AmbitGet startedFAQReferenceGlossaryAll docsGitHubLive demo

Why Ambit

This argument runs ahead of the software. The README describes what runs today; the roadmap identifies what remains unbuilt.

I built Ambit because neither I nor the agents could keep my whole stack in our heads: models, MCP servers, skills, machines, hosted services, credentials and scheduled jobs.

Every component had configuration. Nothing described what the whole arrangement could do.

I wanted to know which addition would open useful work, which providers actually worked, and whether repeated failures shared one cause. Ambit reads those configs into one graph of what the arrangement can do. A person's ambit is what they, their agents and their machines can do together. Exposing that graph over MCP lets agents ask the same questions, without reconstructing the setup every session.

Composition is the part that got interesting

Shell access, a network route, Docker, monitoring, a scheduler and scoped authority might support recovering a service while I am asleep. No individual config entry establishes that outcome.

Ambit keeps structural reach, check results and permission separate:

installed ≠ working ≠ authorized

A configured tool can fail. A working tool can lack permission. Even passing component checks leave the complete recovery workflow untested.

What an agent could notice

When the same missing piece stops an agent more than once, the agent should help find the cause and propose a change later sessions can reuse. A known missing dependency is grounds to add it. A block that recurs with no known cause needs investigating first.

A recurring block leads to investigation and a proposed change. A person may approve or decline it. An approved change is applied and checked; a failed check returns to investigation, while a passing result is retained as evidence.

The useful result is a capability future sessions can reuse, with its dependencies, authority and evidence recorded. If a later check fails, the graph excludes it from current plans.

Accounting, and then a history

Ambit accounts for capacity to act. It records what is configured, what checks have passed, who must participate and what may run without asking.

History makes changes inspectable: adding credentials or a scheduler can complete a modeled capability without adding a dedicated provider for it. Removing one shared dependency can contract several capabilities at once.

Each time Ambit reads the configs it saves a snapshot of every capability's state, kind and verification health, shown with today's names and edges. Reconstructing every route and permission as it stood during an incident needs historical records Ambit does not keep yet.

Effective agency as a governed object

An unchanged model can participate in more consequential work when it gains tools, persistence, credentials or other participants. Those changes deserve review alongside model changes.

The design norm is:

No increase in effective capability without a corresponding increase in legibility, verification, and governability.

An agent can propose expanding its environment. A person decides whether to approve the change; the executor checks authority, and evidence records what worked. Enforcement currently covers apply, routed control-plane execution and Claude Code with the gate hook installed. Other runtimes choose whether to consult it.

The next step is evidence about complete outcomes: what this setup delivered, on which target, within what limits and with how much human participation. The affordance frontier develops that direction across digital, physical and institutional settings.

See what your own agents can reach

Ambit reads the MCP configs of Claude Code, Cursor, OpenCode and thirteen more agents into one local map: every server, which ones work, what one more step would open, and what breaks if a piece goes.

Open the live demoa sample setup in your browser, nothing to install

On your machine: npx ambit-cli, with Node 22.18 or newer on macOS or Linux.

MIT licensed · one local SQLite file, no telemetry · on npm and in the official MCP registry