Docs
Open the console
Changelog

Changelog

What's new in LagrangeCloud and the Lagrange plugin for Codex and Claude Code, newest first. The plugin always downloads the platform's current CLI, so CLI changes reach you without updating the plugin.

Week of October 19, 2026

  • Attach files to runs. Add files in the New run dialog or a follow-up, by button or drag and drop, or with lgr run --attach PATH. The agent finds them under uploads/ in its working directory. See Call your agent.
  • New run is the agent page's main action. The Playground moved to the agent's API tab, next to the caller keys.
  • Markdown results. Run results and Playground replies render as formatted Markdown.
  • See what changed between versions. Each version names what changed since the one before it, and See changes shows the settings, line diffs of the instructions and skills, and changed files. See Versions.
  • Coupon codes. Enter a code when you sign in, or under Settings, to add credit to your account. See Billing.
  • Feedback from the console. The Feedback form sends your message to the team directly.
  • Python callback verifiers. Settings → Callbacks shows how to check a callback's signature in Python as well as Node.js.
  • Running out of credit. Out-of-credit errors link to the page where an owner or admin can add credit, and owners and admins get an email when credit drops below $1 and when it runs out.
  • Fixed: an agent's own scripts can use the OpenAI or Anthropic key you uploaded as a secret, while the agent itself runs on a platform model.
  • Fixed: python works in the sandbox, not only python3.
  • Fixed: the CLI no longer hangs on a stuck request, and prints a one-line error instead of a stack trace.

Week of October 12, 2026

  • Your agent's sign-ins come with it. Deploying from Codex or Claude Code copies MCP sign-ins (Linear, Notion, Sentry and others) and CLI logins, so the first cloud run works. When a sign-in expires, the agent's schedules pause, you get an email, and one Reconnect click lists and renews every expired sign-in. The deploy also tells you what it can't bring, such as local services.
  • More reliable deploys. Multi-line keys, CLAUDE.md imports, symlinked skills and script permissions now carry over, and an MCP server that fails to start shows up in the run's trace.
  • Sandbox tools. The sandbox has uv, uvx and Python 3.12, plus gh, sqlite3, ffmpeg, pdftotext and zip. Codex agents can write files in the cloud.
  • MCP headers. Remote MCP servers can send custom headers, such as X-API-Key, with values taken from the agent's secrets.
  • Scheduled and webhook runs email their results to whoever created the schedule or webhook.
  • Guided demo. Run the demo agent from the console and watch Lagrange find wasted steps, replay the fix and promote a cheaper, faster version.
  • Improvements. An agent can propose a rewrite of its own workflow from its traces. Each candidate is replayed on recent inputs and judged on what each run did, then rolled out as a canary that rolls back on its own if it does worse. See Improvements.
  • Adjust an agent by describing the change. Click Adjust, describe what you want, and review the change file by file before saving it.
  • Platform models. Codex and Claude Code agents run on platform models billed from your credit; any platform model works on either harness. Settings → Models lists them with prices.
  • Buy credit with a card under Settings, in US dollars.
  • Follow up on any run to continue where it left off.
  • Workflows. Run an agent as a series of steps, each on its own model.
  • Docs at /docs, with a quickstart and a reference for the CLI and API.
  • Plugin: the Lagrange harness is marked as coming soon, and deploys run on the harness you use locally.

Week of October 5, 2026

The first release of LagrangeCloud and the Lagrange plugin.

  • Deploy a local agent. The plugin deploys a Codex or Claude Code agent with its instructions, skills, files, MCP servers and keys, and warns about anything that would break its calls in the cloud.
  • Every agent gets an API. Call it over HTTP with streaming, through OpenAI Chat Completions, from the browser with session tokens, or from an embeddable chat widget. Caller keys and pausing are managed in the console or CLI.
  • Try before going live. Deploy a version as a preview, try it in the Playground or with lgr chat, then make it live.
  • Schedules and webhooks start runs on their own, and callbacks post signed results to your URL. See Schedules and webhooks.
  • Isolated sandboxes, with an allowlist for the hosts an agent may call.
  • Traces of every step of every run, in the console.
  • Instances and fleets for always-on agents. See Instances and fleets.
  • Bring your own harness through the harness protocol.
  • Draft an agent from a description in the console.
  • Credit and usage. Runs and instance time are charged to your organization's credit, shown under Settings.