uv tool install langctl
langctl new my-agent
cd my-agent          # add your API key to .env
langctl dev

langctl dev starts the LangGraph Agent Server and your Next.js app, waits for the agent's health endpoint before booting the UI, proxies the agent behind the frontend's own origin, and tears both down cleanly on Ctrl-C. Like next dev, but the backend is an agent.

Long-term memory is on by default and needs nothing running.

Requirements

Two more are installed into your project by langctl new, not onto your machine: langgraph-cli, which runs the Agent Server, and your model provider's LangChain package. If you scaffold with --no-install, run uv sync --extra dev before langctl dev.

For deploying to a remote host you also need ssh and rsync locally, and Docker with the compose plugin on the target machine. Nothing else — no registry, no Kubernetes, no cloud account.

Everything is checked for you:

langctl doctor        # versions, ports, API keys, Postgres, langgraph validate

Platforms. Linux, macOS and Windows are all supported. On Windows, use PowerShell or Windows Terminal rather than the legacy console.

Install

uv tool install langctl     # recommended — isolated environment, on PATH
pipx install langctl        # same idea
pip install langctl         # works, but shares your environment

Upgrade with uv tool upgrade langctl. If that reports no change when you expect one, it is uv's cached index: uv tool install --force --reinstall langctl.

Commands

langctl new

langctl new my-agent                    # interactive
langctl new my-agent --yes              # all defaults
langctl new api-bot --yes --no-frontend
langctl new my-agent --yes --memory-backend postgres --semantic-search

--runtime · --model-provider · --model · --frontend/--no-frontend · --ui · --memory/--no-memory · --memory-backend · --semantic-search · --embeddings · --embedding-model · --yes · --no-install · --no-git

langctl dev

--backend-only · --frontend-only · --port · --backend-port · --no-open · --docker (runs langgraph up, port 8123) · --tunnel · --auto-port/--strict-port

langctl add

langctl add memory --backend postgres
langctl add frontend
langctl add tool "lookup order"

add regenerates the files langctl produced and skips the ones you edited, listing each. A file that still matches what the template last wrote is safe to update; anything else is yours. See Adding to an existing project.

langctl sync

Regenerates langgraph.json and pyproject.toml dependencies from agent.yaml. --check exits non-zero on drift, for CI. --force overwrites owned keys you edited.

langctl doctor

Toolchain, ports, API keys, Postgres reachability, and langgraph validate. Prints a fix for each failure.

Why the proxy

The browser only ever talks to /api/... on the frontend's origin:

localhost:3000                      127.0.0.1:2024
┌────────────────────────┐          ┌──────────────────┐
│ Next.js                │          │ langgraph dev    │
│  /            chat UI  │          │  /threads /runs  │
│  /api/*       ─proxy───┼─────────▶│  /assistants /ok │
└────────────────────────┘          └──────────────────┘
        same origin ⇒ no CORS, ever

Three things fall out of this:

• CORS never applies. There is no cross-origin request to preflight.
• The API key stays on the server. The proxy runs in a route handler and attaches x-api-key there. Nothing secret reaches the browser.
• Dev, tunnel, and production differ by one variable. LANGGRAPH_API_URL points at localhost, then a container on the deploy network. The frontend source never changes.

Chat UI

web/ is LangChain's agent-chat-ui, vendored unmodified at a pinned commit (MIT). You get the real thing: thread history, an agent inbox for human-in-the-loop approvals, artifacts, markdown with syntax highlighting, file and image attachments, and generative UI.

The setup screen never appears. Upstream shows a form asking for a deployment URL and assistant ID when either is missing:

if (!finalApiUrl || !finalAssistantId) {  // ← the setup screen

langctl pre-fills both, so that branch is never reached. The screen is bypassed, not deleted — leaving the source byte-identical means re-syncing with upstream is a plain diff rather than a merge. web/VENDORED.md records the exact commit.

Models

25 providers, plus anything init_chat_model supports:

langctl new my-agent --model-provider openai
langctl new my-agent --model-provider openrouter --model z-ai/glm-5.2
langctl new my-agent --model-provider ollama --model qwen3   # name is required

A custom endpoint — LM Studio, vLLM, a LiteLLM proxy, any OpenAI-compatible gateway:

langctl new my-agent --model-provider openai --model my-model \
  --model-base-url http://localhost:1234/v1

A provider langctl does not know — just say what supplies it:

langctl new my-agent --model-provider mycloud --model m1 \
  --model-package langchain-mycloud

The provider list is open rather than fixed, so a new LangChain integration is usable the day it ships instead of waiting on a langctl release. Providers that need no key (Ollama, ambient cloud credentials) do not demand one.

Some providers have no safe default model. A hosted provider does — everyone with an OpenAI key can reach gpt-5.5. Ollama, LiteLLM, HuggingFace, a custom base_url, and any provider langctl does not know all serve whatever your machine or gateway has, so no name is invented for them. MODEL_NAME comes from .env, and the generated config.py fails at startup with a clear message rather than on the first message with a 404 from a server it never saw.

Everything is also editable in agent.yaml, including per-model options:

model:
  provider: openai
  name: my-model
  base_url: http://localhost:1234/v1
  api_key_env_override: MY_GATEWAY_KEY
  options: {temperature: 0.2}

When base_url or options are set the model is constructed rather than passed as a provider:model string, because a string carries neither.

Memory

Long-term memory is on by default and needs nothing running. Verified against a real restart: without an explicit store, langgraph dev keeps memories in process and loses them all on exit.

Semantic search is off by default — it needs an embeddings vendor and costs per item stored. Three ways to produce vectors:

Two memories, opposite defaults, for a reason: overriding the store is a strict gain (it is otherwise lost on restart), while overriding the checkpointer loses adelete_for_runs, so threads stay server-managed unless you ask.

Middleware

Every new project ships with cost and reliability guards on — an agent with no call limit can loop until it exhausts your budget:

MIDDLEWARE = [
    # limits
    ModelCallLimitMiddleware(run_limit=20),
    ToolCallLimitMiddleware(run_limit=30),
    # reliability
    ToolRetryMiddleware(max_retries=2),
]

langctl add middleware --list              # registry, and what is on
langctl add middleware summarization
langctl add middleware --custom rate_limit # your own class

Order is semantic, not alphabetical. langctl emits a fixed sequence — guardrails → context → limits → reliability → human-in-the-loop → capability → custom — because PII redaction placed after summarization means raw content already reached the summarizing model. Custom middleware always runs last, so it sees a fully prepared request.

Settings live in agent.yaml and are regenerated into middleware/__init__.py by langctl sync.

Custom middleware gets one module per class, mirroring tools/:

middleware/
├── __init__.py          MIDDLEWARE list — built-ins, then yours
└── custom/
    ├── __init__.py      registry, regenerated by `langctl sync`
    ├── rate_limit.py
    └── audit_log.py

Each is scaffolded as a bare class — no hooks. Which of the six hooks a middleware needs is the design decision, so the docstring lists them all with real signatures and you add only what you use.

Adding to an existing project

Upgrading langctl does not touch projects you already created — templates are copied at creation, not linked. Use langctl add.

Projects from 0.1.x/0.2.x used single modules (tools.py, prompts.py) where 0.3+ uses packages (tools/, prompts/). After add, both exist and Python imports the package, so the old file is silently ignored. Move anything you wrote into the package and delete the module — editing it will have no effect.

Sharing and deploying

langctl share — shipped

Puts your locally running app on a public URL. One tunnel covers the whole thing, because the agent already sits behind the frontend's proxy — the agent port is never exposed and no API key leaves your machine.

langctl share                      # cloudflared if present, else ngrok
langctl share --provider ngrok
langctl share --backend-only       # expose the agent API instead (warns first)

cloudflared is preferred because its quick tunnels need no account. Nothing is hosted: close the laptop and the URL dies. Good for demos, client previews, webhook testing, and trying the app on a phone.

The URL is public and unauthenticated. Anyone with it can talk to your agent and spend your API credits.

langctl deploy

Both halves go to one host, in one operation, behind one URL:

langctl deploy                                    # this machine
langctl deploy --host user@1.2.3.4                # a server you own
langctl deploy --host user@1.2.3.4 --domain x.io  # the same, with HTTPS

┌─ one host ────────────────────────────────────┐
│  web       Next.js          ← the only door   │
│   └ /api → agent                              │
│  agent     Agent Server       internal only   │
│  postgres · redis             internal        │
└───────────────────────────────────────────────┘

The frontend is never told an address. It reaches the agent at http://agent:8000 — a service name on the private network. There is nothing to paste into a config and nothing to update later, so redeploying the agent cannot break the UI. This is the failure the command exists to prevent; deploying the two halves separately means re-wiring them every time.

Only the frontend publishes a port. The Agent Server has no route in from outside, so the LangSmith key stays in the server-side proxy exactly as it does in development.

First deploy, in three steps:

langctl deploy --host user@1.2.3.4        # 1. stops, writes .env.deploy
#                                            2. fill in the three secrets
scp .env.deploy user@1.2.3.4:~/my-agent/  # 3. place them on the host, once
langctl deploy --host user@1.2.3.4        #    run again — done

Secrets are checked before anything is built, so a missing key costs you a second rather than ten minutes into an image build. .env.deploy is never uploaded by a deploy and never baked into an image.

Redeploy with the same command: rsync sends only what changed and Docker reuses layers.

With --domain, Caddy joins the stack and obtains and renews a Let's Encrypt certificate on its own. It deliberately does not compress text/event-stream — buffering the token stream to compress it is what makes an agent appear to hang and then answer everything at once.

The startup order is enforced by health checks: Postgres and Redis become healthy → the agent starts and answers /ok → only then does the UI start. A deploy that fails exits non-zero instead of handing back a URL that does not load.

No licence needed, and LangSmith is optional. The default stack runs the in-memory agent server, which has no licence check — verified starting with no LangSmith key at all, in 162 MB. Postgres and Redis are in the stack either way, so your graph gets a real database rather than a file, and switching to --licensed later needs no new infrastructure. Tracing is opt-in: set LANGSMITH_TRACING=true and a key in .env.deploy only if you want traces.

Postgres is only used by a project whose long-term memory backend is postgres — the backend is baked into memory/store.py at scaffold time. deploy says so plainly when it is running but idle, and langctl add memory --backend postgres switches it.

For queue autoscaling, multi-replica scaling and graceful run draining, add --licensed to run LangChain's production Agent Server instead. That one validates a LANGGRAPH_CLOUD_LICENSE_KEY at start-up and refuses to boot without it — the check lives inside LangChain's image, so it cannot be disabled — and it brings Postgres and Redis with it.

Configuration

agent.yaml is the single source of truth; langgraph.json and the dependency list are generated from it. sync merges rather than overwrites, so hand-added keys and packages survive, and drift is reported instead of silently clobbered.

Development

uv venv && . .venv/bin/activate
uv pip install -e ".[dev]"
pytest
ruff check src tests

The suite spawns real child processes rather than mocking Popen: the failures that matter — orphaned grandchildren, ports left held, signals that never arrive — do not exist at the mock level.

Heavier checks need a scaffolded project:

export LANGCTL_E2E_PROJECT=/path/to/project
tests/e2e/dev_runtime.sh          # health gate, proxy, thread creation, teardown
tests/e2e/sse_streaming.sh        # asserts SSE arrives incrementally, not buffered
tests/e2e/memory_persistence.sh   # writes a memory, kills the process group, restarts
tests/e2e/share_tunnel.sh         # live tunnel: public URL, proxy, clean teardown

sse_streaming.sh measures arrival times: a buffering proxy passes every status-code assertion and still ruins the product. memory_persistence.sh kills by process group and requires the port to be released — an earlier version killed only the parent, left the uvicorn child serving, and reported a fake pass.

License

Apache-2.0. Generated projects carry no license obligation to this tool.