Skip to main content

argon-agents

PyPI CI

Argon adapters for AI agent frameworks: sandboxed, versioned MongoDB for LangGraph and Mem0.

Argon versions MongoDB the way Git versions code — branch, time-travel, diff, merge, undo. This package gives agent frameworks the two things plain MongoDB can't:

  1. A disposable copy of state to work against. Fork a sandbox with a TTL, point the agent at an ordinary connection string, and production data stays isolated until you explicitly merge the reviewed changes.
  2. An adopt-or-reject story for what the agent did. Diff the sandbox, preview a merge and explicitly apply its reviewed plan, undo supported captured ranges within retained history, or let the running API reclaim the sandbox after its TTL.

Install

SDK 0.2.0 and 0.3.0 require Argon 2.1.1 or a compatible later release; credential=True needs 2.10.0 and the index report 2.11.0. We recommend Argon 2.11.1, which includes the earlier capture, import and Go module fixes and the fix for GO-2026-6865 in a compression dependency. Version 2.1.1 remains compatible; 2.1.0 has a shutdown bug. Release wheels and source archives are available from GitHub Releases.

python3 -m pip install 'argon-agents==0.2.0'
python3 -m pip install 'argon-agents[langgraph]==0.2.0'

These commands install SDK 0.2.0 from PyPI. The release workflow verifies the registry and installs its wheel in a fresh environment before declaring a publication complete. See release operations.

Requires a running Argon API server (cd api && go run .) backed by MongoDB 7 as a replica set. The managed API waits for capture readiness, synchronizes versioned operations and sweeps expired sandboxes every minute. Stop native writers before release or discard. The public hosted demo does not expose native connection strings.

The client

from argon_agents import ArgonClient

argon = ArgonClient("http://localhost:8080")
argon.get_or_create_project("support-bot")

sandbox = argon.create_sandbox("support-bot", ttl_minutes=60, actor="agent:run-42")
db = sandbox.pymongo_database()        # plain pymongo, isolated copy
db.tickets.insert_one({"_id": "t1", "status": "resolved"})

print(sandbox.diff())                  # what the agent changed
plan = argon.merge_preview("support-bot", sandbox.branch)
print(plan)                            # inspect changes and conflicts
# After reviewing this exact plan, explicitly apply it:
# argon.merge_apply(plan["id"])
# Or reject the proposal with sandbox.discard().

merge_preview creates a plan without changing the target branch. Review that plan before calling merge_apply(plan["id"]); a stale plan must be previewed again. The convenience methods sandbox.merge() and saver.merge() preview and apply immediately, with no pause or approval step. Use them only when that automatic application is intentional in your own workflow.

The actor labels the entire branch/run, not individual MongoDB clients. Use a separate sandbox for each agent. For a protected API, pass ArgonClient(api_url, token=...); never give an agent a production service credential. Inspect argon.capture_status() if capture reports degraded.

From SDK 0.3.0 with Argon 2.10.0 or later, create_sandbox(..., credential=True) and ArgonCheckpointSaver.from_sandbox(..., credential=True) return a connection string for a MongoDB user that can read and write only that sandbox's database. Argon returns its password once and revokes the user when the sandbox is discarded or swept. The deployment must enforce authentication and the engine needs userAdmin on the sandbox databases; otherwise the API refuses the request. An expired sandbox's credential works until a sweep runs. The Mem0 helper does not take the option because it runs collMod, which that user may not do.

With Argon 2.11.0 or later, a sandbox's checkout also builds the secondary indexes recorded for its branch. sandbox.indexes is the number built and sandbox.index_notes names TTL indexes built without expiry and indexes the data violates. Older engines report 0 and an empty list.

New application collections must enable changeStreamPreAndPostImages before rapid updates; the LangGraph and Mem0 adapters do this themselves. An update without exact images or unsupported drop/rename stops capture with an explicit degraded status. Retention limits history unless pinned.

Run the complete review workflow

The two-agent example uses ordinary pymongo and no paid model: both agents start from one pin, propose different order prices, merge the reviewed result, surface the competing conflict, discard it, then undo the adopted change and assert the original data is restored.

ARGON_API_URL=http://localhost:8080 python examples/two_agent_review.py

LangGraph

from argon_agents import ArgonClient, ArgonCheckpointSaver

argon = ArgonClient()
argon.get_or_create_project("support-bot")
saver = ArgonCheckpointSaver.from_sandbox(argon, "support-bot", ttl_minutes=60)

graph = builder.compile(checkpointer=saver)   # any LangGraph graph
graph.invoke(input, {"configurable": {"thread_id": "user-42"}})

plan = argon.merge_preview("support-bot", saver.sandbox.branch)
print(plan)                           # review the checkpoint-store changes
# After review: argon.merge_apply(plan["id"])
# Or saver.discard() to reject, or saver.fork(argon) to try another branch.

ArgonCheckpointSaver is the official langgraph-checkpoint-mongodb saver — same wire format, same semantics — running on an Argon branch. LangGraph's checkpoint ids give step-level rewind within a thread; Argon adds branch-level fork/merge/undo/audit across the whole store.

Mem0

Mem0 speaks MongoDB natively; Argon supplies the versioned sandbox:

Semantic search requires MongoDB Atlas Search or a compatible Search deployment. A plain replica set supports versioned document storage but does not implement $vectorSearch. Provision search indexes separately for each branch; Argon versions documents, not search-index definitions. Configure the LLM/embedder required by Mem0 before running Memory.

from argon_agents import ArgonClient, sandboxed_mem0_config
from mem0 import Memory

argon = ArgonClient()
argon.get_or_create_project("support-bot")
config, sandbox = sandboxed_mem0_config(argon, "support-bot")
memory = Memory.from_config({"vector_store": config})

# ... let the agent read/write memories ...
plan = argon.merge_preview("support-bot", sandbox.branch)
print(plan)                           # review the memory changes
# After review: argon.merge_apply(plan["id"])
# Or sandbox.discard(); the running API also sweeps expired sandboxes.

Reproducible evals: dataset pins

A pin is a named, immutable reference to a branch state. While the pin exists, it protects the history it references from garbage collection and resets. Deleting the pin removes that protection; keep independent backups for loss of the underlying deployment. Pin the eval dataset once, then fork a fresh sandbox from the same retained pin for each run:

argon.create_pin("my-project", "eval-v1", note="golden dataset")

run = argon.sandbox_from_pin("my-project", "eval-v1", ttl_minutes=30)
# ... run the eval against run.connection_string ...
run.discard()          # the pin remains; fork it again while it exists

Tests

pip install -e ".[dev]"
ARGON_REQUIRE_STACK=1 MEM0_TELEMETRY=false pytest

CI checks Python 3.10, 3.12 and 3.14 against Argon v2.11.1 and fails if the stack is unavailable. The dispatch input engine_ref accepts an exact engine commit or release tag; the resolved SHA is logged. Tests exercise mandatory conflicts, actual undo, LangGraph invoke/ainvoke and fork isolation, pinned input, and Mem0's real MongoDB insert/get/update/delete methods. The Mem0 document test bypasses Atlas index creation only; it does not claim semantic-search coverage on plain MongoDB.

Metadata

Release files for argon-agents 0.3.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for argon-agents 0.3.0
File Size Uploaded Provenance Actions
argon_agents-0.3.0.tar.gz 18.8 kB With provenance Details for argon_agents-0.3.0.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for argon-agents 0.3.0
File Interpreter ABI Platform Actions
argon_agents-0.3.0-py3-none-any.whl Python 3 none any Details for argon_agents-0.3.0-py3-none-any.whl

Total release size: 31.1 kB

Release files / argon_agents-0.3.0.tar.gz

Download URL argon_agents-0.3.0.tar.gz
Size 18.8 kB
Tags Source
SHA-256 checksum
How to use checksums
fa3b77d1b9a1fa123a8498b8cd19e6302cdb97785ad7c170a041a560b3909039
BLAKE2b-256 checksum
How to use checksums
9c028b5a8bcfd3ae73d822ed2f1751a011aebe0343b8033b864cd3f06087fcb2
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 21, 2026.

Transparency log

Release files / argon_agents-0.3.0-py3-none-any.whl

Download URL argon_agents-0.3.0-py3-none-any.whl
Size 12.3 kB
Tags Python 3
SHA-256 checksum
How to use checksums
e7ed64b2a571ca1050d355082a9a368a4322f9b101a061fc7744771600ed4f5d
BLAKE2b-256 checksum
How to use checksums
a8defebe5aeac605c9c3a72d092942cfd5007bd3b3442df2545232ff89f65c6d
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 21, 2026.

Transparency log

Security for argon-agents 0.3.0

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for argon_agents-0.3.0.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 21, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page