Loading ToolBench
Git for MongoDB: branch, time-travel, merge and undo - MCP server exposing Argon to AI agents over the Model Context Protocol (stdio, JSON-RPC 2.0, newline-delimited). Supervises change-stream ingesters for sandboxes so agent writes become versioned history.
Argon provides 13 well-named tools with complete JSON Schema input definitions and clear, domain-specific descriptions (avg 156 chars). All tools follow verb_noun naming (argon_sandbox_create, argon_branch_list, etc.). Schemas include required fields, type definitions, and parameter descriptions. However, output schemas are not documented, LLMs cannot predict response structure for chaining. Error handling is minimal; tools lack recovery guidance. No tool annotations (readOnlyHint, destructiveHint) despite clear risk classifications. Parameter descriptions are adequate but could specify constraints (e.g., ttl_minutes range, actor label format). Missing idempotency guidance for write operations.
List a project's branches with their heads, checkout state and TTLs.
Get a MongoDB connection string for a branch, checking it out if needed. Writes through that connection become versioned history.
Show what merging a branch into its parent would change: adopted documents and conflicts.
Apply a pending merge plan. Conflicted plans need a strategy: theirs (take the branch) or ours (keep the target).
Compute and persist a reviewable merge plan (a data pull request). Returns the plan ID.
Pin a branch state under a name: a named, immutable dataset reference that survives garbage collection and resets forever. Pin an eval dataset once, then fork sandboxes from the pin for every run and get identical input state.
Output schemas not documented. LLMs cannot predict response structure (fields, types, pagination) for downstream tool chaining. E.g., argon_branch_list returns branch metadata but structure is opaque.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools are classified by risk (READ_ONLY, WRITE, DESTRUCTIVE, REVERSIBLE) but not exposed to LLM via schema. Agents cannot reason about safety without explicit hints.
Parameter constraints underspecified. ttl_minutes, from_lsn, to_lsn lack min/max bounds. actor label format not documented. LLMs may pass invalid values (negative TTL, out-of-range LSN).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | B | 70 | 2026-07-28+ | v2 |
List a project's pins (name, branch, LSN, note).
Fork a TTL sandbox from a pin and get a connection string. The sandbox starts at exactly the pinned state — the reproducible-eval workflow.
Fork an isolated, disposable copy of a MongoDB branch and get a connection string. Work against it with any MongoDB driver; every write is captured as versioned history. Merge the result back, undo it, or let the TTL reclaim it.
Delete a sandbox immediately and reclaim its storage. Unmerged changes are lost.
Remove a sandbox's TTL, keeping it as a permanent branch.
Snapshot a branch at its current head so reads replay only the delta above it.
Revert a range of a branch's history: every touched document returns to its state before the range. Optionally restricted to one actor's writes. Use dry_run to preview.
Error handling absent. No recovery guidance in tool definitions. If argon_merge_apply fails due to conflict strategy mismatch, LLM has no hint to retry with different strategy.
Idempotency not declared. Write operations (argon_sandbox_create, argon_pin_create) do not state whether repeated calls with same params are safe. Agents may retry and create duplicates.