Harvest

Did you Harvest that? 🌾

Build your AI Workforce with Harvest - the background coding agent that tends to your repositories like farmers tending to crops, cultivating code that grows stronger while you sleep.

Harvest is a background coding agent service built on the architecture that powers Ramp's Inspect, designed specifically for the Relevance AI ecosystem.

---

Status at a Glance

Component	Status	Documentation
Daytona Executor	🚧 MVP Ready	packages/daytona-executor/
Claude Agent SDK	✅ POC Validated	trials/daytona-sdk-poc/
Claude CLI Integration	✅ Ready	docs/ai/
Git Workflow	✅ Ready	docs/ai/shared/git-workflow.md
MCP Servers	✅ Ready	MCP table
API Layer	⏳ Planned	docs/plans/IMPLEMENTATION_PLAN.md
Slack/Web Client	⏳ Planned	docs/plans/IMPLEMENTATION_PLAN.md

---

Core Capabilities

• Autonomous Development: Works continuously across repositories without human intervention
• Sandbox Orchestration: Spins up isolated Daytona sandboxes for each session
• Claude Code CLI Integration: Leverages Anthropic's official Claude CLI with full tool access
• Multi-Model Strategy: Primary Claude Sonnet 4.5 + Gemini for plan validation
• Session Continuity: SQLite-backed conversation state persists across sandbox restarts
• Continuous Harvesting: Generates pull requests, fixes bugs, and improves code while you focus on other work

---

System Architecture

Architecture Status: Daytona + Claude Agent SDK

Current Phase: MVP implementation - Docker snapshot image with pre-installed SDK and MCP servers

See packages/daytona-executor/ARCHITECTURE.md for detailed architecture diagrams.

graph TD
    subgraph "User Interface"
        UI[Relevance Chat UI]
        TV[HarvestToolViewer]
    end

    subgraph "External Triggers"
        SL[Slack Bot]
        GH[GitHub Webhooks]
    end

    subgraph "Relevance Backend"
        TR[TriggerRunner]
        CM[ConversationManager]
        HP[HarvestPresetAgent]
        HR[HarvestRuntime.ts]
        MM[MessageMapper]
    end

    subgraph "Daytona Cloud"
        DS[Daytona SDK]
        SB[Sandbox<br/>harvest-snapshot]
    end

    subgraph "Inside Sandbox"
        SDK[Claude Agent SDK]
        QF[query function]
    end

    subgraph "Claude Agent"
        CA[Claude Sonnet 4.5]
        TL[Tools: Bash, Read, Write, Edit]
        MCP[MCP Servers]
    end

    subgraph "External Services"
        GHA[GitHub API]
        GEM[Gemini API]
        LIN[Linear API]
        GIT[Git Operations]
    end

    UI -->|Send prompt| TR
    SL -->|Event| TR
    GH -->|Webhook| TR

    TR --> CM
    CM --> HP
    HP --> HR

    HR -->|daytona.create| DS
    DS -->|Spawn| SB
    HR -->|codeRun| SB

    SB --> SDK
    SDK --> QF
    QF --> CA

    CA --> TL
    TL --> GHA
    TL --> GIT
    CA --> MCP
    MCP --> GEM
    MCP --> LIN

    CA -->|JSON stream| QF
    QF -->|stdout| HR
    HR -->|Parse| MM
    MM -->|Yield| CM
    CM -->|Stream| UI
    CM -->|Tool calls| TV

    style UI fill:#e3f2fd
    style TV fill:#e3f2fd
    style SB fill:#fff3e0
    style SDK fill:#fff3e0
    style CA fill:#f3e5f5
    style TL fill:#f3e5f5

Loading

Key Architectural Decisions

• Runtime: Daytona sandboxes with Claude Agent SDK
• Session Model: SDK manages session continuity via resume parameter
• Message Format: Structured JSON stream from SDK
• Config Strategy: Baked into snapshot image, loaded at runtime via SDK's systemPrompt
• Security: Non-root harvest user, secrets injected via envVars

Key benefits:

• TypeScript end-to-end
• Structured message types
• Built-in session management and cost tracking
• Fast cold starts with pre-built snapshot

Agent Modes

Harvest uses shared base + mode-specific extensions architecture:

Local Development Mode (Your machine):

• Interactive brainstorming for complex tasks
• Can pause for human feedback
• Human judgment available

Autonomous Agent Mode (Daytona sandbox):

• Maximum autonomy, no human in loop
• Execute without asking
• Fail forward pattern (try alternatives when blocked)

Both modes share execution rules (docs/ai/shared/*.md) but differ in intent/approach.

---

MCP Servers

Server	Package	Required Secret	Permissions	Where to Get
Memory	@modelcontextprotocol/server-memory	None	-	-
Filesystem	@modelcontextprotocol/server-filesystem	None	-	-
Playwright	@playwright/mcp	None	-	-
DevTools	chrome-devtools-mcp	None	-	-
GitHub	@modelcontextprotocol/server-github	GITHUB_TOKEN	Contents: R/W, PRs: R/W, Issues: R/W	github.com/settings/tokens
Linear	mcp-remote → https://mcp.linear.app/sse	LINEAR_API_KEY	Full access	linear.app/settings/api
Gemini	@houtini/gemini-mcp	GEMINI_API_KEY	-	aistudio.google.com/app/apikey
Sentry	@sentry/mcp-server	SENTRY_AUTH_TOKEN	project:read, event:read, issue:read	sentry.io/settings/account/api/auth-tokens

---

Getting Started

Quick Setup

git clone https://github.com/RelevanceAI/harvest.git
cd harvest
pnpm install

Prerequisites

• Node.js 20+
• pnpm package manager
• Docker (for Daytona executor)
• GitHub account with PAT
• Claude Code CLI (claude.ai/download)

Note: Harvest previously used Modal Labs for cloud sandboxes. We transitioned to Daytona because Modal containers run as root, and the Claude Agent SDK blocks autonomous operation (bypassPermissions) with root privileges for security reasons. Daytona supports non-root containers, resolving this architectural constraint.

Claude Code Setup

Enable Harvest AI rules and superpowers skills:

claude plugins add claude-plugins-official https://github.com/anthropics/claude-plugins-official
claude plugins install claude-plugins-official/superpowers
cp .claude/settings.json.template .claude/settings.json

Configure MCP servers in ~/.claude/mcp.json - see MCP table for API keys and permissions.

Testing & Code Quality

# Run validation scripts
pnpm run validate

# Test Daytona snapshot image (requires Docker)
pnpm run test:snapshot

---

Documentation

Directory	Purpose
docs/plans/	Implementation plans
docs/ai/	AI agent rules and workflows
docs/mcp/	MCP server documentation
.claude/	Project-level AI agent rules

---

Contributing

Use Claude, it knows everything it needs to know.

---

License

MIT