ExtendRoom 12 of 13

Every tool declares its risk before it runs.

The catalog covers 63 skills across developer workflows, AI agents, office productivity, media processing, and more. Install more from the Marketplace or let the agent author its own.

Each of the 132 tools cataloged here carries a risk level in its manifest, and the approval gate reads it on every call.

A recreation of the File System skill under an agent's tools in the Sophon Dashboard: 9 tools, each with its declared risk. The agent calls fs.read (None), fs.copy (Low), fs.write (Medium), fs.delete (High). At the default Medium threshold, None and Low run, and Medium and High stop and ask you first.

04 / 04 · fs.delete · High · asks you

Recreation of an agent's tools in the Dashboard. Verdicts at the default Medium threshold.

AWatch a skill go from package to run

One skill, from package to finished run.

From the author's manifest to a destroyed container: what happens to a skill before and while it runs.

A skill folder with manifest.json, SKILL.md and main.py. The manifest declares the tool my_tool with risk level low and a network allowlist of api.github.com.

Step 1 of 5: A skill is a package

01 / 05 · A skill is a package

  1. A manifest.json for the machine: tool names, parameters as JSON Schema, risk levels, required credentials, sandbox config. A SKILL.md for the agent: what the skill does and when to use it. Then the code, in Python or C#.

    The risk level is written down before the tool ever runs.

BMechanism

Five risk levels, written into the manifest.

The level lives in the registry, not the prompt. The gate compares it with your threshold, Medium unless you change it, at Settings › Security › Approvals.
Risk: None

Read-only, deterministic, cannot affect anything.

Runs.

44 cataloged tools · memory.search, fs.read

Risk: Low

Reads with side effects, or minor writes.

Runs.

38 cataloged tools · web.search, fs.copy

Risk: Medium

Writes data, sends messages you could undo.

Asks you at the default threshold.

35 cataloged tools · fs.write, message.send

Risk: High

External, costly, or hard to reverse.

Asks you at the default threshold.

14 cataloged tools · fs.delete, email.send

Risk: Critical

Destructive or privileged.

Always asks, whatever the threshold or any override says.

1 cataloged tool · agent.delete

Overrides, per-call assessment and the Critical floor are applied in that order. How the gate resolves a call

CCatalog

Every skill, every tool, every level.

Filter by category or search for a tool name. Open a skill to see its tools and what each one declares. Jira and Confluence install from the Marketplace rather than shipping in the base install.

Category

63 skills · 132 tools

DWhere skills come from

Bundled, installed, or written by the agent.

Self-authored skills run in the same sandbox as Marketplace skills. Every skill does.
Bundled
Ship with every Sophon install. Enable them per agent in Agents › Tools; newly installed skills start disabled on agents that already exist.
Marketplace

Community-built skills and plugins for Notion, Todoist, Stripe, Salesforce, HubSpot, Spotify, Home Assistant, and more. Every new version waits for a moderator before it is visible. Install from the Dashboard (Skills › Marketplace) or with sophon skills install notion; an install the agent starts is always gated behind your approval.

Browse the Marketplace Marketplace docs

Self-authored
Ask for a skill in plain language and skill.author (High risk) writes the code, the manifest with its risk level, and the SKILL.md, tests it in a sandbox with sample parameters, and asks for your approval before installing it.

ESkills or plugins

Skills call tools. Plugins extend the core.

Since 2.2, the Marketplace also carries agent packages: personas and setup you can review, install and customize. Skills supply their tools. Explore agent packages.

A skill gives the agent a new tool, sandboxed. A plugin gives Sophon itself a new channel type or document format. Plugins run out of process with full trust, are admin-gated, and are distributed through the same Marketplace.

Limits

What a risk level doesn't promise.

  1. The declared level is a default, not a floor. An admin's per-tool override can raise it or lower it, except on Critical tools, which always ask.

    Risk levels
  2. A self-authored skill picks its own risk level. Review the generated manifest before you approve it: that level decides how often you are asked later.

    Self-authoring
  3. The vault keeps a credential out of the model's prompt, not out of the skill. A sandboxed skill holds the real credential while it runs, and its network allowlist is what bounds where it can go. A broker that makes the call on its behalf is roadmap.

    Credential vault

Write the tool Sophon doesn't have.

A manifest, a SKILL.md and a script. The sandbox, the risk gate and the per-agent switch come with it.