# Jazz > Jazz is an open-source AI agent harness that runs a general-purpose AI agent on your own machine — > terminal, scripts, cron, CI, Telegram, Discord. Self-hosted and fully local-capable: > 22 LLM providers, including offline models via Ollama and llama.cpp. > MIT licensed. Install with a single curl command; ships as a self-contained binary for macOS and Linux. Install: curl -fsSL https://github.com/lvndry/jazz/releases/latest/download/install.sh | bash npm package: https://www.npmjs.com/package/jazz-ai Source: https://github.com/lvndry/jazz Full docs in one file: https://lvndry.github.io/jazz/llms-full.txt ## Docs - [Getting started with Jazz](https://lvndry.github.io/jazz/docs/getting-started.md): Install Jazz, configure a model, create your first AI agent, and choose the next capability or deployment surface to add. - [Creating agents](https://lvndry.github.io/jazz/docs/getting-started/create-an-agent.md): Configure a Jazz agent for a specific job: model, provider, persona, skills, and tool policy, saved as one JSON file per agent. - [Run Jazz with local and air-gapped models](https://lvndry.github.io/jazz/docs/getting-started/local-models.md): Run Jazz with self-hosted Ollama, llama.cpp, vLLM, or SGLang; understand offline mode and network isolation. - [Quick start](https://lvndry.github.io/jazz/docs/getting-started/quick-start.md): Install the Jazz CLI with one curl command and run your first AI agent in minutes. Single self-contained binary for macOS and Linux: no Node or npm required. - [Uninstalling Jazz](https://lvndry.github.io/jazz/docs/getting-started/uninstall.md): Remove Jazz completely: scheduled jobs, the daemon service, chat bots, keyring entries, the ~/.jazz data directory, and the binary, in the order that leaves nothing firing. - [CLI reference](https://lvndry.github.io/jazz/docs/commands.md): Jazz command and flag reference for agents, conversations, providers, workflows, and unattended runs. - [Jazz concepts](https://lvndry.github.io/jazz/docs/concepts.md): Explore Jazz capabilities and how they work: agents, tools, skills, memory, goals, automation, model companions, peers, remote handoff, and budgets. - [Peers](https://lvndry.github.io/jazz/docs/concepts/agent-to-agent.md): Connect trusted agents as peers, control what their requests can access, set per-peer budgets, and inspect exchanges with the peer log. - [Agents](https://lvndry.github.io/jazz/docs/concepts/agents.md): Create and configure Jazz agents with their own models, personas, tools, skills, memory, context limits, and media companions. - [Artifacts](https://lvndry.github.io/jazz/docs/concepts/artifacts.md): Find files a Jazz agent produced, understand their generation labels, and retrieve artifact paths from terminal, JSON, and chat output. - [Browser Use](https://lvndry.github.io/jazz/docs/concepts/browser-use.md): Let an agent browse, read, and act on real web pages with the browser tools, and what the browser can and cannot reach. - [Budgets](https://lvndry.github.io/jazz/docs/concepts/budgets.md): Cap a Jazz run by iterations, cost, tokens, or wall-clock time, watch what it is spending with /cost and /limit, and know when a cap can be crossed. - [Compositions](https://lvndry.github.io/jazz/docs/concepts/compositions.md): Create charts, dashboards, calculators, and interactive HTML pages with Jazz compositions, then open or share the saved files. - [Computer use](https://lvndry.github.io/jazz/docs/concepts/computer-use.md): Experimental: let an agent look at and act in desktop apps you grant on your Mac, while you watch. What to set up, what each app allows, and what it can and cannot reach. - [Context management](https://lvndry.github.io/jazz/docs/concepts/context-management.md): How Jazz manages long agent runs: context limits, automatic compaction, saved work state, tool-result retrieval, run budgets, and stall detection. - [Conversations, state, and memory](https://lvndry.github.io/jazz/docs/concepts/conversations-and-memory.md): The five kinds of state a Jazz agent carries, who writes each one, and how long it lasts. - [Daemon](https://lvndry.github.io/jazz/docs/concepts/daemon.md): Run the Jazz daemon for background goals and loops, HTTP requests, webhooks, and peers. Install it as a service and manage authentication and spend limits. - [Deferred work](https://lvndry.github.io/jazz/docs/concepts/deferred-work.md): How a Jazz agent starts work that outlives the turn it was asked in, then resumes the same conversation when that work is done or due. - [Remote handoff](https://lvndry.github.io/jazz/docs/concepts/detach.md): Move a chat conversation to your own SSH server, close the laptop, watch and steer it from anywhere, then bring it and its file changes back. - [Goals and loops](https://lvndry.github.io/jazz/docs/concepts/goals-and-loops.md): Keep Jazz working toward an objective until it is verifiably done with /goal, hand it to the daemon when you leave, and rerun a prompt on a schedule with /loop. - [Lexicon](https://lvndry.github.io/jazz/docs/concepts/lexicon.md): Look up Jazz terms for agents, conversations, runs, tools, memory, goals, loops, approvals, context management, and generated files. - [Model companions](https://lvndry.github.io/jazz/docs/concepts/media.md): Compose one Jazz agent from specialist models for image, audio, and video understanding and generation without changing its identity, tools, or memory. - [Why memory works this way](https://lvndry.github.io/jazz/docs/concepts/memory-design.md): Why Jazz memory learns what you state and the lessons you accept, keeps tastes in scopes, and stays in files you can read, correct, and forget. - [Personas](https://lvndry.github.io/jazz/docs/concepts/personas.md): What a Jazz persona is, how it shapes behaviour and can narrow an agent's tools, the placeholders it can use, and where custom personas come from. - [Skills](https://lvndry.github.io/jazz/docs/concepts/skills.md): What a Jazz skill is, how progressive loading keeps a large skill library off the context budget, where skills come from, and which one wins on a name collision. - [Starting runs](https://lvndry.github.io/jazz/docs/concepts/starting-runs.md): Choose how to start a Jazz run: a terminal command, saved workflow, schedule, loop, webhook, wake trigger, or background job. - [Tools](https://lvndry.github.io/jazz/docs/concepts/tools.md): Choose Jazz tools, control their permissions, review proposed changes before approval, and extend an agent with custom tools or MCP servers. - [Webhooks](https://lvndry.github.io/jazz/docs/concepts/webhooks.md): What a Jazz webhook is: one authenticated URL bound to one agent and a fixed prompt, the payload quoted as data, each delivery run once, the caller bounded. - [Workflows](https://lvndry.github.io/jazz/docs/concepts/workflows.md): A Jazz workflow is a Markdown file: the body is the prompt, the frontmatter says how it runs. Save it once, run it by name, put it on a schedule. - [Where Jazz runs](https://lvndry.github.io/jazz/docs/surfaces.md): Choose where to run Jazz: interactive terminal, scripts, CI, scheduled jobs, chat apps, webhooks, or trusted agent peers. - [Chat platforms](https://lvndry.github.io/jazz/docs/surfaces/chat.md): Run a real tool-using AI agent inside Telegram, Discord, iMessage, WhatsApp, Slack, or your own app: same agent, same tools, same approval rules as the terminal. - [Jazz in CI](https://lvndry.github.io/jazz/docs/surfaces/ci.md): Add Jazz to your CI pipeline for pull-request summaries, code review, release notes, or failed-build analysis, with a GitHub Actions example. - [CLI](https://lvndry.github.io/jazz/docs/surfaces/cli.md): Use Jazz in your terminal: choose an agent, ask it to work on files and repositories, approve actions, steer a run, and resume conversations. - [Headless](https://lvndry.github.io/jazz/docs/surfaces/headless.md): Call Jazz from your own code with jazz run: pass a dynamic prompt, get structured JSON back, control autonomy and timeouts: the contract every integration builds on. - [Scheduled runs](https://lvndry.github.io/jazz/docs/surfaces/scheduled.md): Schedule unattended Jazz runs with launchd or cron: workflow prompts on a clock, output written to a log, and delivered to your phone or chat through a notify target. - [Jazz guides](https://lvndry.github.io/jazz/docs/guides.md): Follow Jazz guides to build a personal agent, receive inbox and research briefings, review pull requests, analyze media, or connect chat apps and webhooks. - [Setting up peers](https://lvndry.github.io/jazz/docs/guides/connect-peers.md): Hands-on setup for peer links between two Jazz machines: approval tiers, the request ledger, and what each side can and cannot do to the other. - [Investigate an attack and approve a Cloudflare WAF rule](https://lvndry.github.io/jazz/docs/guides/contain-cloudflare-attack.md): Investigate an API traffic anomaly with Jazz subagents, then require human approval before applying a real Cloudflare WAF containment rule. - [Reaching your agent from a chat app](https://lvndry.github.io/jazz/docs/guides/deploy-a-chat-agent.md): Step-by-step: reach a Jazz agent from Telegram, Discord, iMessage or WhatsApp. Covers tokens, pairing, allowlists, macOS permissions, and switching providers/models at runtime. - [Build a Goggins accountability agent](https://lvndry.github.io/jazz/docs/guides/goggins-accountability-agent.md): Create a reusable Jazz accountability persona named Goggins, attach it to any model, and run it interactively, headlessly, or on a schedule. - [Start each workday with a read-only inbox briefing](https://lvndry.github.io/jazz/docs/guides/inbox-triage.md): Build a read-only Jazz morning inbox briefing that finds messages needing attention without sending, deleting, archiving, or changing email state. - [Turn incident evidence into a visual briefing with companion models](https://lvndry.github.io/jazz/docs/guides/media-companions.md): Build a mixed-model Jazz incident agent that understands screenshots, recordings, and video, then generates a stakeholder-ready visual summary. - [Build a multi-agent verification council](https://lvndry.github.io/jazz/docs/guides/multi-agent-verification.md): Build a Jazz multi-agent verification workflow with isolated reviewers, structured results, bounded budgets, and one accountable final verdict. - [Review pull requests in CI with your choice of model](https://lvndry.github.io/jazz/docs/guides/pr-review.md): Install Jazz's maintained multi-agent GitHub pull-request reviewer, then choose ChatGPT, an API provider, or a self-hosted model. - [Program status in the terminal](https://lvndry.github.io/jazz/docs/guides/program-status.md): How Jazz tells your terminal what it is doing: OSC 7501 program status (working, blocked, done, error), terminal notifications, and OSC 7 working-directory reports. - [Build a weekly multi-agent research radar](https://lvndry.github.io/jazz/docs/guides/research-digest.md): Schedule a Jazz multi-agent research radar that finds genuinely new work, challenges weak sources, and produces a cited weekly briefing without mutating external systems. - [Wake an agent from another system with a webhook](https://lvndry.github.io/jazz/docs/guides/webhook-endpoint.md): Build an authenticated HTTP door that wakes one Jazz agent with a fixed prompt, test it with curl, then point a real GitHub webhook at it. - [Configure Jazz](https://lvndry.github.io/jazz/docs/configure.md): Configure Jazz agents, models, workflows, MCP servers, web search, email, calendar, output, telemetry, limits, and project overrides. - [Agent configuration](https://lvndry.github.io/jazz/docs/configure/agents.md): Configure Jazz agents: primary and companion models, personas, context limits, tool restrictions, network access, custom tools, environment access, and memory scopes. - [Config file reference](https://lvndry.github.io/jazz/docs/configure/config-reference.md): Every key a Jazz config.json accepts, the values each one takes, which are secrets, and what each does, generated from the schema Jazz validates against. - [Connect email and calendars](https://lvndry.github.io/jazz/docs/configure/email-calendar.md): Connect Jazz agents to email and calendars through maintained CLI skills, with explicit account setup and approval boundaries for mailbox changes and events. - [Environment variables](https://lvndry.github.io/jazz/docs/configure/environment-variables.md): Every environment variable Jazz reads: data location, terminal display, network and catalogs, provider keys and server URLs, telemetry, installer, and the chat bots. - [Jazz configuration](https://lvndry.github.io/jazz/docs/configure/jazz.md): Configure Jazz runtime limits, context management, output, scheduling, notifications, telemetry, webhooks, peers, MCP trust, and project overrides. - [Language Server Protocol plugin](https://lvndry.github.io/jazz/docs/configure/lsp-plugin.md): Configure the optional generic LSP plugin for semantic code navigation and approved refactors. - [Connect MCP servers to Jazz](https://lvndry.github.io/jazz/docs/configure/mcp.md): Connect Jazz agents to local and remote MCP servers with lazy startup, OAuth, deferred tool schemas, explicit trust, and per-agent tool selection. - [Notifications](https://lvndry.github.io/jazz/docs/configure/notifications.md): Send approvals, daemon pauses, results, failures, reminders and spend alerts to the desktop, ntfy, webhooks, Telegram or Discord, with retries and chat approvals. - [Observability](https://lvndry.github.io/jazz/docs/configure/observability.md): Use Jazz's local audit trail and OTLP traces, logs, and metrics with a collector, SigNoz, Datadog, Prometheus, or Langfuse. - [Plugins](https://lvndry.github.io/jazz/docs/configure/plugins.md): Install, inspect, trust, configure, enable, update, and remove optional Jazz plugins safely. - [Configure model providers](https://lvndry.github.io/jazz/docs/configure/providers.md): Configure Jazz model providers including OpenAI, a ChatGPT subscription, Anthropic, Gemini, OpenRouter, Ollama, llama.cpp, vLLM, SGLang, Groq, and other supported APIs. - [CLI themes](https://lvndry.github.io/jazz/docs/configure/themes.md): Choose a built-in CLI theme, switch dark or light variants, or supply your own theme file. - [Configure web search](https://lvndry.github.io/jazz/docs/configure/web-search.md): Configure Brave, Exa, Keenable, Linkup, Parallel, Perplexity, Tavily, or treg for web search, set their fallback order, and understand search options and limits. - [Workflow frontmatter](https://lvndry.github.io/jazz/docs/configure/workflows.md): Configure WORKFLOW.md schedules, approval policies, model bindings, budgets, and delivery to notification targets. - [Jazz tool inventory](https://lvndry.github.io/jazz/docs/tools.md): The complete Jazz tool registry: every tool name, its honest risk tier, and exactly what stands between a tool call and an action on your machine. - [Herdr agent-state reporting](https://lvndry.github.io/jazz/docs/tools/herdr.md): Agent-state reporting when Jazz runs inside a Herdr pane: identity, live working/blocked/idle state, and session restore. - [Jazz security model](https://lvndry.github.io/jazz/docs/security.md): The Jazz authorization model: risk, disclosure, and egress as independent properties of a tool, and how a caller who is not the operator is bounded. - [Approvals](https://lvndry.github.io/jazz/docs/security/approvals.md): How Jazz decides whether a tool call runs, asks, or is refused: risk levels, approval policies, the shell-command classifier, and what happens with nobody there. - [Detach hosts](https://lvndry.github.io/jazz/docs/security/detach-hosts.md): Register and check an SSH host before detaching a Jazz conversation. - [What Jazz sends on its own](https://lvndry.github.io/jazz/docs/security/privacy.md): Every network request Jazz makes on its own, apart from your model and the tools an agent calls: what triggers it, what it sends, and how to turn it off. - [Secrets and egress](https://lvndry.github.io/jazz/docs/security/secrets-and-egress.md): Where Jazz keeps credentials, what it strips from a shell environment, and why a read-only tool that talks to the network is a separate risk from one that writes. - [Surface access](https://lvndry.github.io/jazz/docs/security/surface-access.md): Authenticate every way into a Jazz agent: chat allowlists, webhook tokens and signatures, daemon and operator tokens, peer tiers, and checks before binding a public port. - [Jazz threat model](https://lvndry.github.io/jazz/docs/security/threat-model.md): The Jazz threat model: the failures its approval, disclosure, egress, secret, daemon, webhook, peer, and isolation controls address, and their limits. - [Unattended runs](https://lvndry.github.io/jazz/docs/security/unattended-runs.md): What to decide before a Jazz agent runs with nobody watching: toolset, approval policy, budgets, identity, credentials, and where it actually runs. - [Write confinement](https://lvndry.github.io/jazz/docs/security/write-confinement.md): Write confinement fences model-run shell commands and file tools to the working directory, temp and tool caches with macOS Seatbelt or Linux bubblewrap. - [Runtime data](https://lvndry.github.io/jazz/docs/runtime-data.md): What Jazz writes to disk, which paths and JSON shapes are stable contracts you may build on, and which are internal and free to change. - [Maintaining Jazz](https://lvndry.github.io/jazz/docs/maintainers.md): Trace the Jazz agent harness through its runtime, context, tools, storage, adapters, security boundaries, tests, and evaluation suite. - [Providers & models](https://lvndry.github.io/jazz/docs/maintainers/add-a-provider.md): How Jazz stays provider-agnostic across 25 LLM providers, and what it does about the places providers genuinely differ. - [Feature Flag Service. Example Template](https://lvndry.github.io/jazz/docs/maintainers/add-a-service.md): A self-contained template showing how to design, implement, wire, and test a service in Jazz following the project architecture. - [Code map](https://lvndry.github.io/jazz/docs/maintainers/architecture.md): Find where your change goes in the Jazz codebase: directory structure, Effect/Layer conventions, adapter patterns, and testing practices. - [Context management](https://lvndry.github.io/jazz/docs/maintainers/context-lifecycle.md): Why a long Jazz run doesn't fall off its context window: compaction strategies, priority tiers, and what gets dropped first. - [Design decisions](https://lvndry.github.io/jazz/docs/maintainers/design-decisions.md): Why the Jazz harness is built the way it is: the reasoning behind each architectural choice and what every choice gives up. - [Remote handoff snapshot](https://lvndry.github.io/jazz/docs/maintainers/detach-snapshots.md): Portable snapshot contents and ownership fencing for remote conversation handoff. - [Documentation quality standard](https://lvndry.github.io/jazz/docs/maintainers/documentation.md): Write Jazz documentation that is concise, code-backed, runnable, and useful: editorial standards for concepts, references, tutorials, and feature claims. - [Fullscreen ownership](https://lvndry.github.io/jazz/docs/maintainers/fullscreen-ownership.md): Maintain fullscreen navigation, bounded transcript layout, and terminal runtime ownership without competing scroll or output writers. - [Goal lifecycle](https://lvndry.github.io/jazz/docs/maintainers/goal-lifecycle.md): The contract for inferred goals that outlive individual Jazz runs: routing, plan scope, evidence, limits, pause and recovery. - [Golden sessions](https://lvndry.github.io/jazz/docs/maintainers/golden-sessions.md): Replay recorded model streams through the real agent runner in CI, with no keys, and how to add a scenario and what to assert. - [Interface Design](https://lvndry.github.io/jazz/docs/maintainers/interface-design.md): The visual language of the Jazz terminal interface: one mark, one accent, six hues, and the rules that decide every case not covered explicitly. - [Plugin lifecycle](https://lvndry.github.io/jazz/docs/maintainers/plugin-lifecycle.md): Maintain Jazz's trusted in-process plugin lifecycle, per-run sessions, digest store, consent boundary, advisory routing, and command-risk policy hook. - [Conversation presentation](https://lvndry.github.io/jazz/docs/maintainers/presentation-documents.md): Trace Jazz's semantic conversation document, atomic presentation snapshots, stream identity, renderer projections, and versioned history storage. - [Provider contract](https://lvndry.github.io/jazz/docs/maintainers/provider-contract.md): What Jazz's provider layer relies on from the AI SDK stream: event order, usage, tool-call arguments, reasoning replay, error paths, retries, and empty finishes. - [The agent loop](https://lvndry.github.io/jazz/docs/maintainers/run-lifecycle.md): What happens between pressing enter and getting an answer: the Jazz agent loop, iteration guards, and how long unattended runs stay alive. - [Verifying terminal output](https://lvndry.github.io/jazz/docs/maintainers/terminal-rendering-tests.md): Verify Jazz's actual native ANSI output against an independent terminal grid, including stale cells, resizing, Unicode widths, and input routing. - [Terminal runtime](https://lvndry.github.io/jazz/docs/maintainers/terminal-runtime.md): Maintain Jazz's single terminal writer, demand frames, renderer capabilities, cleanup, and fallback contracts. - [Evals: measuring the harness](https://lvndry.github.io/jazz/docs/maintainers/testing-and-evals.md): How Jazz measures whether a harness change actually makes agents better: the eval suite, scoring, and regression gates for contributors. - [Tools & approval](https://lvndry.github.io/jazz/docs/maintainers/tool-lifecycle.md): How a tool call becomes an action in Jazz: risk tiers, the approval card, auto-approve policies, and what no tier can ever bypass. - [UI performance qualification](https://lvndry.github.io/jazz/docs/maintainers/ui-performance-qualification.md): Requalify the focused failing UI workloads in alternating fresh-process base/head pairs and feed the results to the performance gates. - [Computer use: from phase 1 to complete](https://lvndry.github.io/jazz/docs/design/computer-use-completeness.md): The plan for completing computer use in one PR: full action vocabulary, consent on first reach, captureAfter and wait, and the scenario suite. - [Jazz documentation](https://lvndry.github.io/jazz/docs/index.md): Install Jazz, run an AI agent from the terminal or unattended, connect tools and chat surfaces, and understand how the open-source agent harness works. - [Troubleshooting](https://lvndry.github.io/jazz/docs/troubleshooting.md): Fix common Jazz problems: rejected provider keys, a local model that never answers, empty replies, missing tools, keyring errors, a stopped daemon, and where logs are. - [Upgrading Jazz](https://lvndry.github.io/jazz/docs/upgrading.md): How Jazz versions its releases while it is 0.x, where breaking changes are announced, and what to do when an update rejects your config or changes a default.