StrataDB is built to be learned by agents, not just humans: the wheel and the binary carry their own documentation, errors teach the fix, and one command installs a Claude Code skill into your repo. This page is the shortest path from “use StrataDB for this” to an agent that writes correct code.

One step: install the skill

In any repo an agent works on:

strata agents skill --write

{
  "next": null,
  "path": ".claude/skills/strata/SKILL.md",
  "state": "created"
}

That file is the condensed playbook — opening a database, choosing a primitive, branch isolation, as_of reads, the error contract, and inference — version-stamped from the installed binary. Claude Code loads it automatically whenever a session touches Strata. Re-running is idempotent; it refuses to overwrite a file you’ve edited unless you pass --force.

Cursor and Codex too

One flag installs for every major coding agent, each via its native surface:

strata agents skill --write --for all

{
  "written": [
    {"agent": "claude", "path": ".claude/skills/strata/SKILL.md", "state": "created", "next": null},
    {"agent": "cursor", "path": ".cursor/rules/strata.mdc", "state": "created", "next": null},
    {"agent": "codex", "path": "AGENTS.md", "state": "appended", "next": null}
  ]
}

Cursor gets an MDC rule with the same trigger description (alwaysApply: false, attached when a session touches Strata). Codex reads the repo’s AGENTS.md, so the playbook lands there between <!-- strata-skill:start --> markers — Strata owns the marked region, never the rest of your file, and re-runs update it in place. --for claude, --for cursor, and --for codex target one agent at a time.

The same text ships inside the Python wheel, so an agent (or a setup script) can install it without the CLI:

from pathlib import Path
import stratadb

path = Path(".claude/skills/strata/SKILL.md")
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(stratadb.agents_skill())

The self-describing surface

Everything an agent needs is one hop from wherever it lands:

import stratadb

stratadb.agents_guide()      # the complete Python guide, offline, in the wheel
stratadb.command_index()     # the machine-readable command catalog (the IDL)
stratadb.mcp_config("./mydb")
# {'command': 'strata', 'args': ['./mydb', 'mcp', 'serve']}

strata agents guide            # the CLI guide, version-matched
strata agents commands --json  # command catalog
strata agents errors --json    # every public error code with hints
strata agents init --apply     # plants a pointer block in AGENTS.md/CLAUDE.md

Two rules that prevent most agent mistakes

1. Always target the database explicitly. Pass a path (stratadb.open("./mydb"), strata ./mydb …) or set STRATA_DB. Strata never opens the current directory implicitly — the refusal is invalid_argument.cli.no_database.
2. Recover by error code, never by message text. Failures carry a stable <class>.<area>.<detail> code, a hint, and a per-code URL (https://stratadb.org/e/<code>). A missing key/document/path is None, not an error.

MCP, for chat-based agents

The binary is also an MCP server — the path for Claude Desktop, claude.ai, and any MCP client:

strata ./mydb mcp serve

Client config: {"command": "strata", "args": ["./mydb", "mcp", "serve"]}. Roughly 20 curated tools plus strata_guide (the usage guide as a tool) and strata_command (any cataloged command as raw wire JSON). Same value shapes, same error codes as the CLI and SDK.

For the agent reading this page

Fetch /llms.txt for the curated index, or append .md to any docs URL for CommonMark. The full corpus is at /llms-full.txt.

Next

• For AI agents — the full machine surface, in depth.
• MCP server — the handshake and the tool table.
• Python SDK — the surface the skill teaches.
• Cookbook — agent-shaped end-to-end recipes.