summary to Discord."
  2. Friday checks capabilities. It looks for text, vision, OCR, web search, PDF, file, browser, channel, model, memory, and workflow support.
  3. Friday reports or drafts gap closure. It can search installed skills, trusted catalogs, MCP servers, local files, package registries, OpenAPI specs, and the web for candidate tools or skills, but install/update paths remain policy-gated and under active hardening.
  4. Friday asks when humans are required. API keys, OAuth, paid plans, CAPTCHA, logins, sensitive permissions, and high-risk actions go through a human gate.
  5. Friday runs and verifies. It executes through skills, tools, workflows, browser/desktop control, or channel adapters, then checks the result.
  6. Friday learns safely. It can update memory, routing preferences, setup recipes, generated skills, eval cases, and failure lessons. It does not train model weights by default.

What Friday Can Do

Area What Friday is built to do Boundary
Chat and task execution Answer, plan, execute tool-backed work, show progress, and recover from failures Depends on configured providers and granted tools
Text, vision, OCR, PDF, files Route work to configured providers or built-in parsers, then report what is missing when a lane is unavailable Vision/OCR/TTS depend on provider support and credentials
Web and browser work Use configured web search providers, local browser control, and workflow steps Login, payment, CAPTCHA, and sensitive accounts require the user
Skills and workflows Import, validate, install, run, verify, update, and roll back reusable skills/workflows; generated/self-upgraded flows are WIP Untrusted code must pass review, sandbox checks, and policy gates
Memory and self-improvement Store preferences, lessons, provider routing signals, recipes, evals, and recovery notes User-visible, auditable, and reversible; no hidden model training
Self-healing Detect failures, propose fixes, run low-risk repairs, verify, roll back, and pause after repeated failures High-risk changes require approval
Multi-channel control Connect channels such as Discord, Telegram, Feishu/Lark, Slack-style webhooks, Signal, WhatsApp, and QQ where configured Channels can control Friday, but sensitive actions still require confirmation
Long-running goals Run user-authorized standing goals, create agenda items, gather evidence, and report outcomes Friday is goal-driven by the user; it does not invent unrelated long-term agendas

Setup

Option 1 - npm package

npm install -g @thesongzhu/friday
friday start
# Open http://localhost:3141

Option 2 - from source

git clone https://github.com/thesongzhu/Friday.git
cd Friday
bash scripts/ops/friday-first-run.sh
# Or on macOS, double-click "Friday Setup.command"

The first-run helper installs dependencies, builds Friday, starts the local runtime, and opens the setup page. On macOS it also installs the login startup agent and menu bar companion so Friday can come back after restart. If you run it from Desktop, Documents, or Downloads, it prepares the launchd-safe runtime at ~/Friday automatically before installing the login agents.

Option 3 - manual source start

git clone https://github.com/thesongzhu/Friday.git
cd Friday
npm install
npm run build
npm start
# Open http://localhost:3141

Option 4 - Docker from source

docker compose -f docker/docker-compose.yml up --build
# Open http://localhost:3141

First-run setup guides you through providers, channels, local permissions, and optional capabilities. After setup, Friday should open directly to Home.

Providers And Keys

Friday is BYOK: you bring the model and search/API credentials you want to use.

The setup flow should answer four questions for every capability:

  • Do I have this capability?
  • If not, what exactly is missing?
  • Where do I configure it?
  • After configuration, can Friday verify it with a real task?

Typical provider lanes include OpenAI, Doubao/Volcengine, Moonshot, Anthropic, Google, OpenRouter, Tavily, Serper, local browser/PDF/file tooling, MCP servers, and custom skills. A missing key or account is not treated as success; Friday should show it as a human blocker with the next configuration step.

Capability Self-Acquisition (WIP)

When Friday does not have a capability for a goal, the target closed loop is:

goal -> capability gap -> candidates -> sandbox/test -> approval if required -> install/register -> doctor verify -> execute

This loop is not yet a blanket production guarantee. Current builds can report gaps and exercise parts of the workflow, while generated skills, self-upgrades, and adjustment-fidelity paths remain review-gated and covered by ongoing stress tests. Allowed automatic steps depend on policy. Searching, analysis, draft generation, and sandbox verification are low-risk by default. Downloading code, installing packages, writing config, registering tools, shell access, and external network calls are governed by the autonomy policy. OAuth, payment, CAPTCHA, API keys, sensitive permissions, and production writes always require the user.

Memory, Growth, And Repair

Friday's self-improvement is intentionally boring and inspectable:

  • Memory facts: preferences, project rules, recurring context, and user corrections.
  • Routing preferences: which provider worked best for which kind of job.
  • Recipes: setup and recovery steps that worked.
  • Skills and workflows: generated or improved artifacts with tests and evidence.
  • Eval cases: fixed scenarios that should keep passing.
  • Failure lessons: what broke, what fixed it, and when to stop retrying.

This is not model-weight training. It is auditable state and reusable artifacts that the user can inspect, edit, pause, or delete.

Safety Model

Friday follows a simple rule: automate the boring parts, keep the user in control of irreversible or sensitive parts.

  • Local-first runtime and local state by default.
  • Bring-your-own keys; credentials should live in environment variables, managed secret refs, or OS-backed secret storage.
  • Capability grants have scope, reason, evidence, and expiry.
  • High-risk actions require explicit approval.
  • Tool calls, workflow runs, self-healing runs, and channel actions leave audit evidence.
  • Failed installs or generated capabilities must support rollback.
  • Public network exposure requires explicit auth, CORS, TLS/proxying, and least-privilege access.

What Friday Is Not

  • It is not a guarantee of universal automation.
  • It is not a fully autonomous system that can do every task without you.
  • It does not bypass logins, CAPTCHA, payment, platform rules, or provider limits.
  • It does not safely run arbitrary third-party code without review.
  • It does not remove your responsibility to secure the host, API keys, accounts, and connected channels.

Documentation

Download And Distribution

Platform Method Status
macOS / Linux / Windows npm install -g @thesongzhu/friday Published on npm as 1.0.0
Source git clone + npm install + npm run build Available
Docker docker compose -f docker/docker-compose.yml up --build Available from this repo

The official npm package is @thesongzhu/friday. The unscoped friday package on npm is unrelated.

Community

License

Friday is open-source software under the MIT license.

Third-Party Notices

Friday includes compatibility and adaptation work for third-party agent ecosystem formats and behavior. See NOTICE for preserved upstream copyright and license notices.