Always-on AI workstation that lives inside Claude Code. 142 agents · 366 skills · 200+ prompt templates across 16 categories · auto prompt matcher · auto skill matcher · 6x cheaper token routing.
# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/timdevai/proteus/main/install.sh | bash# Windows
iwr -useb https://raw.githubusercontent.com/timdevai/proteus/main/install.ps1 | iex| Layer | Count | Purpose |
|---|---|---|
| Runtime agents | 6 | Event-bus + orchestrator + memory/research/admin/content/code/trading handlers |
| Agent library | 136 | Curated from rohitg00/awesome-claude-code-toolkit (Apache 2.0) |
| Proteus-original skills | 13 | prompt-matcher, skill-matcher, scope-clarifier, cost-tracker, mcp-finder, permission-auditor, context-compactor, decision-recorder, repo-onboarder, prd-builder, pr-reviewer, commit-curator, tdd-flow |
| Community skills (rohitg00) | 40 | Apache 2.0 — TDD, React, K8s, Postgres, security-hardening, MCP-dev, prompt-engineering, etc. |
| Community skills (alireza) | ~313 | MIT — alirezarezvani/claude-skills: engineering, business-growth, c-level, compliance, finance, marketing, product, research, ra-qm |
| Prompt templates | 200+ | 16 categories — writing, research, coding, thinking, learning, business, creative, productivity, anti-sycophancy, context-discipline, eval-suite, token-optimizer, ai-engineer, trading |
| Token router | 1 | Routes cheap work to Kimi/Haiku/Ollama, premium to Opus. Documented 6x reduction. |
Every borrowed file traces back to its source in ATTRIBUTION.md. Their licenses are in LICENSES/.
Three problems most Claude users hit by month two:
- Wasted tokens. Premium models doing cleanup. Re-sending the same files every turn. 80% of the bill is preventable.
- No memory. Every session starts cold.
- No structure. You write whatever prompt comes to mind. The result is whatever Claude felt like that day.
Proteus is the answer: persistent agent runtime + curated skill/prompt library + routing brain.
curl -fsSL https://raw.githubusercontent.com/timdevai/proteus/main/install.sh | bashiwr -useb https://raw.githubusercontent.com/timdevai/proteus/main/install.ps1 | iexThe installer:
- Clones into
~/.proteus/ - Drops 13 skills into
~/.claude/skills/ - Copies the prompt library to
~/.proteus/_prompts/ - Appends 6 lines to
~/.claude/CLAUDE.mdto wire the auto-matchers - Optionally sets your
ANTHROPIC_API_KEY - Optionally starts the Proteus daemon
git clone https://github.com/timdevai/proteus ~/.proteus
cd ~/.proteus
pip install -r requirements.txt
cp -r skills/* ~/.claude/skills/
# Append CLAUDE.md.snippet contents to your ~/.claude/CLAUDE.md
python proteus.pyYou type:
debug why my agent keeps timing out after 30 seconds
Claude silently picks the Agent Debugger template from _prompts/ai-engineer.md, fills variables from your repo context, asks only what it can't infer, and executes. You get a debugged agent. You didn't write a prompt.
You type:
i need to roll out SOC2 next quarter
Claude scans 366 skills, finds compliance-os/soc2-prep, scores it 10/10, invokes it. You didn't have to remember it existed.
requirements.txt includes LiteLLM. Ships with this config:
default: kimi-2.6-instruct # most work
planning: claude-opus-4-6 # architecture, refactors
implementation: kimi-2.6-instruct # write code
cleanup: claude-haiku-4-5 # rename, format, doc
boilerplate: ollama/qwen3:7b # local, freeBring your own key. Expected savings: 6x vs default Sonnet.
| Agent | Triggers on | Does |
|---|---|---|
| memory | vault writes, file changes | indexes, dedupes, links |
| research | bookmarks, links, articles | summarizes, extracts ideas |
| admin | email, calendar | drafts replies, schedules |
| content | content-creation prompts | generates outlines, hooks, drafts |
| code | code-related prompts | architecture, refactor, review |
| trading | market data, trading prompts | strategy audit, risk check |
Each is a Python module in agents/. Add or swap freely.
Borrowed from rohitg00/awesome-claude-code-toolkit, preserved as-is with attribution. Cover:
- Core development — API designer, backend, fullstack, event-driven, refactoring (13)
- Language experts — TypeScript, Python, Rust, Go, Java, C#, Elixir, Ruby, etc. (25)
- Infrastructure — cloud architect, devops, k8s, SRE, deployment, network (11)
- Data & AI — ML engineer, data scientist, LLM architect, computer vision (16)
- Quality — accessibility, chaos, code-reviewer, compliance, security-audit, test-auto (10)
- Developer experience — CLI, build, docs, developer portal, dependency mgmt (15)
- Business & product — analyst, growth, customer success, legal, content strategy (12)
- Specialized domains — blockchain, fintech, gaming, e-commerce, embedded, IoT (15)
- Research — academic, competitive, market, benchmarking, data research (11)
- Orchestration — multi-agent coordinator, context mgr, error coordinator, task distributor (8)
40 from rohitg00 (skills/community/from-rohitg00/): accessibility, API design, auth patterns, AWS, CI/CD, database optimization, DevOps, Django, Docker, frontend, Git advanced, Golang, GraphQL, K8s, LLM integration, MCP dev, microservices, mobile, monitoring, Next.js, performance, Postgres, prompt engineering, Python, React, Redis, Rust, security hardening, Spring Boot, TDD mastery, testing strategies, TypeScript, WebSocket realtime, and more.
~313 from alireza (skills/community/from-alireza/): organized into 13 domain packs:
business-growth— proposals, contracts, customer successc-level-advisor— board prep, M&A, OKRs, CEO/CTO/CFO/COO/CHRO/CMO/CIO/CTO advisorscompliance-os— SOC2, HIPAA, GDPRengineering— agenthub, code tour, helm charts, LLM cost optimizer (24 packs)engineering-team— team-level engineering frameworksfinance— modeling, forecasting, audit prepmarketing/marketing-skill— campaigns, SEO, ad creativeproduct-team— PRDs, roadmapping, user researchproductivity— focus, weekly review, OKRsproject-management— RACI, kanban, statusra-qm-team— regulatory affairs + quality managementresearch— dossier builder, grant writer, literature review
13 Proteus-original (skills/proteus/): prompt-matcher, skill-matcher, scope-clarifier, tdd-flow, commit-curator, repo-onboarder, prd-builder, decision-recorder, cost-tracker, context-compactor, permission-auditor, mcp-finder, pr-reviewer.
┌──────────────┐
files/email ──│ Event Bus │── tier-0 ──→ skip 80% of noise locally
└──────┬───────┘
│
┌──────▼───────┐
│ Orchestrator │── routes by source + content classification
└──────┬───────┘
│
┌───────┬───────┼───────┬───────┬────────┐
▼ ▼ ▼ ▼ ▼ ▼
memory research admin content code trading
│
┌──────▼───────┐
│ Router │── picks cheapest model that can do the job
└──────┬───────┘
│
Anthropic / Moonshot / Ollama
- Python 3.10+
- Claude Code (free, claude.ai/claude-code)
ANTHROPIC_API_KEY(BYOK)- Optional: Ollama (for tier-0 filter + boilerplate routing)
- Optional: Obsidian vault if you want the memory agent to index into one
Will this break when Claude updates? No. Plain markdown + Python. Skills follow the documented Claude Code spec.
Do I need all 142 agents and 366 skills? No. The auto-matchers only activate what's relevant to your message. Inactive skills consume zero context.
Can I use it without the daemon? Yes. The skills work as standalone Claude Code skills. Agents add the always-on layer.
Why "Proteus"? Greek shape-shifter. The orchestrator changes shape based on the event.
Why MIT? Maximum adoption. Borrowed content keeps its original license (Apache 2.0 for rohitg00, MIT for alireza). See LICENSES/.
Can I use this with GPT / Gemini / Kimi? Prompts work everywhere. Agents call any LiteLLM-supported model. Skills are Claude-specific.
Most of the agent library and many of the skills are borrowed (with permission via their licenses) from outstanding community work. See ATTRIBUTION.md for full credit. The runtime (event bus, orchestrator, tier-0 filter, 6 domain agents, auto-matchers, token router, prompt library) is original.
- v0.1.0 — public launch (this release)
- v0.2.0 — web dashboard for event bus inspection
- v0.3.0 — Slack + Discord source adapters
- v0.4.0 — agent marketplace (community-contributed agents)
- v0.5.0 — local-first mode (zero API spend via Ollama)
PRs welcome. New agent? Add a module to agents/. New prompt? Drop a template in _prompts/*.md. New skill? Drop a SKILL.md in skills/proteus/ (or open a PR upstream to rohitg00 / alireza if it fits there).
Proteus runtime: MIT (LICENSE)
Borrowed agents (rohitg00): Apache 2.0 (LICENSES/rohitg00-LICENSE)
Borrowed skills (alireza): MIT (LICENSES/alireza-LICENSE)
If Proteus saves you 6x on your API bill, star the repo. That's the only thank-you I want.