showagent

<!-- mcp-name: io.github.aytzey/showagent --> <h1 align="center">showagent</h1> <p align="center"><b>showagent</b> — every AI coding session on your machine, in one TUI.<br> Browse, search, resume, branch — and <em>convert</em> a conversation from one agent to another.<br> Codex · Claude Code · Gemini CLI · OpenCode · jcode</p> <p align="center"> <a href="https://github.com/aytzey/showagent/actions/workflows/ci.yml"><img src="https://github.com/aytzey/showagent/actions/workflows/ci.yml/badge.svg" alt="CI"></a> <a href="https://github.com/aytzey/showagent/releases/latest"><img src="https://img.shields.io/github/v/release/aytzey/showagent" alt="Release"></a> <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue.svg" alt="MIT"></a> </p>  > Started debugging in Codex and want Claude's take? Press `x`. Your > conversation moves with you — every user and assistant turn, rewritten in > the target agent's native session format. ## Why You use more than one coding agent now — most of us do. But every agent buries its sessions in its own format under its own dot-directory, and yesterday's context is trapped in whichever tool you happened to start it in. showagent reads the session stores straight off disk and gives you one searchable picker for all of them. It is the only TUI that combines browse + search + resume + branch + convert across agents. - **One list for everything** — sessions from every agent, grouped by workspace, fuzzy-searchable, newest first. - **Resume or branch anywhere** — reopen a session in its own CLI, or fork a local copy to try a different direction. - **Convert between agents** — rewrite a session into another agent's native format so that agent's own resume just works. Originals are never modified; conversions are written atomically. - **100% local** — one static binary that reads your own files. No hosted service, no telemetry, no account. ## Supported agents | Agent | CLI | Sessions read from | Env override | Convert from | Convert to | |---|---|---|---|:---:|:---:| | Codex | `codex` | `~/.codex/sessions/**/*.jsonl` | `CODEX_HOME` | ✅ | ✅ | | Claude Code | `claude` | `~/.claude/projects/**/*.jsonl` | `CLAUDE_HOME` | ✅ | ✅ | | Gemini CLI | `gemini` | `~/.gemini/tmp/<project>/chats/` | `GEMINI_CLI_HOME` | ✅ | ✅ | | OpenCode | `opencode` | `opencode.db`, via the `opencode` CLI | `OPENCODE_DATA_HOME` | ✅ | ✅ | | jcode | `jcode` | `~/.jcode/sessions/*.json` | `JCODE_HOME` | ✅ | ✅ | Notes: - OpenCode stores sessions in a SQLite database, so every OpenCode operation (discover, export, import, delete) goes through your own `opencode` CLI — showagent never writes into the database directly. OpenCode and jcode only appear when their CLI is installed. - The picker only offers hand-off targets whose CLI is on `PATH`, so it never strands a conversion. Scripted conversion to file-backed agents can still prepare a session before their CLI is installed; OpenCode always requires its CLI because imports go through OpenCode itself. - jcode is a niche, experimental agent CLI. Its support is auto-hidden: if no `jcode` binary is on `PATH`, showagent never shows it. - Platforms: Linux and macOS (amd64 + arm64). Windows (amd64) builds are released but **experimental**: resume runs the agent as a child process instead of replacing showagent. ## Install ```sh # Homebrew (Linux/macOS) brew install aytzey/tap/showagent # install script (Linux/macOS, puts the binary in ~/.local/bin) curl -fsSL https://raw.githubusercontent.com/aytzey/showagent/main/scripts/install.sh | sh # Go 1.25.12+ go install github.com/aytzey/showagent/cmd/showagent@latest ``` Or grab an archive from the [releases page](https://github.com/aytzey/showagent/releases/latest) — `linux`/`darwin` amd64 + arm64, `windows` amd64 (experimental). ## Quick start ```sh showagent # open the interactive picker showagent list # plain table of every session showagent list --json # the same, machine-readable showagent resume latest # reopen the most recent session, any agent showagent convert latest --to claude --dry-run # preview exactly what a hand-off would carry/drop showagent info latest # exact resume command + storage location showagent mcp # serve session history to MCP-capable agents (stdio) showagent mcp --read-only # same search/transcript tools, without tools that write copies showagent mcp --allow-secrets # explicitly allow verbatim secret-like transcript values showagent update # update a standalone install (Homebrew: brew upgrade aytzey/tap/showagent) showagent --help # full CLI help ``` ### Keybindings | Key | Action | |---|---| | `↑/k`, `↓/j`, `pgup/pgdn` | Move through sessions | | `/` | Fuzzy search across agent, workspace, session id, and messages | | `enter` | Resume the selected session in its own CLI | | `1`..`9` | Toggle provider visibility, numbered as listed in the header bar | | `p` | Cycle the preview column: first → latest → first + latest message | | `space` | Collapse or expand the selected workspace group | | `o` | Cycle the convert target for the selected session | | `t` | Cycle the convert scope: all turns, or latest 200/100/50/20/10 | | `x` | Preview convert; press `x` again to write and select the new session | | `n` | Branch: create a full local copy of the session | | `y` | Toggle yolo resume (skip the agent's permission prompts) | | `C` | Compound: resume with a learnings-capture prompt (see below) | | `d`, `del`, `backspace` | Delete the session — second press confirms, moving disarms | | `r` | Rescan session stores (keeps cursor, search, and filters) | | `?` | Toggle the full keybinding overlay | | `esc` | Clear search / close overlay / cancel an armed delete (never quits) | | `q`, `ctrl+c` | Quit | ## Scripting `showagent list --json` emits an array sorted newest-first — the field names are a stable contract: ```json [ { "id": "1f7c9a2e-4b31-4c8e-9d02-8a5e3f6b1c44", "provider": "codex", "workspace": "/home/you/code/api-server", "updated": "2026-07-08T19:51:25Z", "first_message": "Add rate limiting to POST /v1/charges", "last_message": "the redis TTL test is flaky - mock the clock" } ] ``` `showagent resume <id|latest> [--yolo]` resumes without the picker, so a shell alias can reopen your last session in one keystroke. `showagent convert <id|latest> --to <provider> --dry-run` prints the hand-off before writing anything: source session, target provider, workspace, scope, transferable turn count, last user ask, and the agent-specific state that will be dropped. Remove `--dry-run` to write the converted session, then showagent prints the resume recipe for the new row. `showagent info <id|latest> [--yolo]` prints the exact resume command, working directory, and storage location for a session. Exit codes: `0` success, `1` error (including "no sessions found"), `2` usage. When stdout is not a terminal, plain `showagent` prints the `list` table, so pipes just work. ## Use it from inside your agent (MCP) `showagent mcp` runs a stdio MCP server, so the agent you are talking to can search every past coding session on your machine — from **any** agent — and pull one in as context or convert it to continue right there. Ask Claude Code "have I solved this rate-limit bug before?" and it can find the Codex session where you did, read the transcript, and hand you the command to resume it — or rewrite it as a native Claude Code session and keep going. Your session history stops being per-tool memory and becomes shared memory. ```sh # Claude Code claude mcp add showagent -- showagent mcp # Codex (~/.codex/config.toml) [mcp_servers.showagent] command = "showagent" args = ["mcp"] ``` An MCP client can send returned transcript text to its model provider. Common secret-like values are therefore redacted by default, and every transcript is marked as untrusted historical data rather than instructions. If verbatim values are required, the user must explicitly start the server as `showagent mcp --allow-secrets`; an MCP tool call cannot bypass redaction. For clients that should never write session copies, register `showagent mcp --read-only`; that mode omits `branch_session` and `convert_session` entirely. Tools: | Tool | What it does | |---|---| | `list_sessions` | Search sessions across all agents — filter by provider, workspace substring, or free text over workspace + first/last user message (default 25, max 100 results) | | `get_transcript` | Read recent user/assistant turns (default 50, hard max 500); secrets are redacted unless the server was explicitly started with `--allow-secrets` | | `branch_session` | Fork a full local copy of a session, same agent; returns the new id, file, and resume command | | `convert_session` | Rewrite a session into another agent's native format; returns the new id, file, and resume command | | `resume_command` | The exact shell command (and cwd) that resumes a session — returned as a string, **never executed** | The default MCP surface is deliberately non-destructive: there is **no delete tool**, and it never launches or resumes an interactive agent. OpenCode storage operations still go through the local `opencode` CLI because its sessions live in SQLite. Deleting sessions stays exclusive to the TUI, where it takes two key presses with a human watching. Branch and convert only add new sessions — originals are never modified — and `--read-only` removes even those additive tools. ## How it compares Great tools exist for *running* agents in parallel — showagent is about the sessions they leave behind. [claude-squad](https://github.com/smtg-ai/claude-squad) and [ccmanager](https://github.com/kbwo/ccmanager) orchestrate multiple live agents in tmux sessions and git worktrees, which is the right choice when you want several agents working at once. [Agent Sessions](https://github.com/jazzyalex/agent-sessions) is a polished macOS app for browsing session history across many agents. showagent is the history-first, terminal-first take: a single cross-platform binary that reads the session stores on disk, resumes and branches from them — and is the only one of the group that converts a session from one agent's format to another's. ## Compound engineering Press `C` on a session and pick an agent. showagent resumes the session there and starts it on a compound-engineering pass: review what was solved, then record the durable learnings as markdown. Learnings are pooled per project but shared across agents: each workspace gets a directory under `~/.showagent/learnings/<project>/` (override with `SHOWAGENT_LEARNINGS_DIR`) that every agent reads and writes. Picking an agent that did not create the session converts it first, so it has full context. `showagent setup` installs the companion [compound-engineering plugin](https://github.com/EveryInc/compound-engineering-plugin) into the Codex and Claude Code CLIs found on the machine. It is idempotent and only installs what is missing. ## FAQ **Is my session data sent anywhere?** showagent itself does not upload session content and has no telemetry or account. An MCP client may send `get_transcript` results to that client's model provider, so MCP transcripts redact common secrets by default; keep that boundary in mind before registering the server. showagent's own HTTP client is used only by the optional release updater and startup update check (disable with `SHOWAGENT_NO_UPDATE_CHECK=1`). `showagent setup` invokes the installed Codex/Claude CLIs, which may download the requested plugin. Message previews also redact password-like strings and API keys before rendering (covered by tests in [`internal/session/session_test.go`](internal/session/session_test.go)). Release archives ship with a `SHA256SUMS` file, and releases after v0.7.0 also carry GitHub build provenance — verify with `gh attestation verify <file> --repo aytzey/showagent`. **How does conversion work?** Conversion extracts the user and assistant turns from the source transcript and writes a brand-new session in the target agent's native format (for OpenCode, via `opencode import`), so the target's own resume command picks it up. Code blocks, newlines, and indentation are preserved. The original session is never modified, and files are private (`0600`) and written atomically — a crash cannot leave a half-written session in another tool's store. Trust is explicit: in the TUI, the first `x` shows the hand-off preview and the second `x` writes it. In scripts, use `showagent convert ... --dry-run` for the same preview. Conversion intentionally does **not** copy tool-call internals, approval history, encrypted reasoning blobs, or provider attachments: those are private to the source agent and would not replay correctly anyway. `t` / `--scope` trims the scope to the latest N turns before converting. **What does delete actually do?** Codex sessions are deleted through `codex delete --force`; OpenCode through `opencode session delete` (which cascades inside its database). Claude Code removes the JSONL plus its matching index entry, jcode removes the JSON plus backup/journal sidecars, and Gemini removes its session file. Delete always takes two presses, and moving the cursor disarms it. **Windows?** Binaries are released and the whole TUI works, but resume semantics are approximated (child process instead of exec), so Windows is labeled experimental until it has seen real use. **A session is missing from the list.** Run `showagent list` with no sessions found and it prints exactly which directories were scanned and which env vars override them. `r` rescans in-place after you start a new conversation. ## Adding a provider A provider is one self-contained file implementing the 8-method interface in [`internal/session/provider.go`](internal/session/provider.go) — around 250 lines including discovery, resume arguments, transcript extraction, and conversion. [`gemini.go`](internal/session/gemini.go) (file-based store) and [`opencode.go`](internal/session/opencode.go) (CLI-based store) are the two templates. Register it in the `registry` slice and the TUI picks up badges, filter keys, and convert targets automatically. Add a matching env override so its tests stay hermetic. Issues and PRs welcome. ## Building ```sh git clone https://github.com/aytzey/showagent.git cd showagent go test ./... go build -o showagent ./cmd/showagent ``` The minimum supported toolchain is Go 1.25.12; CI also runs race tests, golangci-lint, `govulncheck`, and every published cross-compile target. The demo GIF is recorded hermetically with [vhs](https://github.com/charmbracelet/vhs) against fabricated fixtures — see [`demo/README.md`](demo/README.md). Security issues and sensitive-data exposure should be reported privately; see [`SECURITY.md`](SECURITY.md). Contributions are covered by [`CONTRIBUTING.md`](CONTRIBUTING.md). ## License MIT. Built with [Bubble Tea](https://github.com/charmbracelet/bubbletea), [Bubbles](https://github.com/charmbracelet/bubbles), and [Lip Gloss](https://github.com/charmbracelet/lipgloss).