Developer downloads

Build, ship and operate agents

im-cli takes a project from your editor to a running execution. im-sdk is what the agent imports to call the models and services it has been granted. Both are published here.

$ curl -fsSL https://downloads.mgt-platform.com/install.sh | sh

Command-line tool

im-cli

Scaffold a project, link it to an agent, push a version and run it, streaming the console output in your terminal.

Install im-cli →

Python SDK

im-sdk

The library your agent imports: call the models and services it has been granted, read its typed inputs, and trace a run end to end.

Add it to your project →

Documentation

Guides and reference for the CLI, the SDK and the agent workflow.

Not published yet.

Working from a coding agent

Four skills, one per thing you do. Drop one into ~/.claude/skills/ and ask your agent in plain language. llms.txt describes this host for a machine.

Gets a machine ready, checking each step rather than assuming it: im-cli on your PATH, the SDK index resolving, and the Python and uv an agent project needs. No credential, and it stops short of signing in — that needs your organisation’s Portal URL.

curl --create-dirs -fsSL https://downloads.mgt-platform.com/skills/im-setup/SKILL.md -o ~/.claude/skills/im-setup/SKILL.md

Read the skill →

From a plain-language prompt to a running agent: interviews you, reads the grants your agent actually has, writes main.py and the .im/io.json contract, then pushes and runs it.

curl -fsSL https://downloads.mgt-platform.com/skills/im-create-agent.tar.gz | tar -xz -C ~/.claude/skills/

Brings an existing project up to the current im-sdk: migrates the breaking changes in main.py, re-locks against the public index, and stops before pushing.

curl -fsSL https://downloads.mgt-platform.com/skills/im-update-agent.tar.gz | tar -xz -C ~/.claude/skills/

Agents are Python programs that run on your organisation's Portal. You write them locally with im-cli, which links a project to an agent, pushes it and runs it. Inside the agent, im-sdk calls the models and services an admin has granted it. You will need your organisation's Portal URL. Ask whoever administers it.

The shape of it

From an empty folder to a running agent

Five commands. The full guide is in the documentation.

my-agent / im-cli
$ im-cli login --api-url https://portal.example.com/api
✓ signed in, 1 active login

$ im-cli new python-blank my-agent
✓ created ./my-agent

$ im-cli link --agent 8f2c41e0
✓ linked, wrote .im/config.json

$ im-cli push
✓ uploaded as a draft version

$ im-cli run --input message=hello
  streaming console output…
✓ execution succeeded
  1. Sign in to your Portal

    Opens the browser to confirm a code, then stores a token locally. One active login at a time; switch by logging in again.

  2. Scaffold a project

    python-blank is a zero-dependency starter; python-llm calls a model through the SDK.

  3. Link it to an agent

    An admin creates the agent and grants the models and services it may use. im-cli agent models and agent services show what it is allowed to call.

  4. Push a version

    Zips the project and uploads it as a draft you can iterate on. --submit sends it straight to review.

  5. Run it

    Streams the run's console output and exits non-zero if it fails, so it works in a script as well as at a prompt.

Linking needs an agent that already exists. Agents are created in the Portal by an administrator, who also grants the models and services each one may use, so the first three steps need a Portal account, not just the CLI.

Install

Get im-cli

macOS, Linux and Windows. The installer resolves the newest release, verifies its signature against a key pinned in the script and its SHA-256 against the signed manifest, and puts the binary on your PATH.

curl -fsSL https://downloads.mgt-platform.com/install.sh | sh

Installs to ~/.local/bin by default. Pass options after -s --, for example | sh -s -- --bin-dir /usr/local/bin or --version 1.1.0. Windows has no shell for this; use the download below.

Installing from an agent? Follow skills/im-setup/SKILL.md, which adds the checks worth running afterwards.

No release has been published yet.

Python

The package index

Point your project at it once and uv or pip resolve im-sdk from it like any other index.

im-cli new already writes both blocks into the projects it scaffolds. You need them by hand when you are assembling a project yourself.

Point at the index

Two blocks in pyproject.toml: where the index is, and what may come from it.

[[tool.uv.index]]
name = "in-manas"
url = "https://downloads.mgt-platform.com/simple/"
explicit = true

[tool.uv.sources]
im-sdk = { index = "in-manas" }

Depend on it

Pin the version your agent is built against, then uv lock.

[project]
dependencies = [
    "im-sdk",
]

What an agent looks like

The whole runtime contract in one decorator: @im_agent builds and authorizes the client, hands you the run's input, opens a trace, and prints what you return as the execution's output. This one answers the --input message=hello from the terminal above.

# main.py
import asyncio

from im_sdk import im_agent


@im_agent("greeter")
async def main(core, data) -> str:
    reply = await core.llm.call(
        messages=[{"role": "user", "content": data["message"]}],
        alias="gpt-mini",          # an alias granted to this agent
    )
    return reply.text


if __name__ == "__main__":
    asyncio.run(main())

explicit = true is not decoration. Without it, uv searches this index for every dependency in your project, so a public package that happens to share a name with something published here resolves from the wrong place. The [tool.uv.sources] binding is the other half: it says im-sdk, and only im-sdk, comes from here.