Skip to main content
Glama

Argon — MongoDB sandboxes for AI agents

Build Status Go Report License: MIT Homebrew npm PyPI mcp.market

Give every agent run its own real MongoDB database, forked from your data. Review what it changed, merge what you want, discard the rest. Git for MongoDB underneath: branch, diff, merge, time travel and undo. Any driver, real mongod.

Three ideas, thirty seconds:

  1. A sandbox is a real MongoDB database — forked from a branch or a pinned dataset, checked out and given a TTL in one command. pymongo, mongoose, mongosh, indexes, aggregation, transactions: all real.

  2. What the agent wrote is a diff — supported document writes become history while capture is healthy, so a run can be reviewed, merged as a data pull request, undone or thrown away. Your source data is untouched until you merge.

  3. Underneath, it is Git for MongoDB — a branch is a pointer created with a metadata write; checkout separately materializes the data. Diff, merge, undo and pins work within the configured retention and capture guarantees. The commands are side by side with Git below.

Install

For npm, use Node.js 24 LTS with npm 11.16 or newer. The package-scoped approval allows the installer to download Argon's binary on npm 12.

brew install argon-lab/tap/argonctl      # macOS or Linux, builds from source
npm install -g argonctl --allow-scripts=argonctl # cross-platform

# MongoDB must run as a replica set (one-node is fine):
docker run -d --name argon-mongo -p 127.0.0.1:27017:27017 mongo:7 --replSet rs0
docker exec argon-mongo mongosh --quiet --eval 'rs.initiate({_id:"rs0",members:[{_id:0,host:"localhost:27017"}]})'
argon doctor

Use a current supported MongoDB patch release in production. Source builds use Go 1.26.11 or newer, as declared in go.mod. Use a patched toolchain: on the Go 1.27 line that means 1.27.4 or newer, which carry the same standard-library fixes.

Related MCP server: askdb

The flow

your data ──sandbox──▶ mongodb://…  ← the agent, any driver
                           │
          argon diff ──────┤  what the run changed
              ▼            ▼
   merge (a data PR)   or  discard / undo
# 0 · Bring your data in once, from a database nobody is writing to — or: argon projects create myapp
#     (a database that is in use: see "A database that is in use" below)
argon import database --uri mongodb://localhost:27017 --database myapp --project myapp --source-quiesced

# 1 · In terminal A, a sandbox: fork + checkout + TTL + capture in one command.
#     Prints its name and a connection string, then keeps capturing until Ctrl-C
argon sandbox create -p myapp --ttl 1h --watch --actor agent:run-1

# In terminal B, point the agent at the printed URI. Then:

# 2 · Review, then adopt the run — a data pull request
argon diff          -p myapp -b <sandbox>
argon merge preview -p myapp -b <sandbox>
argon merge apply <plan-id>

# …or throw the run away
argon sandbox discard -p myapp -b <sandbox>

Through the MCP server or argon console, capture and expiry are managed for you: nothing stays in the foreground, and expired sandboxes are swept every minute. --watch needs 2.3.0 or newer; on older versions run argon watch -p myapp -b <sandbox> --actor agent:run-1 after creating the sandbox. Long-lived branches work the same way with argon branches create and argon checkout, and argon restore reset rewinds a branch to an earlier time.

A database that is in use

Online import and the argon source commands, introduced in 2.6.0, are now available by default. You can import a replica-set database while writers keep running, then follow its changes on the project's main. Write-back is also available by default. Review the plan in CLI JSON and pass its digest when applying it to the source. The experimental options --auto-resync and --enable-source-images are off by default and each requires ARGON_EXPERIMENTAL_SOURCE=1. Write-back tests have not covered Atlas, deployments with access control, multi-member elections, sharded sources or MongoDB versions other than 7.0. For known issues and limits, see the operations guide.

export ARGON_SOURCE_URI="mongodb://…"     # read per command, never stored
argon import preview --database myapp --online
argon import database --database myapp --project myapp --online
argon source follow -p myapp              # main tracks the source; leave it running

argon sandbox create -p myapp --ttl 1h --watch --actor agent:run-1
argon merge preview -p myapp -b <sandbox> --output json
argon merge apply <plan-id> --write-back --approve <approval_digest>

The import copies ordinary collections at one cluster time, which it prints. Add --dry-run to the import command to check the source and its change stream and see the estimated copy without writing anything. Following ignores changes to views. New merge plans save ordered BSON. CLI JSON preview and list show the order apply writes, preserve BSON types with canonical Extended JSON, identify literal documents with dollar-prefixed fields and show exact bytes wherever JSON loses a BSON distinction. For a pending plan targeting a followed database, inspect this output and pass its approval_digest to --approve. To inspect a saved plan, run argon merge list -p myapp --output json. Plans saved by older versions require a fresh preview. Write-back requires a verified, saved before-image for every replacement or delete and a verified after-image for every insert or replacement. An empty plan completes in metadata. For a nonempty plan, Argon checks capture health and claims the plan with majority acknowledgment, then checks existing documents byte for byte and insert keys for absence in one source transaction. Before commit, it reads each affected key again to verify the exact stored BSON or confirm deletion. A mismatch rolls back the batch and returns the plan to pending. A top-level zero BSON timestamp outside _id must be changed on the editable branch before write-back. Detach refuses while a recent claim is unsettled and reports older unsettled plans for inspection. Write-back and --enable-source-images require importing with a user allowed to run replSetGetConfig, such as one with the clusterMonitor role. Use argon source status to see which identity check applies. The MCP server and REST API do not offer write-back. Argon never drops, renames or rewrites a source database. See the operations guide.

Write-back refuses capped collections and byte-identical replacements before claiming a plan. Ordinary merge also checks the inner length of deprecated BSON binary subtype 2 when decoding stored images.

An older delete can refer to either an empty subtype-2 _id or four zero bytes. Reads that depend on it report the key. If GC removed the relevant WAL and subtype-2 ID evidence survives, reads that need the old snapshot report the unknown LSN range. For a followed branch, stop the follower and run argon source resync -p myapp --force. On an ordinary branch, fork from an earlier readable LSN with argon restore branch or rewind its head with argon restore reset. With argon import database, the copy goes into a new project and the existing branch keeps its history. See upgrade and recovery details.

If a plan stays writing_back, check the affected documents on the source, let following catch up, and run argon merge preview again for the sandbox.

Run argon source status -p myapp to see whether main is verified at its saved source cluster time. A clean stop at an exact checkpoint exits zero, as does restarting while a live follower reopens its stream from that checkpoint. An open catch-up range, a stored capture error or an expired owner exits nonzero even when the state says stopped. Table and JSON output give the reason and next action. argon status and argon doctor report the same health. Write-back preflight requires exact. See the source status table.

Each project can run one follower or resync at a time. Restart following after lease loss and let the replacement verify its checkpoint. argon source follow --auto-resync completes an unfinished resync or repairs a source stream gap. The option checks ARGON_EXPERIMENTAL_SOURCE=1 at startup.

Operators can enable changeStreamPreAndPostImages directly on MongoDB or use --enable-source-images on online import or follow. When requested, source follow --enable-source-images checks the lease before altering each collection. An error after a change lists the affected collections.

Prefer clicking? argon console serves a local web console (UI + REST API), supervises capture, reaps expired sandboxes every minute, and opens your browser. The console's MIT-licensed source is in web/; see the rebuild instructions. For a complete managed workflow, run the two-agent pinned dataset example.

For AI agents

claude mcp add argon -- argon mcp        # 13 tools: sandbox, diff, merge, undo, pins
# Install the published Python SDK and LangGraph adapter.
python3 -m pip install 'argon-agents[langgraph]==0.2.0'

Start argon console --no-browser in another terminal, then:

from argon_agents import ArgonClient, ArgonCheckpointSaver

argon = ArgonClient("http://127.0.0.1:1818")
argon.get_or_create_project("myapp")
saver = ArgonCheckpointSaver.from_sandbox(argon, "myapp")
# Compile and run your LangGraph graph with checkpointer=saver, then review:
plan = argon.merge_preview("myapp", saver.sandbox.branch)
print(plan)
# After reviewing this exact plan, apply it explicitly:
# argon.merge_apply(plan["id"])
# Or reject the run: saver.discard()

argon.create_pin("myapp", "eval-v1")                        # pin the dataset once
run = argon.sandbox_from_pin("myapp", "eval-v1")            # identical state, every eval run

saver.merge() is a convenience call that previews and applies immediately; use the separate calls above when approval is required. Pins protect captured document state while they exist; keep independent backups.

To keep a person in the loop, withhold tools from the MCP server: argon mcp --disable-tools argon_merge_apply leaves applying a merge plan to argon merge apply <plan-id>, and argon mcp --read-only offers only the tools that inspect state.

Agent-written code that runs in a code sandbox can write to an Argon sandbox through a MongoDB user made for one run: examples/boxd-sandbox does that with an isolated boxd machine, and boxd's documentation has a page for the integration.

The full agent workflow: docs/AGENTS.md.

Git for MongoDB

Underneath the sandbox, Argon does for MongoDB documents what Git does for files. Side by side:

In Git

In Argon

What it does to your data

git branch

argon branches create

adds a pointer into shared history; no documents are copied

git worktree add

argon sandbox create

gives a new branch its own working copy: a real MongoDB database, with a TTL

git diff

argon diff

lists the documents a branch would add, change or remove

git merge

argon merge preview, then argon merge apply

saves the change as a plan to review, then applies exactly that plan

git checkout <commit>

argon time-travel query

reads the data as it was at an earlier position in retained history

git revert

argon undo

reverses earlier writes with a new change; history is not rewritten

git tag

argon pin create

names one fixed state and protects its history while the pin exists

Where it differs from Git: there is no commit step, because a running capture process records each document change. A merge goes from a branch into its parent and compares whole documents. Checking out a branch copies its data, so it takes time and storage that depend on the dataset. Documents are versioned; indexes and collection options are not. History older than the retention window can be reclaimed unless a pin protects it.

What you get

Command

In one line

Sandboxes

argon sandbox create --ttl 1h

fork + checkout + TTL in one step — disposable agent workspaces

Dataset pins

argon pin create / pin sandbox

named captured states protected from GC and resets while the pin exists

Data PRs

argon merge preview/apply

three-way merges as reviewable plans; conflicts never silent

Undo

argon undo --actor <a>

revert a range or actor label; append-only, conflict-aware

Real databases

argon checkout / argon proxy

any driver, real mongod; proxy serves stable mongodb://host/<project>~<branch> URIs

Write capture

argon watch

exact change-stream images → history, one actor label per branch

Branching

argon branches create

a metadata write — instant, zero copy

Time travel

argon time-travel query

query captured document states at retained LSNs; restore/pin also accept timestamps

Restore

argon restore preview/reset/branch

rewind a released branch or fork retained history

Web console

argon console

local UI + REST API in one command

Data history covers document inserts, updates, replacements and deletes. Collection drop/rename produces degraded capture; indexes and collection options are not versioned. Native writes are asynchronous; control operations drain capture. Stop writers before release. Existing pins protect their referenced states, while GC can expire audit and undo history. See operations for recovery and credentials.

Two things to know before handing a sandbox to an agent. A collection the agent creates is captured exactly while capture is running, or when it is next synchronized within MongoDB's snapshot history (five minutes by default); the one exception is a document that is updated and then changed again inside one transaction on a collection created moments earlier. Before 2.3.1 such a collection needed argon collections prepare first. And a sandbox is a full physical copy with capture ready, so creating one takes longer as the data grows: the benchmarks measure it as sandbox_capture_ready_ms.

Storage retention uses snapshots + retention-window GC to keep state plus a window of history, not every write forever. Details and the consistency model, stated honestly: ARCHITECTURE.md.

Documentation

Quick start

install → first merge, step by step

CLI reference

every command

Agents

sandboxes, pins, MCP, REST API, argon-agents, proxy

Architecture

how it works and what it guarantees

Operations

deployment, chunk stores (S3/FS), GC, v1→v2 migration

Performance

every number lives in the reproducible benchmarks — none here, by policy

Community

Jake Wang maintains Argon.

Issues · Discussions · Contributing · argonlabs.tech · Blog

Contributors

Thank you to everyone who has made Argon better.

What each person shipped, and the release it shipped in, is listed at argonlabs.tech/about#contributors. Every release also names its new contributors in the release notes.

New here? Start with a good first issue and the first-contribution guide.


Give your MongoDB a time machine. Branch without fear. ⭐

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables natural language querying of MongoDB data by transforming AI agent questions into MongoDB aggregations. Supports secure data separation, semantic vector search, and advanced lookup capabilities for database interactions.
      5,254 npm
      6
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Self-hosted MCP server that bridges your MongoDB or PostgreSQL database to AI agents, with sandbox isolation and field-level control.
      71
      MIT
    • A
      license
      A
      quality
      D
      maintenance
      Enables AI agents to connect to Databend databases safely with a session sandbox that keeps production read-only while agents experiment.
      13
      9
      Apache 2.0