# deepagents deploy
## Project layout
```
src/
AGENTS.md # required — system prompt + read-only /memories/AGENTS.md
skills/ # optional — seeded under /skills/
mcp.json # optional — HTTP/SSE MCP servers
deepagents.toml
```
## `deepagents.toml`
```toml
[agent]
name = "my-agent"
model = "anthropic:claude-sonnet-4-6"
# [sandbox] is optional — omit to run tools in-process.
[sandbox]
provider = "langsmith" # none | langsmith | daytona | modal | runloop
scope = "thread" # thread | assistant
# template = "deepagents-deploy"
# image = "python:3"
```
That's the entire surface. Skills, MCP servers, and model deps are
auto-detected.
## CLI
```bash
deepagents init # scaffold deepagents.toml in cwd
deepagents dev --config src/deepagents.toml [--port 2024]
deepagents deploy --config src/deepagents.toml [--dry-run]
```
## Runtime
- **System prompt:** `src/AGENTS.md` verbatim, baked in at build time.
- **Memories:** `/memories/AGENTS.md` in the LangGraph store, namespace
`(assistant_id, "memories")`. Read-only at runtime — edit the source
file and redeploy.
- **Skills:** `/skills/<skill>/...` in the store, namespace
`(assistant_id, "skills")`. Also read-only.
- **Sandbox:** default backend. Per-thread cache by default; set
`[sandbox].scope = "assistant"` to share one sandbox across all threads
of an assistant. Omit `[sandbox]` entirely to fall back to an in-process
`StateBackend`.
- **MCP:** HTTP/SSE only. Stdio is rejected at bundle time.
## Gotchas
- `/memories/` and `/skills/` are read-only. Edit source files and
redeploy.
- `deepagents deploy` creates a new revision on every invocation (full
cloud rebuild). Use `deepagents dev` for iteration.
- The in-process sandbox cache does not survive process restarts;
thread-scoped sandboxes get re-provisioned if the server recycles.
- Custom Python tools are not supported — use MCP servers.
---------
Co-authored-by: Mason Daugherty <github@mdrxy.com>
Co-authored-by: Mason Daugherty <mason@langchain.dev>
Sydney Runkle