Replay-mode managed-agents for finance teams

Replay-mode managed-agents for finance teams. Earnings briefs, audits, reconciliations. Part of the subagentdispatch fleet — same chassis, same replay layer, vertical-specific corpus. (2 cassettes on this site.)

See the demoscurl the replay endpoint →API docs
Replaying · 0 API calls · 0 tokens

Earnings-reviewer — Pre-print brief for an enterprise SaaS Q1

claude-haiku-4-5
user ▸ Produce a 4-bullet pre-print brief: consensus vs. whisper, net retention, AI monetization, bear thesis.
• Consensus vs. whisper — street EPS is $0.42 with buy-side whispers near $0.46, so a print below $0.44 triggers immediate guide-down pricing.
• Net retention — NRR has slipped 110bps over the last two prints; another step-down through 115% invalidates the durable-growth narrative.
• AI monetization — investors want a hard number on AI-attached ACV; commentary without a dollar figure will be read as a soft miss.
• Bear thesis — federal contract concentration plus the looming GSA renewal is the swing factor; unfavorable renewal language could compress NTM revenue ~3%.
Streamed from /api/v1/replay/finance/earnings-reviewer · R2-backed cassette · zero tokens spent.
0
API tokens at runtime
0
Cassettes ready
0
Fleet sites live
Native
MCP server

Platform capabilities

Record once. Replay everywhere. Stream cassettes from Cloudflare R2 with zero tokens spent.

SSE replay from R2

Every demo streams a recorded cassette through /api/v1/replay/<slug>. The chunk shape, delta_ms timing, and tool-use boundaries match the original Anthropic response byte-for-byte.

MCP server

Connect from Claude Desktop, Cursor, or any MCP-compatible AI tool to ingest cassettes, query the replay layer, and trigger record-mode against your own Anthropic key.

Per-vertical fleet

subagentlegal, subagentfinance, subagentengineering — each fleet site runs the same chassis with a SITE_VERTICAL var that filters cassettes and themes the hero. Add a vertical in one commit.

Built on

Edge-boundary recording (Netflix replay pattern). Cassettes on R2. SSE streamed verbatim.

R2
Cassette storage

HAR 1.2 + _anthropic SSE chunks, keyed by sha256(method + url + body)

Workers
Streaming runtime

OpenNext on Cloudflare. Replay endpoint emits SSE with original delta_ms timing

AI Gateway
Record-mode routing

Live / record traffic goes through Cloudflare AI Gateway — retries, observability, cost capture

Demo cassettes

2 in the subagentfinance vertical — every one streamed live from /api/v1/replay/<slug>.

Earnings-reviewerfinance
/api/v1/replay/finance/earnings-reviewer
claude-haiku-4-5
KYC-screenerfinance
/api/v1/replay/finance/kyc-screener
claude-haiku-4-5

How it works

Three steps from a live Anthropic API call to a zero-token replay.

[●]
1

Record once

REPLAY_MODE=record routes through Cloudflare AI Gateway and captures every Anthropic request + SSE response chunk as a HAR cassette in R2.

Edge-boundary capture (Netflix pattern)
</>
2

Match by request class

Cassettes key on sha256(method + url + normalized_body). Request IDs, UUIDs, and timestamps are stripped so the same logical request matches across runs.

HAR 1.2 + _anthropic SSE extension
{=}
3

Replay everywhere

REPLAY_MODE=replay never touches the network. GET /api/v1/replay/<slug> streams the recorded SSE chunks with original delta_ms timing. Zero API key, zero tokens.

SSE · 0 tokens · 0 API calls

Search the corpus

BM25-lite over every demo cassette. Type to jump to the matching replay.

Try: