Skip to main content
Every Skydive agent is backed by a private git repository. It’s the source of truth for everything that makes the agent this agent rather than a generic one: its persona, what it knows, the tools it has, and the code it runs. You can explore it, edit it, and commit to it just like any other repository.

​
Where the repository lives

Inside the agent’s sandbox, the repository is mounted at /home/user/.self. The agent commits and pushes its changes automatically, so everything it learns and configures is persisted across sandbox restarts.

​
What’s in it

​
soul.md

The agent’s persona. It’s the same content you can edit through the dashboard, but here you can version it, diff it, and PR it. The agent reads this at the start of every conversation.

​
mcp.config.json

Which MCP servers the agent has connected. Add a server here and the agent picks it up on its next turn:
See Tools & MCP for the full connection patterns.

​
tools/

Custom in-process tools the agent can call. Each .ts or .mjs file here exports a ToolDefinition. The agent picks up new files without a restart. Files starting with _ or . are skipped.

​
skills/

One subdirectory per skill. Each contains a SKILL.md with YAML frontmatter (name, description) and the workflow instructions. The agent re-scans this on every conversation, so editing a skill file takes effect immediately. See Skills.

​
.memory/

The agent’s persistent memory, organized as plain markdown files. The directory structure reflects how the agent groups what it knows:
You can read these to understand what the agent has learned, and you can edit or delete them to correct stale facts. The agent reads from this at the start of each conversation. See Memory.

​
Dockerfile

Packages and system dependencies baked into the agent’s image. Anything installed here survives sandbox restarts. Use it for tools the agent needs reliably:
Changes take effect on the next image build.

​
start.sh

Setup that runs at boot, after credentials are available. This is where you put steps that need secrets or network identity, like cloning a private external repository. The Dockerfile runs at build time with no credentials; start.sh runs at runtime with full access to secrets and the git credential helper.

​
Exploring an agent’s state

To see what an agent remembers, inspect .memory/ in the repository. To see which tools it has, look in tools/ and mcp.config.json. The persona is in soul.md.

​
Making changes

Changes committed and pushed to the agent’s repository take effect on its next run. The agent’s sandbox is rebuilt from the Dockerfile on push, so package changes need a push to persist. In-conversation changes (memory writes, tool additions, MCP config edits) are committed by the agent automatically.
The .memory/ directory uses plain markdown files. Editing them directly is safe, but the agent may overwrite a file if it writes a newer version in its next conversation. Treat direct edits as corrections, not long-term overrides.

Tools & MCP

How the agent extends its toolset.

Memory

How agents accumulate and recall knowledge.

Skills

Reusable workflows stored in the repository.

Persona

The standing instructions in soul.md.

Platform CLI

The runtime commands the agent runs inside its sandbox.