AI agents that wake when production breaks.

agentsfleet is an open-source runtime for AI agents that wake on production events. An agent starts on an event — a pull request, an incident, a deploy — reads your code, telemetry, internal docs, and live control-plane state, finds the root cause, and opens a scenario-backed fix, on the platform's model or a key you bring. A human approves, then it ships the fix or drafts the customer reply. Every step is a replayable log.

• Human approval, by design — the agent investigates and proposes; a person approves before anything ships
• Replayable event logs — audit every action and decision
• Bring your own provider keys — no vendor lock-in on inference
• Runs locally or against production — same agent, same evidence

Agents are defined in Markdown playbooks with tools, triggers, and investigation steps. Open-source runtime, hosted control plane; bring your own model key or run on the platform's.

---

Quick start

Need access? Join the waitlist. The website's Dashboard links currently open the waitlist, and the shell installer command is hidden while unavailable.

npm install -g @agentsfleet/cli
agentsfleet login

Define an agent in Markdown, connect a webhook, and get an evidenced diagnosis and a proposed fix on your next pull request or incident. Full walkthrough at docs.agentsfleet.net/quickstart — free to try, no card, under five minutes.

---

What's in this repo

Directory	What
rustd/	Rust workspace — the agentsfleetd control plane (afd_* crates, one per plane) and the agentsfleet-runner execution daemon (afr_* crates).
ui/packages/app/	Dashboard — Next.js, Clerk auth
ui/packages/website/	Marketing site — agentsfleet.net
ui/packages/design-system/	Shared UI components
cli/	Command-line interface (CLI) — install, manage agents, tail runs
public/openapi.json	OpenAPI spec — committed and edited by hand
schema/	Postgres migrations
playbooks/	First setup, deployment, verification, and operations

---

Local development

Prerequisites: Rust 1.99.0 (pinned by rustd/rust-toolchain.toml, which rustup installs on the first cargo run inside rustd/) · Docker (Postgres, Dragonfly, QStash) · Bun ≥1.4 · Clerk dev project · 1Password CLI for secrets

git clone https://github.com/agentsfleet/agentsfleet.git
cd agentsfleet

# Populate .env before running make up. See playbooks/founding/01_bootstrap/001_playbook.md for the full bootstrap.
make up           # Postgres + Dragonfly + QStash + agentsfleetd (auto-migrates DB)

# Install in this order. The root install alone does NOT reach the workspace
# packages, and the coverage lane then fails resolving `next/headers` — a
# failure that reads like a code defect and is not one.
bun install
(cd cli && bun install && bun run build)
for pkg in ui/packages/*/; do (cd "$pkg" && bun install); done

cd ui/packages/app && bun run dev

Run provision-env-1password before local tests. The shell sources ~/.config/agentsfleet/.env, which names the UI, runner, and daemon files under that directory through AGENTSFLEET_UI_ENV_FILE, AGENTSFLEET_RUNNER_ENV_FILE, and AGENTSFLEET_RUSTD_ENV_FILE. Local UI and daemon tests read their files directly; the runner path is available to tests that need it. No worktree environment-file symlinks are needed. The app refuses to guess a backend when NEXT_PUBLIC_API_URL is unset.

Verify:

make lint-all
make test-unit-all
make test-integration-rustd   # needs Docker: live Postgres + Dragonfly

---

CLI

agentsfleet defaults to production. Point it at your local stack with the --api flag, or persist it via the environment:

agentsfleet --api http://localhost:3000 <command>
export AGENTSFLEET_API_URL=http://localhost:3000   # or set it once

---

Contributing

git config core.hooksPath .githooks

Project facts live in AGENTS.md. The operating model and the deterministic gate scripts are committed alongside them — .orly/AGENTS.md, .orly/dispatch/, and .orly/audits/ — materialised by orly and recorded in .orly/orly.json. A clone carries its own rules; nothing resolves out of a developer's home directory.

Install orly at the version this repository records, so a local run and Continuous Integration grade against the same rules:

npm install -g "@agentsfleet/orly@$(jq -r .orly_version .orly/orly.json)"

orly doctor                     # report drift between the lock and the tree
orly update --no-hooks          # re-materialise at the installed version
orly gate work                  # what the pre-commit hook runs

--no-hooks is not optional here: this repository owns .githooks/pre-commit and .githooks/pre-push, and orly declines to overwrite hooks it did not write.

make harness-verify runs shared gates from .orly/audits/ and repository-owned checks from audits/.

Read about architecture, start with the operator playbooks, or jump into local development.

---

MIT — Copyright (c) 2026 agentsfleet.