Browse the documentation
Documentation
Everything the agent can do, and what it cannot.
Written against the code and dated where it was checked. Where a behaviour is defined but has not been watched working, the page says so instead of rounding it up.
Getting started
Install Snaga Code, sign in, and give an agent its first task.
Working with the agent
The REPL, Act and Plan modes, memory, and undoing what the agent did.
- Act and Plan modesPlan mode makes the agent read-only. Start there if you are letting an agent into a repository you care about.
- Undo with checkpointsIn a git repository, Snaga snapshots your working tree before every destructive action. /rewind puts it back — and the rewind is itself reversible.
- Slash commandsEverything you can type at the prompt that is not a request, grouped by what you are trying to do.
- MemoryWhat the agent remembers between turns and between sessions — SNAGA.md, stored facts, and the extraction pass that runs on its own.
- Context and compactionHow much the agent can hold at once, what happens when it fills up, and how to feed it the right things.
Agent tools
What the agent can do on your machine, and what it deliberately cannot.
CLI reference
Every subcommand and the flags that matter, including the machine-readable output modes.
Permissions and safety
The sandbox, the shell policy, and how the agent asks before it acts.
- What it can do without askingOut of the box, dangerous commands are blocked outright, dangerous tools wait for a person, and where the platform has a sandbox backend writes are confined to your project — with nothing configured to make any of it true.
- Changing the rulesTurn a hard block into a prompt, grant an exception for a session or forever, and see what is currently granted — without disarming the built-in guards.
Bridge — drive the agent from your phone
Pair a machine, send it work from the web or a phone, and approve what it asks.
- Run tasks from your phoneConnect a machine to your account and send it work from the web or a mobile app. The agent still runs on your computer, against your files.
- Approvals and safetyA dangerous action needs a human to say yes. Silence is not a yes — a timeout, a dropped connection and a refusal all deny.
- AttachmentsFiles attached to a message can reach the agent — on some paths, within a size limit, and landing on disk rather than in the conversation.
Extending the agent
Writing tools, the SPEC registry, and MCP in both directions.
- The agent writes its own toolsWhen a capability is missing, the agent can author a tool at runtime — generate Rust, compile it to WASM, and register it without a restart.
- Installing tools from the registryFind a tool someone already wrote, install it into a running session, and publish your own — with a trust policy deciding what may install without asking.
- Skills and MCPConstrain the agent to a role with a skill, and connect it to external tools — in both directions — with MCP.
Autonomous runs
Goals in TOML, run profiles, watching a run, and rolling one back.
Configuration
snaga.toml, trust.toml, environment variables, and what lives in .snaga/.
- snaga.tomlProject and global configuration — agent behaviour, memory, permissions, hooks, MCP servers — and why a config file from a repo you cloned asks before it does anything.
- Files and environment variablesWhere Snaga keeps things, what is safe to delete, and the environment variables worth knowing.