Incident replay exposes weak agent tooling fast. The first fetch works. The rerun stalls because the prompt is still carrying pagination, overlap handling, gap checks, and state reconstruction. At that point you have a wrapper around raw historical glue.
This guide sits beside the broader native-versus-managed decision. The narrower question here is when an MCP, CLI, or agent loop needs replayable historical state instead of only raw venue access.
Short answer: if the job is a one-off quote or a recent snapshot, raw API access is often enough. If the workflow has to rerun cleanly, reproduce an incident window, or stay deterministic across repeated jobs, the historical layer starts to matter much more than the first fetch.
In 30 Seconds
Agent workflows need stable primitives, not prompts that rediscover historical glue every run.
Replayability matters when you want deterministic reruns, incident reconstruction, or repeated batch jobs.
0xArchive already exposes public CLI, MCP, replay/backtesting docs, coverage endpoints, and tiered historical depth.
The safe operating model is managed history for repeated work and native venue checks for disputes.
The prompt smell test
If any of these rules live in the prompt instead of the tool, the surface is still too raw for repeated agent work:
Pagination or cursor walking logic.
Boundary overlap handling between windows.
Gap checking or coverage decisions before a pull.
State reconstruction before the actual workflow can start.
What you can check today
0xArchive's agent-readable llms.txt covers CLI, MCP, replay/backtesting support, coverage routes, and order-level historical endpoints. The public MCP repo lists 73 tools, including get_summary, get_orderbook_history, get_data_incidents, and coverage tools. The public CLI repo and the published CLI walkthrough both show command shapes such as oxa auth test, oxa summary, and oxa orderbook history.
Where raw API access is enough and where replayable history helps
Job | Raw API is enough when... | Replayable history is worth testing when... | Public surface |
|---|---|---|---|
Morning market check | You only need a current quote, funding, or summary snapshot. | The job must rerun reliably across symbols or operator handoffs. | MCP |
Incident replay | You only need to confirm current venue state. | You need to reproduce a suspicious historical window and inspect it again later. | Backtesting docs, replay event types, checkpoint history routes |
Daily batch agent | The job is shallow and can tolerate manual retries. | The workflow needs coverage checks, repeatability, and stable output every day. | Coverage endpoints, CLI, MCP |
Migration evaluator | You are still testing by hand and only need one recent pull. | You want to measure whether the second run removes real glue work. |
Two repeatable examples
MCP incident check
Tool: get_data_incidents
Question: "Any data incidents this month?"
Fallback: compare the affected window back to native venue truth
Use when the job repeats and needs a stable tool contract.CLI daily summary
oxa auth test
oxa summary --exchange hyperliquid --symbol BTC
oxa orderbook history --exchange hyperliquid --symbol BTC --start 2026-03-01T00:00:00Z --end 2026-03-01T01:00:00ZWhen this framing is overkill
If you are still exploring ideas by hand and only need recent-window pulls, this is heavier than you need. The same goes for teams that already run mature collectors and replay loaders internally. The point is simpler than that: repeated workflows reveal whether the tool matches the job or hands the plumbing back to the operator.
How to test one workflow without overcommitting
Pick one job that recurs: a market summary, incident replay, or parity evaluation.
Define the output you need before you wire tools together.
Use the public CLI or MCP surface first so the workflow starts from an explicit contract.
Track whether the prompt still owns pagination, state prep, or coverage decisions.
Keep one native venue check in the loop for disputes.
Final takeaway
Raw API access proves the call path. Repeated agent work asks for more: a workflow that reruns cleanly without hiding historical plumbing in the prompt. Managed surface for repeated work, native venue checks for disputes, is the safer operating model once the job starts repeating.
Start the 14-day Build trial with no credit card and test one recurring MCP, CLI, or replay workflow before you automate anything larger.
Related reads
Hyperliquid Historical Data API for Backtesting: Native API vs 0xArchive
How to Run a One-Symbol Parity Test Before You Migrate Historical Data
When Checkpoint History Is Enough and When You Need Order-Level Depth