Every decision on the record.Every session readable.
Enterprise deployments include a console for the people who run your agents: full transcripts, answers about your traffic, session summaries, content filters, usage, plans and keys.
The transcript explorer.User, session, thread, turn, step.
Sessions are named the way your clients name them: Claude Code's session id, Codex's prompt cache key, pi's session affinity. Subagents and side calls get their own threads. Each step stores what was new, the full answer, and the routing decision behind it.
now add tests for the empty page and the last page
Added two tests: an empty page returns an empty list with no next cursor, and the last page has no next link. Both pass.
Six views of a session
Conversation, Prompts and answers, Reasoning, Tool calls, Routing and Raw, each linked to the turn it came from.
Full-text search
Every prompt, answer, reasoning block and tool call is indexed. Find the session where an error first appeared.
Filters in the URL
User, app, tier, model, complexity, kind, decision, tool and errors. Share a filtered view with a link.
Ask the data.Get answers with sources.
Pick users, sessions and a time window, then ask a question in plain words. An agent investigates with read-only tools scoped to your selection, and the answer cites turns and sessions as links.
- Where did sessions get stuck this week, and why?
- Which tasks went to Smart and could have run on Normal?
- What did the team ask about the billing service?
Three sessions stalled on the same fixture: the tests expect a fixed clock and the fixture reads the system time session 7f3a91 turn 12.
Two recovered once the agent pinned the time session c20e4d. The third stopped after the user switched tasks session 91b7aa turn 8.
- Title
- Pagination for the orders API
- Goal
- Paginate /orders and cover the edge cases with tests
- Status
- achieved
- Progress
- Endpoint paginated with a cursor, two tests added, suite passing
- Problems
- The first attempt changed the default page size
- Next
- Document the cursor parameter
- User mood
- satisfied
Session summaries.Every session at a glance.
Once a session goes quiet, it is summarized from its previous summary plus only the new turns. Every summary has the same shape, so you can scan a hundred.
Content filters.Log it, or stop it at the door.
A filter is a yes or no rule that reads the sides of the traffic you choose: prompt, response, reasoning, tool call or tool result, with an optional regex prefilter. Its action is log or block. A blocked input is never forwarded and returns 403; a blocked reply is cut and the client gets an error event.
| Filter | Decides | Reads | Mode | Action |
|---|---|---|---|---|
| Private keys | Pattern BEGIN [A-Z ]*PRIVATE KEY | prompt, tool result | regex | block |
| Destructive commands | "Does this command delete data outside the project, or force-push a shared branch?" after a prefilter on tool name bash | tool call | classifier | log |
| Customer data in replies | "Does this reply contain a customer's personal data?" | response | classifier | block |
| Off-task use | "Is this request unrelated to software work?" | prompt | classifier | log |
| Mode | Who decides | Cost a check | Time | On a streamed reply |
|---|---|---|---|---|
| regex | the pattern, locally | none | 12 microseconds on a 14 KB item | a 512-character tail is held back, so a match is cut before it is sent |
| classifier | a small decision model | $0.000015 to $0.000023 | 165 ms p50, 199 ms p95 | the reply is scanned before it is released |
On real agent sessions, logging everything cost 1.2% of model spend. A block gate on the way in cost $0.0006 for 40 calls.
Usage analytics.Per request, per token.
Every request is metered with its input, output and cached tokens, its time to first byte, and its cost. Slice it by the same filters as the explorer.
Plans and keys.Limits per team.
Give each team or product its own client, plan and keys. A plan sets the tiers, requests per minute, concurrent requests, monthly token and spend caps, and per-tier prices. A key is shown once and stored as a SHA-256 hash.
| Plan | Tiers | Req/min | Concurrent | Tokens/month |
|---|---|---|---|---|
| Starter | fast, normal | 60 | 4 | 20M |
| Pro | fast, normal, smart | 300 | 16 | no cap |
| Enterprise | all four | yours | yours | yours |
- Transcripts
- on
- Raw parameters
- on
- Retention
- 30 days
- After that
- bodies deleted; tokens, timing, model and decision kept
- Credentials
- stripped from logged headers
Your own lanes.Your own console.
Keys are issued on request. Tell us what you are building and which clients you use, and we will send a key by email.