Skills

A skill is a packaged set of tools that agents can call. Bundled skills ship with Sophon. Marketplace skills are downloaded and installed. Self-authored skills are generated by the agent from a natural-language description. Every skill runs in a Docker + gVisor sandbox with resource limits, no host filesystem access by default, and explicit network rules.

Anatomy of a skill

A skill is a folder with three files:

• manifest.json — machine-readable: tool names, parameters (JSON Schema), risk levels, required credentials, sandbox config
• SKILL.md — natural-language guide for the agent: what the skill does, when to use it, examples
• main.py / main.cs — the code that implements each tool

skills/my-skill/
├── manifest.json
├── SKILL.md
└── main.py         # or main.cs for C# skills

See Skill Manifest for the full schema.

Bundled skills

These ship with every Sophon install. Enable them per agent in Agents → Tools.

Developer tools

Note: Jira and Confluence install from the Marketplace rather than shipping in the base install (sophon skills install jira / confluence). Their source is maintained in the open-source sophon-skills repository, alongside google-gmail.

Productivity

Media

Extraction

Browse the full bundled list and per-tool details on the Skills page.

Core built-in tools

Some tools aren't in a skill — they're part of Sophon itself and always available:

• datetime.now — current time, always-available
• memory.write, memory.search, memory.list, memory.forget, memory.graph, memory.diagram — see Memory
• workflow.* — see Workflows
• document.* — including document.retrieve (verbatim top-k passages with source ids) and document.ask (grounded answers with citations); see Documents
• connection.configure, connection.list, connection.status — see Connections
• discussion.run — see Discussion
• insights.query — see Insights
• remote.control, node.present_canvas, desktop.activity — see Sophon Node
• tool.create_chain — compose existing tools into a reusable chain
• session.spawn — sub-agent spawning for multi-agent routing

Installing marketplace skills

From the CLI:

sophon skills install notion
sophon skills install todoist

From the Dashboard: Skills → Marketplace. The catalog is served by the Sophon Marketplace; installation downloads the .sophon-skill archive, verifies its SHA-256 checksum (installs fail closed if the checksum is missing or mismatched), unpacks it to ~/.sophon/skills/installed/<name>/, and registers its tools. See Installing Packages for the full pipeline, updates, and air-gapped configuration.

Uninstall:

sophon skills uninstall notion

Enabling skills per agent

Every skill can be toggled per agent. By default, newly installed skills are disabled on existing agents — you have to enable them explicitly. This prevents a skill from surprising an agent that wasn't configured for it.

From the Dashboard: Agents → <agent> → Tools.

From the CLI:

sophon agents tools <agent-id> --add github.repos github.create_issue
sophon agents tools <agent-id> --remove code.lint

Risk levels and approval

Each tool declares its risk level in manifest.json:

The gate routes to the Dashboard, a mobile push notification, or the originating channel. See Approval Gates.

Self-authoring

Agents can generate their own skills from a natural-language description:

"Create a skill that fetches the weather for a given city and returns temperature + forecast."

The skill.author tool:

1. Generates Python code (main.py) implementing the tool.
2. Generates a manifest.json with parameter schema and risk level.
3. Generates a SKILL.md describing when to use it.
4. Tests the skill in a sandbox with sample parameters.
5. Requests your approval.
6. On approval, installs to ~/.sophon/skills/self-authored/<name>/.

Self-authored skills run in the same sandbox as marketplace skills. Review the generated manifest before approving — the risk level it picked determines how often you'll get prompted later.

Sandbox

Every skill runs in a Docker container with:

• gVisor for kernel isolation
• Resource limits declared in the manifest (CPU, memory, timeout)
• No host filesystem by default
• Network disabled by default — skills must declare "network": true and an allowlist
• Declared credentials — when a skill's manifest declares a credential, the stored credential is delivered into its sandbox for that one run, then the container is destroyed. It stays out of the model's prompt, but the skill code does hold it, so the network allowlist above is what bounds where it can go. ("access": "proxy" is a deprecated alias of "injected" and behaves the same; a brokered proxy is roadmap. See Credential Vault.)

The sandbox image includes Python 3.13 (with async support), .NET 10 runtime, ffmpeg, Tesseract, and a Playwright-capable Chromium.

Agent and workflow templates

Sophon ships with pre-configured agent templates (each comes with a curated skill set + SOUL.md):

sophon agents create --template devops-engineer my-devops

Plus workflow templates: Approval flow, Data pipeline, Email digest, Morning briefing.

CLI

sophon skills list
sophon skills search <query>
sophon skills install <name>
sophon skills update <name>
sophon skills uninstall <name>
sophon skills get <name>

Where to go next

• DevStudio — IDE for skill / agent / plugin / app authoring
• Skill Manifest Reference
• Approval Gates — how risk levels gate execution