LoopPrism

See what your coding agents actually do.

An inspector for Claude Code, Codex, Gemini CLI, OpenCode, DeepSeek Harness and WorkBuddy: every session, step by step.
What filled the context, where the time went, what it cost. Local-first, zero setup.

English · 简体中文

Download · Features · Get started

Chat — the conversation, with every tool call expandable	Trajectory — every step on a timeline, loops and failure streaks flagged
Context — what filled the context, where the time and the money went	Agents — subagents branching off and merging back, with what was passed between them
Cost — everything your agents spent, across every app, in one place

_{Made-up demo sessions, generated by scripts/make_demo_sessions.py and tests/support_agents.py.}

Features

🗂 Sessions Every Claude Code, Codex, Gemini CLI, OpenCode, DeepSeek Harness (dsh) and WorkBuddy session on your machine. Search all messages, group by day or project, export to Markdown. Claude Code and Codex sessions can also be resumed in a terminal or moved to the Trash from here; the other others are read-only.	🧭 Trajectory Every step on a timeline: prompt, injected context, thinking, tool call, result. With durations, search, and flagged patterns: the same call again and again, runs of failures.
🧮 Context What filled the model's input at each step: system prompt, tool schemas, your messages, injected context, skills, tool results. Window fill, trend, cache hit rate.	💸 Cost Real token usage priced at API list prices, split by token type, by what filled the context, by model and by turn. The Cost page adds up every session of every app: per day, model, agent and project, plus the most expensive sessions.
⏱ Timing Thinking vs answering vs tool time, and tokens per second.	🔌 MCP tracing & live runs An optional proxy records every MCP tool call; hooks record what an agent does as it happens (the run viewer is linked from the Cost page).
⚡ Optimize Finds where you spend more than you need to (low cache hits, a big model on a simple task, repeated calls, huge tool outputs, endless sessions), estimates the saving of each, and shows the total as a percentage of your spend. Mark a suggestion as adopted and LoopPrism compares your cost per call before and after. It also points to open-source tools that help, found on GitHub. Optional AI analysis uses a model and API key of your own choice.
🤖 Agents A session that spawns subagents gets a flow view: the main agent as a line, each subagent branching off and merging back, with what it was asked, what it returned, its cost, and how much ran in parallel. Open any subagent's own conversation. (Claude Code; Codex agent trees are not read yet.)
🎨 Look & language Dark, light or follow-the-system theme, and an interface in English or 简体中文 (or follow the system). Both are chosen in Settings and apply to the whole app, including the dashboard.

Supported apps and models

Claude Code	Codex	Gemini CLI	Cursor
OpenCode	DeepSeek Harness	WorkBuddy	More apps coming

LoopPrism runs on macOS and Windows with the same features on both: the desktop app, the lprism CLI, the hooks and the dashboard behave identically (only the system's own wording differs, such as Trash and Recycle Bin).

App	Sessions (read from disk)	Live runs (hooks)	Resume / delete here
Claude Code	~/.claude/projects (also subagent trees)	yes, plugin	yes
Codex	~/.codex/sessions	yes, ~/.codex/hooks.json	yes
Gemini CLI	~/.gemini/tmp/<project>/chats (JSON and JSONL), $GEMINI_CLI_HOME	yes, ~/.gemini/settings.json	read-only
Cursor	Cursor/User/globalStorage/state.vscdb (no cost: Cursor records no token usage)	yes, ~/.cursor/hooks.json	read-only
OpenCode	~/.local/share/opencode/opencode.db	yes, a plugin in ~/.config/opencode/plugins	read-only
DeepSeek Harness (dsh)	~/.dsh/sessions (zstd), $DSH_HOME, and DSH Desktop's own dsh-desktop/harness/sessions in its app-data folder	yes, via its hook plugin (one manual step)	read-only
WorkBuddy	~/.workbuddy/projects, $WORKBUDDY_CONFIG_DIR	yes, ~/.workbuddy/settings.json	read-only

Cost is priced from each model's published pay-as-you-go API price (cache reads and writes priced separately). Built in: Claude (Fable, Opus, Sonnet, Haiku), OpenAI (GPT-6, GPT-5 and Codex models), DeepSeek (chat, reasoner, V3.x, R1 and later versions), MiniMax (M2 and later), GLM (4.5 to 5.x), Qwen (Max, Plus, Flash, Coder), Kimi (K2), Gemini (2.5, 3), and free tiers such as OpenCode Zen's -free models. A newer version of a family is priced like the latest known one; OpenCode's own reported cost is used when we have no price, and models with no price at all (for example Tencent's hy3) show "no price" and are left out of totals. Set or correct any price in ~/.loopprism/pricing.json. Costs are what the work would cost on the API: subscriptions and free tiers bill differently.

Everything stays on your machine: SQLite on disk, no Docker, no cloud, no account.

Get started

Desktop app (macOS and Windows) — recommended

macOS

1. Download the .dmg from the latest release (arm64 for Apple Silicon, x64 for Intel) and drag LoopPrism into Applications.
2. Open it. First launch: right-click the app → Open (it is not notarized yet), or xattr -dr com.apple.quarantine /Applications/LoopPrism.app.

Windows (x64)

1. Download LoopPrism-<version>-x64.exe from the latest release and run it. The installer lets you choose the install folder and whether to install for just you (default, no administrator prompt) or for everyone on the PC.
2. It is not code-signed yet, so SmartScreen may warn: click More info → Run anyway.

That's it. The app bundles everything (no Python, no terminal), opens on your sessions, connects Claude Code and Codex for live runs, and keeps running from the menu bar (macOS) or system tray (Windows). Details: desktop/README.md.

Command line

pip install "git+https://github.com/atlasmarkz/LoopPrism.git"   # or: uv tool install git+https://github.com/atlasmarkz/LoopPrism.git
lprism proxy                                                    # then open http://localhost:8080

Record MCP tool calls through the proxy

lprism proxy (and the desktop app) runs an MCP proxy on port 4317. To record an HTTP MCP server's tool calls, give your MCP client the proxy's address instead of the server's:

http://127.0.0.1:4317/<name>?backend=<server url>

• <name> only labels the server in the dashboard; it is not sent to the server.
• <server url> is the server's full endpoint, path included, for example http://localhost:3000/mcp. URL-encode it if it contains ? or &. Anything after /<name>/ and any other query parameters are passed on to the server, and so are your headers (such as Authorization).
• A client that can set headers can use http://127.0.0.1:4317/<name> with the header X-LoopPrism-Backend: <server url> instead.
• It works with Streamable HTTP servers, including answers streamed as server-sent events. It does not proxy stdio servers or the older HTTP+SSE transport: there the server tells the client to post its messages to an absolute path (/messages/…), which bypasses /<name> on the proxy.
• The proxy and the dashboard only answer requests from this machine; a web page from another site gets 403.

For example, in Claude Code:

claude mcp add --transport http myserver "http://127.0.0.1:4317/myserver?backend=http://localhost:3000/mcp"

Upgrading from an earlier version: the <name> segment used to be appended to the backend URL, so …/mcp?backend=http://localhost:3000 reached http://localhost:3000/mcp. It is now a label only. Put the full endpoint in backend (…/myserver?backend=http://localhost:3000/mcp).

Live runs from Claude Code, Codex and WorkBuddy

Sessions are read after the fact. To also watch an agent as it works:

npx loopprism install     # adds the hooks to Claude Code, Codex, Gemini CLI, Cursor, WorkBuddy, OpenCode and/or DeepSeek Harness, whichever you have
lpa status                # newest runs

Claude Code can also use the plugin directly (/plugin marketplace add atlasmarkz/LoopPrism, then /plugin install loopprism@loopprism). In Codex, run /hooks once to trust the hooks; restart WorkBuddy, Gemini CLI and OpenCode after installing (OpenCode gets a one-file plugin). DeepSeek Harness sessions are read from ~/.dsh (or $DSH_HOME) and from DSH Desktop's own data folder, and WorkBuddy's from ~/.workbuddy (or $WORKBUDDY_CONFIG_DIR), and OpenCode's from its database, with nothing to set up. For live runs, DeepSeek Harness needs one manual step: lpa install dsh writes ~/.loopprism/dsh-hooks.json; in DSH Desktop install the plugin dsh-hooks-claude-code, set its configPath to that file and restart. Privacy options and uninstall: npm/README.md.

License

MIT

Star History