TypeScript
v1.4 · current
First-party. Used by 70% of customers.
$ npm install @lattice/sdk
Documentation · v1.4
Documentation engineers actually finish. Six sections. None of them filler.
The doc was written by the team who built the runtime. We rewrite it whenever a customer tells us a section was confusing. The full text is open source on GitHub; pull requests welcome.
01 · Quickstart5 min
From zero to a running agent in 14 seconds.
Install the SDK, drop in your API key, define an agent, run it. The whole quickstart fits on one page; no boilerplate, no scaffold generator.
Open →
02 · Concepts12 min
Runs, steps, schedules, and traces.
What Lattice models, why each primitive exists, and how they compose. Read this once and the rest of the API will make sense.
Open →
03 · Guides8–20 min
Production-tested patterns.
Long-running RAG, tool-augmented chat, multi-agent handoffs, batch enrichment, scheduled cron agents, human-in-the-loop approvals. Each guide is a real customer pattern, distilled.
Open →
04 · API referencebrowse
Every method, typed.
Generated from source. Full TypeScript types, Python types, request/response shapes, error codes, and rate-limit behavior.
Open →
05 · SDKs—
TypeScript, Python, Go.
First-party SDKs for the three runtimes our customers actually use. REST + Webhooks if you want to roll your own client.
Open →
06 · Examples—
Reference repos you can clone.
Twelve example apps, one per agent pattern. Each one builds, deploys, and runs against Lattice cloud or self-host with no edits.
Open →
Four steps · 14 seconds ↓
From zero to a running agent in four steps.
1
Install the SDK
$ npm install @lattice/sdk
2
Set your API key
$ export LATTICE_API_KEY="lt_..."
3
Define an agent in agent.ts
$ lattice.agent.define({ name: "summarize", steps: [...] })
4
Run it
$ lattice.runs.create({ agent: "summarize", input: {...} })
The vocabulary, in order ↓
Five primitives, defined once.
- Agent
- A named, versioned definition of work — the steps an LLM call should perform. Defined once, invoked many times.
- Run
- A single invocation of an agent against a specific input. Has a unique ID, a status, a trace tree, and durable persistence.
- Step
- One unit of durable work inside a run. An LLM call, a tool invocation, or a persisted intermediate. The unit Lattice bills on.
- Schedule
- A queue + concurrency policy. Controls how many runs of an agent can be in flight, with what rate limits, against which providers.
- Trace
- The complete record of a run — every step, every retry, every prompt, every tool response. Native OpenTelemetry, replayable.
Six guides. One per real customer pattern.
Updated monthly · last 2026-04-22
RAG
Long-running retrieval-augmented generation
14 min →
TOOLS
Tool-augmented chat with rate-limit handling
11 min →
MULTI-AGENT
Sub-agent handoffs and result merging
18 min →
BATCH
Nightly batch enrichment of CRM records
9 min →
SCHEDULED
Cron-like scheduled agents
6 min →
HITL
Human-in-the-loop approval gates
13 min →
Pick a runtime ↓
Python
v1.3 · 1 minor behind
First-party. Pinned to TS SDK for parity.
$ pip install lattice-sdk
Go
v0.9 · pre-1.0
First-party. Smaller surface, focused on runners.
$ go get github.com/lattice-run/go
REST + Webhooks
stable
Roll your own client in any language.
$ no install
API reference
Generated from source.
Every method documented with the actual TypeScript and Python types. Searchable by symbol. Includes error code listings, retry behavior, and rate-limit semantics for every endpoint.
Browse API ref →Examples
Twelve clone-ready repos.
One reference app per agent pattern. Each one builds clean, deploys to Vercel/Modal/Render, runs against Lattice cloud, and is also runnable against the OSS runtime with no edits.
Browse examples on GitHub →The docs are open source. PRs welcome.
Spotted a typo or a confusing sentence? Open a PR on github.com/lattice-run/docs. We merge and ship the same day if the change is right.
$git clone github.com/lattice-run/docs