Eridian unifies your Claude Code and OpenCode sessions into a single desktop app: live session timelines, subagent trees, per-session change inspection (files, diffs, commands, risk tags), full-text search across all history, a built-in syntax-highlighted file viewer with git time-travel, token/cost & context-fill rollups, and a durable local archive that survives the agents' own transcript purges.
It is strictly read-only against your agent data and everything stays on your machine — no telemetry, no cloud, no network calls except to a local OpenCode server you start yourself (and, only if you opt in, GET-only fetches of public skill/MCP catalogs — see Privacy & safety).
Agent sessions are scattered across ~/.claude/projects/*.jsonl, an OpenCode server, and
opencode.db, in formats that change between releases and get pruned over time. When you
want to review what an agent actually did — which files it changed, what commands it ran,
which subagents it spawned, how much context it burned — there's no single place to look.
Eridian is that place. It ingests everything into a local SQLite index (its own,
rebuildable, 0600-permissioned database), normalizes the differing agent formats behind
one event model, and gives you a fast, searchable, review-oriented UI over all of it.
- Unified sessions — Claude Code ingested live via a filesystem watcher + byte-offset
tail; OpenCode via REST bootstrap + SSE, plus a read-only cold-import from
opencode.dbso history shows even when the server is down. Sessions are grouped by agent, with plugin-generated sessions (e.g. claude-mem) in their own section. - Timeline — per-kind cards (prompt / assistant / thinking / tool call + result /
system), Markdown + XML rendering, kind-filter chips, inline image previews, token &
context-fill badges, and one-click copy of the session id or a
--resumecommand. - Change inspection — per session, split into Subagents / Commands / Files tabs: GitHub-style diffs, full-content previews with syntax highlighting, CLI commands with heuristic risk tags (danger / notable / safe, clearly labeled as heuristic), a subagent activity graph, and a session activity graph. Large sessions paginate.
- Skill & command detection — surfaces which skills (the
Skilltool) and slash commands ran in a session. - File viewer with time-travel — open any file, skill, or config in an overlay: syntax highlighting + line numbers, Markdown source/rendered toggle, image rendering, copy-path, and a git history picker to view the file as of any past commit.
- Full-text search — FTS5 across every agent's history; jump straight to the event.
- Token usage — a per-day usage chart (input incl. cache vs output) over 7/30/90 days.
- MCP & Skills panels — read-only, parsed from on-disk config with secrets masked; click a row to open the source file at the relevant block.
- Server control — start / stop / force-kill an Eridian-managed
opencode servewith live stdout, from the app. - Durable archive — keeps sessions after the agent purges its source transcript, with an "archived — source purged" badge; configurable retention and rebuild-from-disk.
- Resizable, persistent layout — drag-resize the sidebar and session list; pin sessions to the top.
- Automatic updates — signed in-app updates; new versions install with one click.
- Review an agent's work before you trust it — every file it changed, command it ran, and subagent it spawned, with diffs.
- Audit for risky actions — surface destructive commands and sensitive-file writes across a long session.
- Find that thing again — search all your past agent sessions and jump to the moment.
- Keep a durable record — retain history after the agents purge their own transcripts.
- Watch cost & context — see per-session token/context fill and daily usage.
| Agent | Live ingest | Offline history |
|---|---|---|
| Claude Code | ✅ filesystem watcher + tail | ✅ (reads ~/.claude/projects) |
| OpenCode | ✅ REST + SSE (opencode serve) |
✅ cold-import from opencode.db |
Codex, Gemini CLI, and read-only panels for other agents are on the roadmap.
These are hard guarantees, enforced in code (see CLAUDE.md):
- Read-only against agent data. Eridian never writes, renames, or truncates anything
under
~/.claude/,~/.config/opencode/, or~/.local/share/opencode/. Files are opened read-only; the OpenCode DB is openedSQLITE_OPEN_READ_ONLY. The only write action is starting/stopping an OpenCode server you explicitly launch from the app. - Local only. No telemetry, no update checks, nothing phones home. The only network
destination is
http://localhost:<opencode_port>. Optional catalog browsing: when you enable it in Settings (default off), Eridian may make GET-only requests to a compiled-in allowlist —registry.modelcontextprotocol.io,api.github.com,raw.githubusercontent.com— to download public catalog metadata for the Skills and MCP “Discover” tabs. Nothing is uploaded; responses are cached locally. - Your data stays private. Transcript bodies are never logged; Eridian's own database
is created
0600.
Download a build from the Releases
page — macOS (Apple silicon / Intel .dmg) and Linux (.AppImage / .deb) — or
build from source.
macOS: builds aren't Apple-notarized, so macOS quarantines the download and may say the app is "damaged and can't be opened." It isn't. Move Eridian.app to Applications, then run:
xattr -cr /Applications/Eridian.appand open it normally. (For the "damaged" message, right-click → Open does not work — use the
xattrcommand.)
Windows: builds aren't code-signed, so SmartScreen may warn about an "unrecognized app." Click More info → Run anyway to install.
- macOS, Linux, or Windows
- Rust (stable) and Tauri 2 system dependencies
- Node 20+ and pnpm
- Optional but recommended: Claude Code (transcripts in
~/.claude/projects/) and/or OpenCode
pnpm install
pnpm tauri dev # run the app in development
pnpm tauri build # produce a release bundlepnpm tauri dev # app (Vite :1420 + Rust)
pnpm test / pnpm test:cov # frontend unit tests (Vitest) + coverage
pnpm tsc --noEmit # frontend typecheck
cd src-tauri && cargo test # Rust tests
cargo clippy -- -D warnings # Rust lint
cargo cov / cargo cov-gate # Rust coverage (cargo-llvm-cov)Logic modules are covered to ≥90% on both sides (gate: pnpm test:cov, cargo cov-gate). Some tests read your real agent data and are #[ignore]d by default:
cd src-tauri
cargo test real_backfill -- --ignored # Claude Code backfill (large)
cargo test opencode_live_bootstrap -- --ignored # needs `opencode serve` running
cargo test cold_import_real -- --ignored # reads ~/.local/share/opencode/opencode.dbSee CONTRIBUTING.md for architecture, conventions, and how to extend the ingesters.
Each ingester normalizes its agent's format into one shared event model at the edge, then writes to a single-writer SQLite store (WAL, FTS5, atomic batches). The Rust core exposes a read-only Tauri command surface; the React UI reads only through those commands and receives live updates via events — it never touches agent files directly. Full map in CONTRIBUTING.md.
- Now: Claude Code + OpenCode, full change inspection, search, time-travel viewer, daily token usage.
- Next: packaged releases (signed builds + CI); per-commit diff in the viewer.
- More agents: Codex (JSONL rollouts), Gemini CLI, read-only panels for Antigravity / Claude Desktop.
- Later (opt-in, carefully): config assistance (MCP config sync, key management via the OS keychain) — only after the read-only foundation is rock-solid.
Ops-console, not SaaS dashboard: near-black slate base, monospace-first data typography,
and agent-colored identity accents only — Claude Code coral #E8825A, OpenCode green
#3ECF8E, subagent violet #A98CFF, one shared alert amber #E7B75F.
Contributions are welcome — see CONTRIBUTING.md. Please keep the read-only and local-only guarantees intact; they are the point of the project.
MIT © Kresna