12 — Platforms, Hosted Sharing, Observability
Hosted capture adds collaboration above native stores, while telemetry and orchestrator history omit full durable transcripts.
Summary
Compares hosted capture, native sharing, analytics, telemetry pipelines, and orchestrator-owned conversation history. Verification cutoff: 2026-07-03 (tier B/C unless noted). This chapter covers commercial capture products, agent-native sharing, analytics dashboards, and telemetry pipelines; local OSS exporters are covered separately. SpecStory Cloud is covered in 01.
Question and scope
Can hosted products, native sharing, telemetry, or orchestrator history satisfy the archive requirement?
Method
The comparison reviews product documentation, public repositories, and capture mechanisms at the verification cutoff.
Findings
Commercial capture products
Mantra — the most direct SpecStory alternative
Local-first desktop "time machine" that auto-captures Claude Code, Cursor, and Gemini CLI sessions by reading their native stores, then maps every conversation turn to git history with automatic per-turn snapshots: scrub a timeline of user messages / AI responses / commits, diff or restore code state at any message. Full-text search, secret detection, optional web publish. Free, no sign-up, but closed source (releases-only repo, 8★, no license; release 2026-07-01). No Codex/Copilot coverage. Sources: mantra.gonewx.com, mantra-releases, docs: vs-alternatives.
The turn↔git-state correlation is the same problem SpecStory's unreleased provenance engine attacks (±5 s window matching, 02) — two independent products converging on "conversation turns joined to code changes" as the next frontier past raw capture.
LeanMCP AI Gateway — capture by proxy
Commercial SaaS: point ANTHROPIC_BASE_URL at their gateway and every request/file-read/token is logged server-side into a hosted dashboard. Sees wire-level traffic including system prompts — at the price of routing all code and prompts through a third party. The associated 2,103★ repo is an awesome-list funnel, not the product. Sources: docs.leanmcp.com, github.com/Leanmcp/superview.sh.
Sharing and persistence built into the agents
The vendors are absorbing capture natively — the baseline any third-party feature must exceed:
| Agent | Native story | Sources |
|---|---|---|
| Amp | The strongest: threads are server-side objects with URLs (ampcode.com/threads/T-…), workspace-shared by default, agent-readable as context (@T-<id>). Public discoverable sharing retired 2026-06-02 — now unlisted-URL or workspace/enterprise visibility only | end-of-public-threads, manual |
| opencode | /share mints a public link per conversation (opencode.ai/s/…); modes manual/auto/disabled; /unshare deletes | opencode.ai/docs/share |
| Codex | cloud task history (codex cloud list) and cross-surface session sync CLI↔Desktop↔Cloud (2026); still no transcript export or share links | CLI reference, cross-surface sync |
| Claude Code | /export → clipboard/plain text only; web sessions + /remote-control; OTel built in (off by default). No markdown archive, no share links | sessions docs, observability |
| Cursor | per-conversation "Export Chat" to markdown; no bulk export, no sharing | forum guide |
| Copilot (VS Code) | Chat: Export Chat... → JSON; copy-as-markdown; known bug dropping code blocks | chat-sessions docs, bug |
| aider | oldest native archive in the space: auto-writes .aider.chat.history.md markdown transcript | aider FAQ |
| Cline | per-task export button → markdown | discussion |
Usage analytics dashboards (transcript-adjacent, content-light)
| Tool | What | Signals (2026-07-03) |
|---|---|---|
| ccusage | token/cost reports from Claude Code JSONL (npx ccusage); feeds leaderboards | 16,791★, pushed 2026-07-03 — category giant |
claude-code-templates --analytics | real-time session dashboard inside a huge config/template ecosystem | 28,426★, pushed 2026-07-03 |
| Claude Code Usage Monitor | live TUI quota babysitter | 8,352★, pushed 2026-06-27 |
| sniffly | usage + error-pattern dashboards, shareable | 1,242★, pushed 2025-08-08 — dormant |
| vibe-log | Claude Code + Cursor productivity reports/coaching | 335★, 2026-04-19 |
| viberank | public usage leaderboard (Claude Code, Codex, Gemini) | 105★, 2026-06-28 |
| tokenhabit | offline token-waste linting of logs | 13★, 2026-06-15 |
These read the same native stores but extract numbers, not conversations — proof the stores support multiple derived products, and none of them is an archive.
Telemetry pipelines (vector C in Déjà Vu terms)
- Grafana Cloud Claude Code integration (first-party): Claude Code OTel → Grafana OTLP endpoint, official dashboard ID 25052, org-deployable. Metrics/logs, no conversation content. Docs, dashboard.
- Langfuse coding-agent ingestion: three routes — Claude Code plugin, Stop-hook → trace converter (this one does capture full conversations), or OTLP env vars. Caveat: Claude Code emits OTel logs, not traces, so bridges are needed (discussion #9088). Integration docs.
- Pydantic Logfire plugin (official repo, 12★, pushed 2026-07-02) and claude_telemetry (
claudiawrapper → Logfire/Sentry/Honeycomb/Datadog, 26★). - Official Grafana path is the supported Claude Code OTel integration; ColeMurray/claude-code-otel is an older community reference (454★, pushed 2025-06-17).
- Hook fan-out: claude-code-hooks-multi-agent-observability (disler, 1,474★, pushed 2026-02-08) — hooks POST every event to a local server + live dashboard; the reference implementation of vector B liveness.
Orchestrators whose captive history is not an archive
Every agent orchestrator persists conversations for its own workflow, none as an exportable cross-agent archive — directly relevant since jackin❯ sits in this category:
| Tool | Category | Session story | Signals |
|---|---|---|---|
| vibe-kanban (BloopAI) | kanban orchestrator (Claude Code, Codex, Gemini, Amp, opencode) | per-task-attempt logs in its own DB, for review only | 27,248★, pushed 2026-04-24 |
| Happy | mobile/web client for Claude Code + Codex | E2E-encrypted live session sync to relay; continuity, not archive | 22,370★, pushed 2026-07-03 |
| opcode (ex-Claudia) | GUI over ~/.claude/projects + checkpoints | session browsing as side effect | 22,133★ but pushed 2025-10-16 — stalled |
| Crystal/Nimbalyst | parallel worktree runner (Claude Code, Codex) | per-worktree conversation + diffs, in-app | 3,095★, 2026-02-26 |
| Conductor (Melty Labs) | macOS parallel workspace orchestrator | in-app chats/diffs, closed, no export | active product |
| claude-squad | tmux/worktree manager | no capture at all — panes only | 8,005★, 2026-06-17 |
| omnara (YC S25), cui | remote session relays | both archived (2026-01/03) | dead |
Category read-out
- Nobody ships the cross-agent, host-side, purge-surviving archive Déjà Vu targets. Mantra is closest in spirit but closed, desktop-bound, 3-agent. Orchestrators keep history captive; analytics tools drop content; telemetry pipelines drop content by design (OTel never exports thinking; GenAI semconv content fields are opt-in).
- Native vendor sharing is rising (Amp threads, opencode share, Codex sync) — but it is per-vendor, cloud-tied, and inconsistent; Amp even withdrew public sharing. Cross-agent normalization remains unclaimed ground.
- The provenance frontier is live: Mantra's git time-travel and SpecStory's provenance engine both join turns to code state — the capability layer above capture, and one jackin❯ is structurally positioned for (it owns the container, the workspace mounts, and the git materialization).
Implications for jackin❯
Separate the transcript archive from optional sharing, analytics, telemetry, and orchestrator-history layers.
Limitations and unknowns
Closed products expose limited implementation evidence, and hosted features, pricing, and retention policies can change.