langchain-inis

Native inis.run sandbox backend for LangChain Deep Agents (deepagents.backends.sandbox.BaseSandbox), plus a lower-level LangGraph recipe for teams not using Deep Agents.

Beta. Verified against deepagents==0.7.1 — pin langchain-inis and re-check before assuming it tracks a later deepagents major without changes.

Install

pip install langchain-inis deepagents

Quickstart: native Deep Agents backend

from inis import Client
from deepagents import create_deep_agent
from langchain_inis import InisSandboxBackend

backend = InisSandboxBackend.create()  # owned: this app destroys it
try:
    agent = create_deep_agent(model="anthropic:claude-sonnet-5", backend=backend)
    result = agent.invoke({"messages": [{"role": "user", "content": "List /workspace, then write hello.txt"}]})
    print(result["messages"][-1].content)
finally:
    backend.destroy()

No custom lifecycle tools in the model prompt — passing backend=backend is the whole integration. Deep Agents auto-registers ls/read_file/write_file/edit_file/glob/grep and execute against it, all bounded.

Owned vs. attached

• InisSandboxBackend.create(...) — a brand-new session this app owns and must backend.destroy() when done; SandboxBackendProtocol has no lifecycle hooks of its own.
• InisSandboxBackend.attach(session_id, ...) — binds to a session this app does not own. Never destroys it; call backend.close() when done, which only releases this backend's local HTTP transport.

backend = InisSandboxBackend.attach(existing_session_id)
try:
    agent = create_deep_agent(model=..., backend=backend)
    ...
finally:
    backend.close()  # local transport only — the session is not this app's to destroy

One agent thread/conversation should map to one session — resolve session_id per thread using whatever key your app already tracks conversations by. See examples/langgraph_recipe.py's ThreadSessionStore for a worked example.

What it does

• execute/aexecute (shell)
• upload_files/download_files (+ async) — base64 for binary
• Every derived file op (ls/read/write/edit/grep/glob/delete) — BaseSandbox's own defaults, unmodified
• Owned/attached lifetime split — see above
• Structured error mapping — InisError.code mapped onto FileOperationError literals where it matches

enable_capture_offload is False, forced — see below, this is the most important thing to read here.

Notes

True remote cancellation of an in-flight execute() is not implemented: BaseSandbox.execute(command, timeout=...) is timeout-only, with no abort hook. A timeout still ends the call. Streaming output is not implemented either — ExecuteResponse has no streaming shape.

Large execute() output is not capped on this backend

deepagents' automatic eviction of oversized tool results (/large_tool_results) requires writing to a directory hardcoded off the filesystem root — which fails Permission denied on inis.run's non-root, /workspace-confined session, and deepagents silently keeps the full-size message rather than raising. enable_capture_offload=False reflects that this eviction path is a no-op here, not a preference.

The practical ceiling is instead inis.run's own guest-side capture cap (up to 4 MiB per stream), well past what Deep Agents' ~80 KB budget assumes. Nothing is silently dropped — the risk is an unexpectedly large chunk of context from one verbose command. Bound noisy commands yourself (head -c/tail -c, or write to a file and read_file back a slice) until deepagents supports a configurable offload root for a plain BaseSandbox.

ExecuteResponse.truncated is set directly from inis.run's own per-stream truncation signal, so a caller reading it gets an honest answer.

LangGraph recipe (lower-level, secondary path)

examples/langgraph_recipe.py is a complete, runnable example using plain LangGraph primitives (@tool, ToolNode, StateGraph) for teams not using Deep Agents. One session per graph thread ID, a bounded tool surface, and a small file-backed ThreadSessionStore demonstrating a session ID surviving a process restart.

python examples/langgraph_recipe.py           # creates (or reattaches to) demo-thread-1
python examples/langgraph_recipe.py --end      # ... and cleans it up

Development

pip install -e ".[dev]"
pip install -e ../python   # inis core SDK, editable
pytest

Tests subclass the upstream SandboxIntegrationTests conformance suite (langchain-tests) against a fake local-subprocess inis.run HTTP backend — the network transport is faked, execute()/file ops run for real. Run on Linux (or Linux CI); several conformance tests assume GNU grep semantics that a default macOS shell doesn't have.

Links

• Docs: docs.inis.run · LangChain integration guide
• Source: github.com/inis-run/sdk/tree/main/langchain-inis
• Homepage: inis.run

Built and run in the EU. Your code and data never leave Europe.

License

MIT — see LICENSE.