# Harbor

> Harbor is an all-in-one agent execution layer. It moves tools, credentials, code execution, memory, artifacts, jobs, apps, workflows, and traces out of the model context into one workspace-scoped runtime. Surfaces span a web dashboard (apps/web), a runtime API (apps/api), the hrbr CLI (apps/cli), and the cli-mcp stdio adapter (packages/sdk/cli-mcp). Exec paths create workspace-scoped run records with spans for tool calls, runtime output, plugin dispatch, and orbit access where applicable.

> This is the FULL variant of Harbor's llmstxt.org index. For the short index see [/llms.txt](/llms.txt).

## Product Overview

- [Homepage](https://tryharbor.ai/): Marketing entry point. Tagline: an all-in-one agent execution layer. Move tools, credentials, code execution, memory, artifacts, jobs, apps, workflows, and traces outside the model context with workspace-scoped state built in.
- [Why Harbor](https://docs.tryharbor.ai/documentation/why-harbor): Concrete failure modes of doing tool wiring inside the model context — repeated auth, weak audit, no shared state across agents, no durable shared execution record, no workspace tenancy.
- [Mental Model](https://docs.tryharbor.ai/documentation/mental-model): Plugin then Tool then Run then Trace. Three layers — workspace management (web + API), execution layer (isolated hrbr exec code with workspace-scoped credential access), and ingress (hrbr CLI, cli-mcp, remote MCP).
- [Documentation Overview](https://docs.tryharbor.ai/documentation/overview): Top-level docs navigation covering concepts, for-agents, plugins, guides, and reference.

## Getting Started

- [Quickstart](https://docs.tryharbor.ai/documentation/quickstart): From signup to first plugin invoke in under five minutes. Install with `npm i -g @zonko-ai/harbor`, run `hrbr login`, then submit your first `hrbr exec`.
- [Sign up and Install hrbr](https://docs.tryharbor.ai/guides/get-started/sign-up-and-install-cli): Create your Harbor account, install the CLI, log in. Two minutes end to end.
- [Install Your First Plugin](https://docs.tryharbor.ai/guides/get-started/install-your-first-plugin): Pick a service from the catalog, complete OAuth, run your first tool.
- [Your First Exec](https://docs.tryharbor.ai/guides/get-started/your-first-exec): Write multi-step TypeScript that calls real plugin tools.
- [Your First Function](https://docs.tryharbor.ai/guides/get-started/your-first-function): Promote a working exec snippet into a named, versioned, reusable Function.
- [Your First App](https://docs.tryharbor.ai/guides/get-started/your-first-app): Wrap a Function in an App and get a public URL.
- [Your First Skill](https://docs.tryharbor.ai/guides/get-started/your-first-skill): Author a workflow-backed skill the agent can read and run.
- [Connect Endpoint](https://tryharbor.ai/connect): Public deep-link surface that agent harnesses send users through to connect Harbor.

## Concepts

- [Workspaces](https://docs.tryharbor.ai/documentation/concepts/workspaces): The top-level isolation boundary for every Harbor resource. Memberships, roles, shared credentials, shared state, and audit logging are all scoped per workspace.
- [Plugins](https://docs.tryharbor.ai/documentation/concepts/plugins): Workspace-scoped integrations (Linear, GitHub, Cloudflare, Notion, Google services, Sentry, Slack, Postgres, Stripe, Vercel, Convex, and more). Each plugin exposes one or more sources and a typed tool schema.
- [Exec](https://docs.tryharbor.ai/documentation/concepts/exec): One-off TypeScript execution against your workspace tools. `hrbr exec` submits code to a workspace-scoped sandbox where plugin tools, orbit primitives, and shared state are available, and every run is traced.
- [Functions (Beta)](https://docs.tryharbor.ai/documentation/concepts/functions): Named, versioned, schedulable TypeScript that runs against your workspace. Published with `defineOrbitJob` and `hrbr exec -f ./job.ts`. Callable from other exec runs, apps, or schedulers.
- [Apps (Beta)](https://docs.tryharbor.ai/documentation/concepts/apps): Long-lived routed apps deployed via `defineOrbitApp`. Stable URLs, app-local state, public or internal auth modes, backed by workspace runtime primitives.
- [Skills](https://docs.tryharbor.ai/documentation/concepts/skills): Agent-facing markdown that teaches an LLM when and how to use Harbor. Bundled and rideable through the hrbr install.
- [Orbit Runtime (Beta)](https://docs.tryharbor.ai/documentation/concepts/orbit-runtime): Workspace-scoped primitives available inside hrbr exec, functions, and apps: `hrbr.storage`, `hrbr.cache`, `hrbr.db`, `hrbr.ai`, `hrbr.tools`, `hrbr.jobs`, and app-handler `hrbr.state`.
- [Runs and Traces](https://docs.tryharbor.ai/documentation/concepts/runs-and-traces): Exec paths record runs with spans for tool calls, sandbox stdout/stderr, plugin dispatch, and orbit access where applicable. Durable and inspectable.

## Plugins

- [Plugins Overview](https://docs.tryharbor.ai/plugins/overview): 1,000+ integrations (992 live as of 2026-05-29) spanning MCP servers, REST/GraphQL APIs, and CLIs. Install in one command, run their tools from any agent. How Harbor models third-party services as MCP-style sources with discovery, authentication, and tool dispatch.
- [Plugin Catalog](https://docs.tryharbor.ai/plugins/catalog): Browse every Harbor plugin. Filter by kind, category, or search. Lists canonical sources and exposed tool surfaces.
- [Custom MCP Plugin](https://docs.tryharbor.ai/guides/authoring/custom-mcp-plugin): Register your own MCP server as a Harbor plugin.

## Guides — Authoring

- [Apps (Beta)](https://docs.tryharbor.ai/guides/authoring/apps): Deep dive on defineOrbitApp: routes, auth modes, public and internal split.
- [App UI (Beta)](https://docs.tryharbor.ai/guides/authoring/app-ui): @hrbr/orbit/app-ui templates and primitives — server-rendered HTML without React.
- [Functions (Beta)](https://docs.tryharbor.ai/guides/authoring/functions): Deep dive on defineOrbitJob: shape, context, error handling, return types.
- [Local CLI Bridge (Sand)](https://docs.tryharbor.ai/guides/authoring/sand-cli-bridge): Bridge a local CLI (git, gh, wrangler) from a remote Harbor run.
- [Skills — Template Anatomy](https://docs.tryharbor.ai/guides/authoring/skills-template-anatomy): Walk through the canonical SKILL.md template field by field.
- [Skills — Wiring Tool Calls](https://docs.tryharbor.ai/guides/authoring/skills-frontmatter-and-tool-calls): Bind workflow steps to exact plugin tools, with argument templating.
- [Skills — Validation and Publish](https://docs.tryharbor.ai/guides/authoring/skills-validation-and-publish): Run the validator, fix common errors, publish to personal then workspace scope.
- [Skills — When to Write One](https://docs.tryharbor.ai/guides/authoring/skills-when-to-write): Decide between a Skill, a Function, an App, or just exec.

## Guides — Operations and Recipes

- [Observability](https://docs.tryharbor.ai/guides/operations/observability): Find runs, drill into traces, and inspect workspace execution history.
- [Sharing Across Team](https://docs.tryharbor.ai/guides/operations/sharing-across-team): Workspace versus personal scope, role permissions, common sharing flows.
- [Versioning and Pinning](https://docs.tryharbor.ai/guides/operations/versioning-and-pinning): How @vN versions work for Functions and Apps, plus rollback patterns.
- [Disable vs Delete](https://docs.tryharbor.ai/guides/operations/disable-vs-delete): When to disable a Function, App, Skill, or Plugin, and when to actually delete.
- [Incident Triage](https://docs.tryharbor.ai/guides/recipes/incident-triage): Sentry then GitHub then Linear then Slack. End-to-end working example.
- [Daily Digest](https://docs.tryharbor.ai/guides/recipes/daily-digest): Scheduled Function posts a daily roll-up of workspace activity to Slack.
- [PR Review Bot](https://docs.tryharbor.ai/guides/recipes/pr-review-bot): Always-on PR reviewer that listens to GitHub events, summarizes the diff, posts inline comments.
- [Release Notes](https://docs.tryharbor.ai/guides/recipes/release-notes): Weekly release notes auto-generated from GitHub commits.
- [OAuth Approval Workflow](https://docs.tryharbor.ai/guides/recipes/oauth-approval-workflow): step.waitForEvent — park a run until an approver clicks a button.
- [Multi-step with Retries](https://docs.tryharbor.ai/guides/recipes/multi-step-with-retries): step.do + step.sleep for durable, retryable, time-paced pipelines.
- [HTML Report from Orbit](https://docs.tryharbor.ai/guides/recipes/html-report-from-orbit): Generate a styled HTML report and serve it from a signed Orbit storage URL.
- [Public Intake Form](https://docs.tryharbor.ai/guides/recipes/public-intake-form): Public feedback form that drops submissions into Linear.

## For Agents

- [For Agents — Overview](https://docs.tryharbor.ai/documentation/for-agents/overview): How AI agents consume Harbor. Distinguishes control plane (`hrbr inspect`) from runtime (`hrbr exec`), and explains why Harbor splits the two.
- [Connecting Agents](https://docs.tryharbor.ai/documentation/for-agents/connecting): Wire agent harnesses to Harbor via the `cli-mcp` adapter (`hrbr serve mcp`) and remote endpoints. Covers Claude Code, Cursor, Codex, Gemini CLI, and generic MCP clients.
- [Agent Skills (well-known)](https://docs.tryharbor.ai/documentation/for-agents/agent-skills): Canonical Harbor agent skill bundle served at /.well-known/agent-skills/harbor/, plus topic skills for plugins, exec, jobs, apps, and orbit.
- [llms.txt for Agents](https://docs.tryharbor.ai/documentation/for-agents/llms-txt): How Harbor's docs site exposes an llmstxt.org-style index, and how agents should consume it.
- [Harbor Meta Skill (markdown)](https://tryharbor.ai/skill.md): The canonical Harbor agent meta-skill body served by apps/web/app/skill.md/route.ts. Also available as /skills.md. Used by Claude Code and other harnesses on install.
- [Harbor SKILL.md (well-known)](https://docs.tryharbor.ai/.well-known/agent-skills/harbor/SKILL.md): Static well-known skill copy for harnesses that prefer the .well-known location.
- [Connect Claude Code](https://docs.tryharbor.ai/guides/get-started/connect-claude-code): Point Claude Desktop at Harbor's MCP server.
- [Connect Cursor](https://docs.tryharbor.ai/guides/get-started/connect-cursor): Point Cursor at Harbor's MCP server.
- [Connect Codex](https://docs.tryharbor.ai/guides/get-started/connect-codex): Point Codex CLI at Harbor's MCP server.
- [Connect Gemini CLI](https://docs.tryharbor.ai/guides/get-started/connect-gemini-cli): Point Gemini CLI at Harbor's MCP server.

## Reference

- [Glossary](https://docs.tryharbor.ai/documentation/reference/glossary): Every Harbor term, defined once — workspace, plugin, source, tool, run, request, agent, orbit, registry, job, app, span, MCP, cli-mcp, hrbr.
- [Limits](https://docs.tryharbor.ai/documentation/reference/limits): Plan-level and technical limits across the Harbor surface — execution timeouts, payload sizes, sandbox memory ceilings, plugin call concurrency, workspace quotas.
- [Security Model](https://docs.tryharbor.ai/documentation/reference/security): Credential handling, sandbox isolation, OAuth trust boundaries, WorkOS / AuthKit, tool policy, workspace isolation.
- [Harbor SDK on GitHub](https://github.com/zonko-ai/harbor-sdk): Public SDK and CLI source.

## Endpoints

- [Protected MCP Endpoint](https://mcp.tryharbor.ai/mcp): Harbor's protected MCP endpoint. It advertises OAuth protected-resource metadata; verify authenticated tool workflows for the specific client.
- [API Catalog](https://tryharbor.ai/.well-known/api-catalog): RFC 9727 linkset describing Harbor API surfaces on the marketing domain.
- [Agent Skills Index](https://tryharbor.ai/.well-known/agent-skills/index.json): JSON manifest of agent skills exposed on the marketing domain.

## Site

- [Blog](https://tryharbor.ai/blogs): Engineering and product writing about Harbor, the MCP ecosystem, agent runtime design, and Cloudflare-native architecture.
- [Skills Index](https://tryharbor.ai/skills): HTML index of Harbor agent skills exposed at this domain.
- [Docs](https://tryharbor.ai/docs): Server-side redirect to docs.tryharbor.ai. The /docs path on the marketing domain is a static-export redirect to the canonical Harbor documentation site.
- [Templates](https://tryharbor.ai/templates): Hub of starter recipes for hrbr exec — coming soon as GitHub repos. Cards list planned repos with a Coming-soon badge; underlying repos do not exist yet.
- [Integrations](https://tryharbor.ai/integrations): Live MCP plugin catalog rendered on the marketing domain from the compiled-in @hrbr/registry-catalog package. Every connection is workspace-scoped, per-tool authorized, and traced.
- [Community](https://tryharbor.ai/community): Harbor's community surfaces — GitHub (source, issues, discussions), X / Twitter (updates), and the Harbor Discord at https://discord.gg/zMH7XKgaz5 (chat).

## Use Cases

- [Use Cases Hub](https://tryharbor.ai/use-cases): Workflow-shaped surfaces of the same Harbor execution layer: one workspace, scoped tools, typed exec, artifacts, and traced runs. Indexes the four agent workflows Harbor markets — coding-agent (canonical) plus ops-agent, data-agent, and customer-ops (coming soon).
- [Coding Agent](https://tryharbor.ai/use-cases/coding-agent): Canonical workflow page. Harbor gives your coding agent workspace-scoped tools: GitHub, package managers, and your private services arrive as scoped tools, runs execute in Harbor, and every action is traced. Includes problem/solution/secondary/FAQ sections.
- [Ops Agent](https://tryharbor.ai/use-cases/ops-agent): Coming-soon shell. A scoped workspace for the agents that page on-call, file incidents, and act on your platform tools, on the same execution layer as the coding-agent surface.
- [Data Agent](https://tryharbor.ai/use-cases/data-agent): Coming-soon shell. Workspace-scoped credentials, typed exec, and a per-run trace for every query and transform, on the same primitives Harbor already exposes to coding agents.
- [Customer Ops](https://tryharbor.ai/use-cases/customer-ops): Coming-soon shell. Scoped helpdesk, CRM, and billing tools, exec in a sandbox, every interaction recorded for review. Underlying workspace + plugin model already shipped.

## Pricing and Sales

- [Pricing](https://tryharbor.ai/pricing): Three tiers — Free, Workspace, Enterprise. Pricing in private beta; numeric prices intentionally omitted until a unit and rate ship.
- [Enterprise](https://tryharbor.ai/enterprise): Workspace-scoped agent execution layer for orgs needing SSO via WorkOS / AuthKit, dedicated workspace, per-tool policy governance, audit log export, and Cloudflare Workers for Platforms isolation for Orbit jobs. SOC 2 on roadmap; self-host on roadmap.
- [Security](https://tryharbor.ai/security): Concrete security posture — WorkOS / AuthKit, workspace tenancy, tool policy, KV-backed credential state under SECRETS_ENCRYPTION_KEY, audited runs with spans where applicable, sandbox isolation via Cloudflare codemode, HSTS preload + XFO + XCTO + RP + PP edge headers, Content-Signal ai-train=no / ai-retrieval=yes. Roadmap: SOC 2, HIPAA, on-prem, BYOK. Vulnerability disclosure: founders@zonko.ai.

## Compare

- [Compare Hub](https://tryharbor.ai/compare): On-domain index of every Harbor comparison page (Composio, Nango, Runlayer, Arcade, Infisical, Pipedream) with a one-line factual positioning summary per competitor. Backed by CollectionPage + ItemList + BreadcrumbList JSON-LD pointing at each /compare/<slug> route.
- [Harbor vs Composio](https://tryharbor.ai/compare/composio): Honest side-by-side. Harbor differentiates on workspace tenancy, Cloudflare-native runtime, and typed exec; Composio leads on toolkit breadth, JTBD use-cases, and published trust evidence. Includes TechArticle + FAQPage + ItemList JSON-LD plus primary source citations.
- [Harbor vs Nango](https://tryharbor.ai/compare/nango): Honest side-by-side. Different caller — Nango is an integration SDK for backend code; Harbor is an all-in-one execution layer for agents. Harbor leads on workspace tenancy and traced exec; Nango leads when backend-owned integrations and self-hostable open-source core are the requirement.
- [Harbor vs Runlayer](https://tryharbor.ai/compare/runlayer): Direct shape peer comparison. Runlayer composite 58.60 (rank 3) vs Harbor 39.57. Runlayer leads on security-domain vocabulary (OWASP MCP Top 10, tool poisoning, DCR), named advisors, and structured-data depth. Harbor leads on Cloudflare-native runtime and AI-crawler citation share (Runlayer's robots.txt blocks GPTBot / ClaudeBot / CCBot; Harbor allow-lists them).
- [Harbor vs Arcade](https://tryharbor.ai/compare/arcade): Honest side-by-side. Arcade composite 51.80 (rank 4) vs Harbor 39.57. Arcade leads on technical (74.4, rank 2) and content / E-E-A-T (62.0 with named LangChain customer). Harbor differentiates on workspace tenancy, runs/traces as a first-class product object, and Cloudflare-native operations.
- [Harbor vs Infisical](https://tryharbor.ai/compare/infisical): Cross-layer comparison flagged explicitly — Infisical is not in the Phase 1b benchmark cohort. Infisical is an open-source secrets management platform (vault, dynamic secrets, rotation, PAM, PKI, Kubernetes Agent Injector) with two newer agent products: Agent Vault (HTTP credential proxy that attaches secrets at the edge so agents never hold tokens) and Agent Sentinel (governs outbound MCP servers for agent workloads, including OAuth dynamic client registration). Harbor is an all-in-one agent execution layer where credentials are one primitive among several. The two products own different layers and can compose: Infisical holds long-lived secrets while Harbor runs the agent workspace.
- [Harbor vs Pipedream](https://tryharbor.ai/compare/pipedream): Pipedream is a developer-first integration / workflow-automation platform with a large pre-built connector catalogue (2,500+/3,000+ apps), serverless Node.js / Python / Go / Bash code steps, Pipedream Connect (per-end-user OAuth for SaaS / AI-agent products), and hosted MCP servers at mcp.pipedream.com. Pipedream announced on 2025-11-19 a definitive agreement to be acquired by Workday. Harbor is an all-in-one agent execution layer; the two products enforce different tenancy boundaries (per-end-user vs. per-workspace).

## Legal

- [Privacy Policy](https://tryharbor.ai/privacy): How Harbor handles user data and workspace content. Workspace data is tenancy-isolated; credential state is KV-backed.
- [Terms of Service](https://tryharbor.ai/terms): Terms governing use of Harbor and the hrbr CLI / @zonko-ai/harbor package.

## Key Facts

- Product name: Harbor.
- Company: Zonko AI (zonko-ai org on GitHub).
- Tagline: An all-in-one agent execution layer.
- Core thesis: Give agents tools, credentials, code execution, memory, artifacts, jobs, apps, workflows, and traces in one workspace-scoped runtime.
- Architecture: Cloudflare-native workspace management and runtime — Workers, D1, KV, R2, Vectorize, Workflows, Durable Objects.
- Primary surfaces: apps/web (dashboard), apps/api (control plane + runtime), apps/cli (hrbr), packages/sdk/cli-mcp (cli-mcp stdio adapter).
- Execution layer: server-side TypeScript execution for hrbr exec code, with workspace-scoped credential access, plugin tools, and orbit primitives available; exec paths create auditable run records.
- Plugin model: workspace-scoped MCP-style sources, with shared OAuth client config and per-tool authorization.
- Auth: WorkOS / AuthKit for web sign-in; protected-resource metadata for the inbound MCP endpoint.
- Distribution: `npm i -g @zonko-ai/harbor` installs the hrbr CLI and bundled agent skills.
- Protected MCP endpoint: https://mcp.tryharbor.ai/mcp advertises OAuth protected-resource metadata; authenticated client workflows should be verified per client.
- Public docs: https://docs.tryharbor.ai
- Public SDK: https://github.com/zonko-ai/harbor-sdk

## Contact

- Website: https://tryharbor.ai
- Docs: https://docs.tryharbor.ai
- GitHub: https://github.com/zonko-ai/harbor-sdk