11 — Open-Source Exporters and Viewers

Summary

Compares open-source session exporters and viewers by capture mechanism, coverage, output, license, and maintenance. Verification cutoff: 2026-07-03; stars and activity dates are GitHub API observations at that cutoff (tier B/C — repositories and readmes; source-audited only where noted). Records follow the research brief. Comparison baseline: SpecStory CLI, 1,268★, Apache-2.0.

Pattern across the whole category: every serious OSS tool reads the agents' native stores (vector A in Déjà Vu terms). None wraps a PTY; one (claude-trace) patches fetch to see wire traffic. The market has independently converged on SpecStory's mechanism.

Question and scope

Which open-source capture tools provide reusable mechanisms, roster coverage, and durable maintenance signals?

Method

The catalog reviews repository code, READMEs, licenses, release activity, and public discovery threads at the verification cutoff.

Findings

Multi-agent tools (closest to SpecStory scope)

claude-replay — replay as a portable artifact

• What/maker: session logs → self-contained interactive HTML replays; solo dev es617. MIT, free.
• Surfaces: Claude Code, Cursor, Codex, Gemini, OpenCode.
• Mechanism: reads native local stores (JSONL incl. tool calls, thinking, timestamps); read-on-demand, no daemon.
• Output: one static HTML file per session — step-through player, timeline, tool-call inspection; shareable by copying the file.
• Maintenance: 743★, pushed 2026-06-02; Show HN 2026-03-06, 105 points (HN).
• Differentiator: replay-as-artifact vs SpecStory's continuous markdown pipeline.
• Sources: github.com/es617/claude-replay.

coding_agent_session_search — the retrieval heavyweight

• What/maker: unified Rust TUI/CLI full-text search over local session history "across 11+ providers"; Jeffrey Emanuel (Dicklesworthstone). Free (non-standard license — GitHub reports NOASSERTION).
• Mechanism: indexes each agent's native store read-only; no capture step at all.
• Maintenance: 938★, pushed 2026-07-03 (same-day activity).
• Differentiator: broadest provider coverage in the compared evidence; solves the retrieval half of SpecStory's proposition without owning capture. SpecStory only reached indexed search in v2.0.0 (2026-06-29).
• Sources: github.com/Dicklesworthstone/coding_agent_session_search.

claude-code-tools / aichat search

• What/maker: agent-productivity toolkit whose aichat search builds a Rust/Tantivy full-text index over sessions of Claude Code, Codex, "and similar CLI agents"; Prasad Chalasani. MIT.
• Maintenance: 1,885★, pushed 2026-06-30.
• Sources: github.com/pchalasani/claude-code-tools.

agentlore — team aggregation (embryonic)

• What/maker: "team visibility layer" aggregating agent logs to a Go server with shareable permalinks; clkao. MIT. 11★, pushed 2026-03-19.
• Differentiator: the only OSS attempt at SpecStory-Cloud-style team sharing found; too early to lean on.
• Sources: github.com/clkao/agentlore.

Long tail (all under 10★, verified alive 2026-07-03)

recall (Apache-2.0, desktop browser), tessera, agents-session-viewer, chathistory (HTML/PDF export).

Claude Code-specific

Smaller: claude-notes (76★), claude-insight (75★), ccrecall (transcripts → SQLite, 22★, pushed 2026-07-02), claude-session-export, claude-code-log-analyzer (21★).

Cursor-specific

The graveyard rate here is the Cursor-schema-churn story from 01 seen from the OSS side: unfunded maintainers stop chasing state.vscdb changes.

Copilot- and Codex-specific

• Copilot cluster (native export is JSON-only, so small tools fill the gap): fengzehan's Copilot Chat Exporter and Copilot Chat to Markdown (VS Marketplace), github-copilot-chat-exporter (PowerShell, 5★, 2026-04-26), copilot-chat-export-formatter (dead, 2023). All read workspaceStorage/**/chatSessions/*.json or drive the built-in export command.
• Codex cluster (no dominant viewer exists): @nogataka/codex-viewer (npm v0.0.11, 2026-02-05, 35★), codex-session-viewer (2★), codex-history-list (10★), plus ~5 more ≤2★ repos. All parse ~/.codex/sessions/**/rollout-*.jsonl. The niche is unconsolidated — nobody owns Codex history tooling.

Category read-out

1. Mechanism monoculture: native-store reading won everywhere; wire interception (claude-trace) exists only for depth (system prompts/thinking), and PTY scraping appears nowhere.
2. Claude Code has a rich, healthy tool belt; every other agent's ecosystem is thin or stale. Cross-agent tools are new (2026) and retrieval-focused.
3. Sharing is the unsolved half: gist upload (simonw) and static HTML files (claude-replay) are the state of the OSS art; only embryonic agentlore attempts team-level aggregation.
4. Maintenance risk correlates with store stability: Cursor exporters die fastest; Claude Code JSONL readers live longest — matching Déjà Vu's "pair unstable formats with a vendor-blessed hedge" stance.

Implications for jackin❯

Borrow proven schemas, adapters, and export patterns, but own unsupported providers and the durable host-side archive contract.

Limitations and unknowns

Repository activity and star counts are weak maintenance proxies, and tools discovered after the cutoff are absent.

Sources

• claude-replay
• claude-replay Show HN

Related work

• Session-capture dossier
• Déjà Vu conversation-capture design