Friday
A local control plane that routes AI work with approvals, verification, and audit evidence.
- Source repo
- thesongzhu/Friday
- Stars
- ★ 839
- Last updated
- 2mo ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 53/100 · Major gaps
At a glance
- Works with
- Universal · cross-platformCodex · Claude.ai
- You'll need
- Typical use
- A solo developer using several model APIs can hand Friday a long-running software goal and inspect or override the model chosen for each step.
- Main limitation
- It currently runs from source; there is no publicly released native Friday.app or documented consumer installer.
- Source review
- 53/100 · Major gaps
What does this agent do, and when should you use it?
Friday is a private control plane for AI agents, not a model or chatbot. Its controller, keys, data, and memory are intended to live on the local machine, while BYOK connects to cloud AIs such as Codex, Claude, and DeepSeek. Its stated loop is route, execute, verify, and remember: select a model, govern risky work through approval, require evidence of completion, and propose memories for confirmation. The README describes hash-chained audit records for routed turns, including actor, content, and metering information. The current product runs from source; the npm package is explicitly not a complete consumer installation path, and the native Friday.app is not publicly released.
After a user supplies a goal, Friday follows the documented route → execute → verify → remember loop. It selects an AI for a step, executes governed work, and puts risky actions into an approval-or-rejection path. Completion requires a proof receipt rather than a model's assertion alone; the README describes this as enforced through a persisted mission/work-item code boundary. It presents memories as candidates that require user confirmation before becoming stored memories, and says recalled content is scrubbed for known secret patterns before being shown to a model. Routed agent turns are described as all-or-nothing entries in a hash-chained audit log with execution and cost details.
- A solo developer using several model APIs can hand Friday a long-running software goal and inspect or override the model chosen for each step.
- A technical lead automating operational work can allow ordinary read-only actions while requiring a separate approval for actions that change external state.
- A team that needs stronger evidence for agent output can require a proof receipt before a work item is treated as complete.
- A BYOK user managing spend across providers can route work before execution and retain a record of the model that actually answered after a fallback.
- A privacy-conscious individual can review each proposed memory candidate before allowing it to influence later work.
How do you install or deploy this agent?
The README provides this macOS source-run path:
brew install node@22
git clone https://github.com/thesongzhu/Friday.git
cd Friday
open "Friday Setup.command"It requires Node.js 22 or later and your own model-provider keys. The supplied evidence does not document a first CLI invocation, credential configuration format, or a non-macOS installation path. It also explicitly states that npm install does not install the full Friday product.
How do you use this agent?
Run "Friday Setup.command" after the documented source checkout. Then submit a goal through the described chat interface; Friday processes it through route, execute, verify, and remember, pausing risky steps for approval. You must provide your own keys for the named cloud AIs, including Codex, Claude, or DeepSeek. The supplied evidence does not document a copyable first-goal command, chat endpoint, or configuration-file format.
What are this agent's strengths and limitations?
- Keeps the controller, keys, data, and memory on the local machine while supporting BYOK cloud models from multiple vendors.
- Separates risky actions into human approval instead of allowing an agent to approve its own side effects.
- Treats proof receipts as the basis for task completion instead of relying only on an agent's self-report.
- Describes routing, metering, and hash-chained audit records for tracing execution and model cost.
- It currently runs from source; there is no publicly released native Friday.app or documented consumer installer.
- The source tree is 1.0.3 while the npm package is 1.0.2, and the README says the package does not carry the complete current engine.
- It requires Node.js 22+, a local runtime environment, and user-supplied cloud-model keys rather than local model execution.
- The README says live behavior has so far run only on internal and self-test traffic, not real end-user workloads.
- More autonomous capabilities are off by default, while capability acquisition and self-upgrade remain review-gated work in progress.
How does this agent compare with similar options?
Friday positions itself as a local, cross-vendor control layer: Codex and Claude are presented as individual agent providers, while Friday handles routing, approval, verification, memory-candidate confirmation, and auditing across them.
Key facts side by side with the most closely related agents.
| Agent | Source review | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|
| Friday This agent | 53 · Major gaps | ★ 839 | 2mo ago | TypeScript | Codex · Claude.ai |
| taOS Self-Hosted Agent OS | 71 · Some gaps | ★ 546 | 1d ago | Python | Claude Code · OpenAI API · Claude API |
| Kontext | 83 · Good | ★ 221 | 1d ago | Go | Codex · Claude Code |
| Node9 | 82 · Good | ★ 216 | 1d ago | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence: README emphasizes local-first, BYOK, approval-first, human-controlled, evidence-backed; code shows approval flows, audit logs, hash chains, secret pattern checks. Deductions: data flow transparency only partially shown (e.g., memory candidate mechanism), rollback not explicit, source attribution relies on README claims.
Evidence: CI includes build, test, migration, security, contract gates; tests cover unit, integration, adversarial. Deductions: failure message quality not evident in source, dependency availability only indirectly via CI retry.
Evidence: README describes multiple use cases (chat, coding, channels, scheduled), capability boundaries (skill sandbox, trust dial). Deductions: trigger precision (natural language triggers) only mentioned, environment fit (macOS, Node 22) documented but not all platforms.
Evidence: README, LICENSE, package.json, docs directory, install notes (Quickstart), known limitations (npm not consumer path). Deductions: naming stability (version 1.0.3 vs npm 1.0.2) inconsistent, no changelog, maintenance responsibility unclear.
Evidence: output usability (receipts, audit logs) designed, marginal value (cross-vendor, verification) articulated. Deductions: cost-benefit (token spend, pricing) not quantified, output usability (UI screenshots) not in source.
Evidence: README claims correspond to code structure (tests, CI), adversarial tests and evidence policy exist. Deductions: cross-source corroboration limited (single repo), fact-inference separation unclear (claims vs implementation not verified line-by-line).
- This is an early candidate; npm package version lags source, and no real end-user workloads have been run.
- Publisher identity unverified; dependency security relies on CI checks, no vulnerability scan report provided.
- Rollback mechanism and failure message specifics not evident in source; further review needed.