callimachus

Local index & search for your AI coding-agent threads across 13 tools (Claude Code, Codex, Cursor, Gemini CLI, Qwen Code, Goose, OpenCode, Continue, Cline, Roo Code, Kilo Code, Copilot Chat, Copilot CLI), plus a provider-agnostic chat, an MCP server, a CLI, and a VS Code / Cursor extension. Your history stays on your machine.

Named for Callimachus, who built the first catalogue of the Library of Alexandria.

Download

Grab the latest signed build from Releases: macOS (.dmg, Apple Silicon), Windows (.msi or -setup.exe), or Linux (.AppImage, .deb or .rpm). The app auto-updates from there on. Prefer to build it yourself? See Develop.

macOS: drag Callimachus into Applications and open it from there before enabling the agent integration. While it runs from the disk image or a quarantined download, Settings won't register it, because that path disappears when the app quits.

Windows uninstall: the -setup.exe uninstaller removes what the app added (the /recall skill, the Claude Code, Codex, Gemini CLI, GitHub Copilot and Cursor hooks, the MCP entries for Claude Code, Codex, Cursor and Gemini CLI, and cal.exe), but the .msi uninstaller doesn't. Before removing an MSI install, click Remove for Claude Code and under Other agents in Settings → Agents (the Other agents Remove also takes out the Codex, Gemini CLI, GitHub Copilot and Cursor memory hooks), or run callimachus.exe --uninstall-integration; otherwise the hooks keep pointing at the deleted app and each agent reports a hook error at every session start.

The same release also carries standalone cal-<target> and callimachus-mcp-<target> binaries for CLI/MCP-only use, with their SHA-256 sums in SHA256SUMS. Linux builds and binaries need glibc 2.38 or newer (Ubuntu 24.04+, Debian 13+, Fedora 39+); Ubuntu 22.04, Debian 12 and RHEL 9 are too old for the bundled ONNX Runtime.

Updates: the first launch after an update upgrades the index in the background; the splash keeps animating, and on a large index this can take tens of seconds. An older build can't open an upgraded index: an older app shows "This index is from a newer Callimachus", and an older cal, callimachus-mcp or pinned npx callimachus-mcp@<version> stops with an error. Whichever of them opens the index first does the upgrade, so keep the app, the npx launcher and any standalone binaries on the same release.

Related MCP server: Memory MCP

What it does

• Indexes every conversation from 13 coding agents into one local SQLite store: Claude Code, Codex, Cursor (the editor's chats, its agent transcripts and the Cursor CLI), Gemini CLI, Qwen Code, Goose, OpenCode, Continue, Cline, Roo Code, Kilo Code, Copilot Chat (the built-in chat panel of VS Code, Cursor, VSCodium and Windsurf), and the Copilot CLI. Adding another source is a small, documented contract.

• Searches them with hybrid ranking: keyword (SQLite FTS5 / BM25) fused with on-device semantic similarity (sqlite-vec KNN, no cloud) via Reciprocal Rank Fusion. Filter by source, project, subagents, starred, and tags.

• Finds code-aware: type file:embed/mod.rs in the search bar (or cal files <path>) to find every thread that touched a path; backed by a file-mention index built at index time.

• Distills knowledge: free heuristic TODO extraction (always on, no LLM: - [ ] tasks and TODO / FIXME markers, an agent's checklist only when it names a file, symbol or identifier, plus the work an agent listed under "Follow-ups", "Not done", "Deferred" or "Remaining" in a session's latest such list from the last 30 days), plus LLM distillation of decisions, gotchas, and summaries behind one consent click that names the engine and where transcripts go (the Turn on memory card in Projects or Chat's Ask mode preselects a detected Claude Code or Codex CLI, then local Ollama, then a saved key; Settings > AI engine has the same switch), with cross-thread semantic recall of past decisions/gotchas. Optional auto-distillation works through threads active in the last 30 days in the background, up to a daily limit you set next to its toggle (50 threads by default, 0 pauses it; Settings shows what's left today), so memory self-populates without an open-ended bill. Runs on local Ollama (keyless), your logged-in Claude Code or Codex CLI (keyless, billed to whatever that CLI is signed in with), or a cloud API key.

• Imports what agents already write: Claude Code's own auto memory (the topic files under ~/.claude/projects/<project>/memory/) joins each repo's project memory with no LLM: feedback memories become gotchas and project memories decisions, tagged "Claude Code memory", refreshed when a file changes and removed when it goes (Settings > Agents turns it off and removes them). The managed AGENTS.md / CLAUDE.md block and .callimachus/memory.md leave them out, since Claude Code already loads them and those files get committed. When auto-distillation is on, a SessionEnd hook moves the session that just ended to the front of the queue, so its memory lands minutes after it ends.

• Curates the facts: pin, edit, or hide distilled facts so your edits survive re-distilling, record a decision or gotcha yourself with Remember in the Projects tab (saved pinned), plus an LLM "Review conflicts" pass that flags decisions that contradict each other.

• Remembers per project: a Projects tab gives each repo one page: its recent sessions, its decisions, gotchas, and open TODOs aggregated into durable memory (grouped by a canonical project key, so worktrees / symlinks / ~ don't split one repo), its snapshots, and exactly what its agents are given (the SessionStart hook's text and the managed AGENTS.md block, with their sizes), plus an LLM brief and a managed .callimachus/memory.md. That memory is prepended when you "Open in CLI", and you can inject it into any agent automatically: Update AGENTS.md (or cal agents) writes a managed block into the repo's AGENTS.md / CLAUDE.md, and cal hook feeds a size-budgeted version to a Claude Code SessionStart hook, labeled as notes from past sessions rather than instructions. Every line is dated by when its session happened (not when it was distilled) and says where it came from, e.g. (2026-10-03, codex, thread 1234; replaces 2026-10-01) (the date only in the committed AGENTS.md block); a replaced decision is left out, and two that disagree with nothing settled are marked disputed. A repo with no memory yet still gets something on day one, with no LLM: a digest of its last few sessions (title, agent, branch, age, the latest summary or prompt, touched files, open TODOs, and the latest snapshot to resume), also appended under the facts when they leave room. The hook runs on startup, /clear and compaction (a shorter, facts-only refresher), stays quiet on a resumed session, and skips a repo where Claude Code already loads the managed block (a CLAUDE.md, CLAUDE.local.md or a file one @-imports, or AGENTS.md when no CLAUDE.md exists), so the memory is injected once. cal hook --explain shows what a new session would get, cal hook --mute turns the digest off for one repo, and Settings > Agents turns it off everywhere. Codex, Gemini CLI, GitHub Copilot and Cursor get the same memory through their own hooks (Settings > Agents > Other agents > Memory hooks): cal hook --for codex / --for gemini / --for copilot / --for cursor when a session starts (Codex and Gemini CLI also on /clear, Codex after compaction), wrapped in each agent's hook output format, plus their imported Claude Code memories, marked as such, since none of them loads Claude's auto memory. GitHub Copilot's hooks live in their own file, ~/.copilot/hooks/callimachus.json, which Copilot CLI and VS Code's Copilot sessions both read; Cursor's go into ~/.cursor/hooks.json next to yours, and its hook takes the repo from the workspace root. Cursor does not wait for its session-start hook, so the first chat after Cursor starts (while its hooks load) or a first prompt sent right away can miss the memory. The hook stays quiet in a repo whose AGENTS.md (Codex), root AGENTS.md or CLAUDE.md (Cursor), GEMINI.md (Gemini CLI) or any instruction file Copilot loads (AGENTS.md, CLAUDE.md, GEMINI.md, .github/copilot-instructions.md, ~/.copilot/copilot-instructions.md) already holds the managed block. Their SessionEnd hooks queue the ended session for auto-distillation like Claude Code's (for Copilot, Copilot CLI sessions only: VS Code Copilot sessions are not indexed from Copilot's session store yet), and a Codex PreCompact hook snapshots the session. Codex asks you to trust new hooks once (/hooks in Codex). Cursor also runs the hooks in ~/.claude/settings.json (its third-party import, on by default); the Claude Code ones stay quiet there, so a Cursor chat gets the memory once, from its own hook. Neither Copilot nor Cursor lets a prompt-submit hook add context, so proactive recall stays Claude Code only. The /recall skill is also copied to ~/.agents/skills/, the shared skills folder other agents read. Gemini CLI loads every context file it is configured with plus their @ imports, and Codex its fallback filenames too, so the hook stays silent wherever the managed block already lives.

• Memory that travels with the repo: Publish to repo on a project's Memory section (or cal memory publish) writes its pinned facts and the ones you or an agent recorded (with --all, or the switch in the preview, the distilled ones too) to .callimachus/memory/decisions.jsonl and gotchas.jsonl in the checkout, one JSON object per line sorted by a stable id, so branches merge line by line. The app shows the counts and every added, updated or removed fact before anything is written (cal memory publish writes, then lists the same changes), secrets are redacted, Claude Code's auto memory is never included, lines a teammate added are kept, and nothing is committed or published on its own. Everyone with Callimachus who works in the repo gets those facts as the app (or cal index) indexes a session there, or with cal memory import (--review lists the changes first): marked "From the repo", never pinned, included in the hook and the AGENTS.md block marked repo under a line naming the files, and with your own edit of one winning over the file (hiding one is local to you: publish never removes a teammate's line). A fact removed from the files is hidden for everyone on their next import (once no other clone of the repo on their machine still has it); the ones you published show as "Shared".

• Asks your history (RAG): the Ask mode in Chat gives a synthesized, cited answer over your own threads, with [thread N] citations that open the sources it used; it is saved with your chats and follow-ups continue as a normal chat. Needs an LLM engine (memory turned on in Settings > AI engine).

• Organizes into collections: star threads and attach free-form tags, then filter the list by starred or by tag.

• Chats with an in-app agent (Anthropic / OpenAI / Gemini / OpenRouter / Ollama, local or Cloud, or your logged-in Claude Code / Codex CLI with no key) that can search your own history and run shell commands with your approval; streaming, cancellable, with live model lists. Chats are saved and become searchable too.

• Carries context across tools: open any thread in any agent CLI ("Open in Claude / Codex / Gemini …", seeded with the packed transcript), resume a Claude Code / Codex thread in its native CLI, copy context, or export a thread to Obsidian (optionally AI-summarized with decisions / gotchas / TODOs).

• Links threads to commits: infers on-device which git commits a thread produced, by overlapping the files a thread discussed with git log's changed files inside the thread's time window (shared-file count = confidence). See it as a thread→commit timeline (cal commits), per-thread in the desktop UI ("Produced commits"), or via the linked_commits MCP tool. Run it backwards from code with cal why <file>:<line>: git blame names the commit behind the line, and you get the sessions that produced it (the ones that edited the file first) with their decisions and gotchas; the same answer comes from threads_for_file with a line, and from Why This Line in the VS Code / Cursor extension, which also marks the lines a session wrote in the gutter and explains them on hover. It speaks Agent Trace (an open RFC for AI code attribution) both ways: when a repo carries trace records in .agent-trace/, cal why and threads_for_file show the agent, model and conversation another tool recorded for a line, even one Callimachus has no transcript of, and cal trace export writes records for the commits your sessions produced. The background watcher links new sessions' commits as they are indexed, and records which files each session read, edited or wrote from its tool calls (sessions indexed before the update are filled in once, in the background).

• Snapshots agent sessions: durable, resumable checkpoints of a thread (packed transcript + carry-forward project memory) for handoff across a context-window compaction or across tools; auto-captured by a Claude Code PreCompact hook right before a context compaction, which also keeps Claude Code's compaction summary as the thread's summary. Take, list, and resume them (cal snapshot / cal snapshots / cal resume, or the snapshot_session / list_snapshots / load_snapshot MCP tools), or in the app, where a project's page lists its snapshots to read, delete, or resume in any agent CLI.

• Guards decisions: decisions can carry a rationale ("why"), and an active guard surfaces settled decisions on a topic before an agent re-litigates one (cal check "<proposal>" / the check_decision MCP tool), each with its date and status: a decision a later session reversed comes back labeled superseded, naming the decision that replaced it, so "go back to X" meets X's replacement first. A reversal is recorded as a link, never inferred from dates or similarity: the memory pass declares it from a session that plainly says so, you set it in the app, or an agent passes replaces to record_decision (cal remember --replaces); one aimed at a fact you pinned or edited, or an agent's past five in a project in a day, is shown as disputed for you to settle instead. Replaced decisions stay as history, and one Undo brings a fact back however many facts replaced it. Today's Memory to review lists what agents and memory passes did that nobody looked at yet (stale or wrong flags, open disputes, and what agents replaced or recorded since you last opened the app) with Keep, Undo and Hide; it shows only when there is something, and cal memory review prints the same list.

• Recalls before you redo: an opt-in Claude Code hook quietly injects "you may have solved this before" into the agent's context on each prompt when it strongly matches prior work (Settings → Agents → Proactive recall, off by default, signal-floored and per-session deduped).

• Mines recurring issues: surfaces the errors you keep hitting across all your tools (normalized signatures, last 180 days; a Claude Code subagent's tool output counts for the session that started it) so you fix the pattern, not the instance (on the Project page, or cal issues).

• Remembers how you fixed them: when a session that hit a recurring error then edited the file the error named and a commit linked to that session changed that file, that commit is recorded as the fix in that repo (the first such commit after the newest sighting that leads to one, kept while later sightings lead to none: it says where the error was last dealt with, not that it stopped). The Project page shows "Last fixed in <sha>" with a link to the session, cal issues and the recurring_issues MCP tool return the fix with the sessions behind the error, and the fix becomes a playbook gotcha in the repo's memory ("<error>: seen N times; fixed in <sha> (<subject>), see thread <id>"), refreshed in place, never pinned. A playbook is listed with the repo's gotchas on the Project page and in cal memory, cal issues and recurring_issues return it with the fix, and the opt-in hook below cites it; it takes no slot in the memory injected at session start (or the AGENTS.md block, or project_memory) unless you pin it. Generic errors (a bare error:, a CI exit code, a path that wasn't there) get no playbook and stay out of the lists (cal issues --all shows them); a Python traceback counts as its exception line, and file content read with a line-number gutter (the Read tool, cat -n), source lines that only name an error variable and a script's echo lines are not errors. Fixes need the files a session edited, which come from its tool calls; sessions indexed before the update are filled in once, in the background, with no reindex. An opt-in Claude Code hook (Settings → Agents → Error playbooks, off by default) adds that line to Claude's context when a shell command it runs fails with an error fixed before in the same repo, once per error per session; it runs after every shell command, and output without an error line costs a few milliseconds. The hook ships for Claude Code first; Codex and Gemini CLI do not get it yet.

• Shows what it did for you: a Stats card counts, per week, the sessions that started with injected memory or a recent-sessions digest, the prompts where recall surfaced prior work, and the MCP tool calls answered (cal stats prints the same as one line). The counts are appended to ledger.jsonl next to the index (<name>.ledger.jsonl for any other index file, such as the dev build's index.dev.db), and once a day is over the app moves that day's counts into the index itself; neither leaves the machine. The same card shows how many indexed Claude Code sessions are still searchable after Claude Code deleted their transcripts (its cleanupPeriodDays, 30 days by default), with a reversible "Keep transcripts for a year" that sets it to 365 in Claude Code's settings.json after you confirm.

• Shows what it cost: captures per-message token usage and model at index time (Claude Code, Codex, Gemini CLI, Qwen Code, OpenCode and Copilot), then estimates dollar spend by model, by session, by the commits those sessions produced, by project and by day. Each session gets short neutral notes on where its spend went (cache reads, turns past 50, subagents, a tool failing over and over). Figures are list-price equivalents from an embedded price table (apps/desktop/src-tauri/core/prices.json), so for a subscription they show what the same usage would cost at API rates; models without a flat published list price (open-weight models, Qwen's prompt-tiered models, a release not priced yet, Claude Code's placeholder turns) are listed as unpriced with the reason. See it in the Stats card (with the last 30 days and the priciest commits), on a Project page, or with cal cost --by thread|commit|project|day. Codex, Gemini CLI, Qwen Code and OpenCode sessions indexed before the update get their usage filled in on their own, once, in the background (by the app, or by cal index), with no reindex.

• Sums it up: "Your agents, so far" puts the whole index on one screen: sessions per agent and estimated hours (over the sessions with usable timestamps), active days, spend by agent and model (all time and the last 30 days) with the priciest commits, the most frequent recurring errors and the commits that fixed them, the tasks you asked again on another day (sessions whose first prompt matches, once the index has 50 sessions), project memory and how often a session started with it, the files agents changed most, and the value ledger. Open it from Today's card (after the first index) or from Stats, then copy it as markdown or save a 1200 x 630 PNG card (headline numbers, agents, the top error from tool output with paths shortened to file names, and its fix; no prompt text) to your Downloads folder. cal report prints the same markdown (--since DAYS narrows it, --json gives the object). Built from the index alone; nothing leaves the machine unless you share it.

• Feeds PR audits: cal audit-pr <repo> --changed-files … --shas … returns one JSON bundle (per-commit provenance back to the session that produced it, per-file prior threads + reasoning, repo recurring errors, project memory) so an external local PR-review tool can show the history behind a diff.

• Surfaces to your agents: a bundled MCP server (callimachus-mcp) exposes the index as tools any agent can call mid-session, and it's two-way: agents can write back into Callimachus's own memory (close TODOs, record decisions/gotchas, snapshot a session) without ever touching your files. The /recall skill teaches them when to use it.

• Stays current via a background file watcher, plus a catch-up pass at launch for the sessions your agents wrote while the app was closed; stays private: the index never leaves your machine, API keys live in the OS keychain, conversation content goes only to an AI provider you pick, and usage analytics are opt-in and content-free (see Security & privacy).

Stack

• Shell: Tauri 2 (Rust) + React 19 + TypeScript + Vite 8

• Store/search: bundled SQLite + FTS5 (rusqlite); on-device embeddings via fastembed (bge-small-en-v1.5, 384-dim); KNN in SQL via sqlite-vec (vec0)

• Watcher: notify + debouncer

• Chat: multi-provider via the genai crate (Anthropic / OpenAI / Gemini / OpenRouter / Ollama local + Ollama Cloud), each with a configurable base URL, streaming tokens over a Tauri Channel, cancellable, with agent tool-calls (history search + approved shell)

• Secrets: OS credential store via the cross-platform keyring crate: macOS Keychain, Windows Credential Manager, Linux Secret Service

• Entry points: the app binary also runs as cal (through the ~/.local/bin/cal link or cal.exe) and as the MCP server (--mcp); the standalone cal and callimachus-mcp release binaries are the same code built on their own. All of them run the Tauri-free core crate (callimachus-core, which the standalone binaries link without Tauri, GTK or WebKit) against the same index.db

• Editor: a VS Code / Cursor extension (apps/vscode, published to the Marketplace + Open VSX) that shells out to cal

Monorepo

This is a Turborepo + pnpm workspace.

apps/
  desktop/            # the Tauri 2 desktop app; src-tauri/core is the Tauri-free core (index,
                      # cal, MCP server, jobs), src-tauri/standalone the cal + callimachus-mcp bins
  vscode/             # VS Code / Cursor extension (search history from the editor)
  web/                # callimachus.app, the marketing + download site (TanStack Start)
packages/
  callimachus-mcp/    # the npm launcher behind `npx -y callimachus-mcp`
.changeset/           # version + changelog management
scripts/              # version-sync, release tagging

Releases, versioning, and the auto-updater are documented in RELEASING.md.

Develop

pnpm install
pnpm desktop:dev      # launches the desktop window (tauri dev) against a separate dev index

# from the repo root, across all apps:
pnpm build            # turbo: build every app's frontend
pnpm typecheck        # turbo: typecheck every app

First launch: the index is empty, so the Search screen offers Index my history (or use Reindex in Settings → Sources); then Build semantic index enables semantic search. Later launches also index whatever your agents wrote while the app was closed. Reindex runs as a background job with a per-source progress bar, separate from Build semantic index. The two are mutually exclusive (one pauses while the other holds the write lock).

Tests

cd apps/desktop/src-tauri
cargo test                                   # fast unit tests, every crate in the workspace
cargo test -p callimachus-core -- --ignored --nocapture   # real-data + model + keychain smoke tests

The --ignored tests touch live data on this machine: each source has a real_<source>_index smoke test that indexes your real history read-only (~/.claude, ~/.codex, Cursor, ~/.gemini, ~/.qwen, Goose, OpenCode, Continue, Cline/Roo/Kilo), plus the embedding-model download (first run, needs network) and a Keychain round-trip.

Use your history anywhere

Beyond the desktop window, the same local index is reachable from your agents, terminal, and editor, all reading one index.db.

Claude Code plugin: the one-step setup for Claude Code. It bundles the /recall skill (as /callimachus:recall), the MCP server (the callimachus-mcp launcher ships inside the plugin and runs with node, pinned to the plugin's release) and the SessionStart, PreCompact and SessionEnd memory hooks (plugins/claude-code):

claude plugin marketplace add BetaBots-LLC/callimachus
# then, inside Claude Code:
/plugin install callimachus@callimachus

The hooks run the cal CLI at ~/.local/bin/cal, so open the desktop app (which builds the index) and click Install the cal CLI in Settings → Agents once (the Claude Code button's label while the plugin is enabled): that installs only cal and removes the app's own skill, MCP entry and hooks, so memory is never injected twice and only one server starts; Settings then says the integration is managed by the plugin. Until cal exists the hooks do nothing. Without the desktop app (macOS and Linux), npx callimachus-mcp --install-cal downloads the release's cal (checked against the same SHA256SUMS as the MCP server) and links it at ~/.local/bin/cal; run cal index once to build the index, and from then on the plugin's SessionStart hook (cal index --if-stale) starts a catch-up pass in the background whenever the index is more than an hour old and neither the app nor cal watch keeps it fresh. You give up the app's windows (the Projects page, Chat, the harvest controls and Settings, where memory distillation is turned on); search, recall, the memory hooks and the MCP server work the same (plugins/claude-code). The plugin needs the app at 0.11.4 or later: the marketplace serves it from the latest release tag, with the MCP server pinned to that release, so keep the app on the latest release too. On Windows, use Settings → Agents instead of the plugin: Claude Code runs plugin hooks through Git Bash (or PowerShell, where they fail with a hook error), and the app keeps its own skill, MCP entry and hooks there, so the plugin would load them twice.

Cursor plugin: the same setup for Cursor (plugins/cursor): the recall skill, the MCP server (npx -y callimachus-mcp@<version>, pinned to the plugin's release) and the sessionStart and sessionEnd memory hooks, which run cal hook --for cursor and cal session-end --for cursor through the cal at ~/.local/bin/cal (the app 0.12.1 or later; the hooks stay silent until it is there). Install it from the Cursor Marketplace once it is listed there (Customize, search for Callimachus), or until then copy plugins/cursor into ~/.cursor/plugins/local/callimachus and reload the window. On cursor.directory, Add to Cursor adds only the MCP server. Alongside the app's own Cursor hooks (Settings → Agents) the memory still arrives once per chat: cal hook --for cursor claims each conversation with a marker under ~/.callimachus/recall. On Windows, use Settings → Agents instead: the plugin's hooks are sh scripts.

Codex plugin (experimental): the same setup for Codex (plugins/codex): the recall skill, the MCP server (npx -y callimachus-mcp@<version>, pinned to the plugin's release) and the SessionStart, PreCompact and SessionEnd memory hooks, which run cal hook --for codex, cal snapshot-hook --for codex and cal session-end --for codex through the cal at ~/.local/bin/cal (the app from the plugin's release or later; the hooks stay silent until it is there). It is built from Codex's plugin loader source and the plugins bundled with Codex and not yet install-tested; Codex runs its hooks only after you trust them in /hooks. Add the marketplace with codex plugin marketplace add BetaBots-LLC/callimachus, then install the plugin in the ChatGPT desktop app, as Codex's plugin docs describe. While it is enabled and installed in Codex the app removes its own Codex MCP entry and puts it back as it was once the plugin is off. The app's Codex hooks (Settings → Agents) stay: with the plugin's trusted too, Codex runs both, and the first to run claims each event while the plugin is enabled, so the memory still arrives once (a second clear or compaction of one session within 5 seconds injects nothing). Without the app, the plugin's cal index --if-stale hook keeps the index fresh, as the Claude Code plugin's does. Codex reads this repository's .agents/plugins/marketplace.json before .claude-plugin/marketplace.json, so adding the repository to Codex offers this plugin, not the Claude Code one. On Windows, use Settings → Agents instead: the plugin's hooks are sh scripts.

MCP server: let any agent search its own past work mid-session. It reads the index the desktop app builds and does not index anything by itself, so install the app and let it index once first (or, without the app, run cal index; see below). The quickest way to register it is npx (it downloads the prebuilt binary on first run and verifies its SHA-256); it's also listed on the official MCP registry as io.github.BetaBots-LLC/callimachus:

claude mcp add callimachus -- npx -y callimachus-mcp    # or any MCP client

Already running the desktop app? You don't need npm: Settings → Agents → Enable for Claude Code registers the app itself (launched with --mcp) as Claude Code's MCP server, and Settings → Agents → Other agents does the same for Codex, Cursor and Gemini CLI. The app keeps those entries pointed at itself after an update or a move. You can also download the standalone callimachus-mcp-<target> binary from Releases, or build from a checkout with cargo install --locked --path apps/desktop/src-tauri/standalone --bin callimachus-mcp (add --force over a source install from an earlier version).

Tools (23), read and write. Reads (18): search_threads, search_current_project (auto-scoped to the repo it runs in), recent_threads, get_thread, list_tags, list_open_todos, get_thread_knowledge, recall_decisions, recall_gotchas, find_prior_work (the "have I done this before?" guard: prior sessions similar to a task), project_memory (a project's aggregated decisions / gotchas / open TODOs), ask_history (a cited RAG answer over your history), threads_for_file (which sessions touched a path; with a line, the commits behind it and the decisions of the sessions that wrote them), check_decision (surface settled decisions before re-litigating a proposal), linked_commits (the commits a thread likely produced), recurring_issues (errors you keep hitting across sessions), list_snapshots (a project's session snapshots), and load_snapshot (restore a saved checkpoint). Writes (5, into Callimachus's own memory, never your code): complete_todo (close an open TODO), record_decision (optionally with a rationale, and replaces when it reverses a recalled decision), record_gotcha (persist a fact into a project's memory), flag_fact (mark a recalled fact stale or wrong, which keeps it out of memory until you review it; capped per project, limited to the repo the server runs in, and a fact you pinned or edited stays in), and snapshot_session (checkpoint a thread for handoff). The bundled /recall skill (resources/recall/SKILL.md) tells agents when to reach for them.

Results are compact JSON sized for an agent's context. The search tools return one entry per thread with its matching snippets, get_thread and load_snapshot take an optional budget (default 24,000 chars, max 96,000), and project_memory returns size-budgeted markdown by default (format: "json" for data, plus limit and kinds). Until the app has indexed something, every read tool answers with setup instructions instead of an empty list. The server never upgrades the index itself: when its version and the index's schema disagree, every tool says to open the app once (the index needs an upgrade) or to update callimachus-mcp or the app (the index is newer). Memory that is not distilled yet still answers: when nothing matches in a scope that isn't fully distilled, recall_decisions, recall_gotchas and check_decision return the distillation coverage and up to three closestSessions by search, project_memory adds the repo's recent sessions while it has no decisions or gotchas, and a search with zero hits says which sources the index holds, when it last indexed, and its blind spots (cloud sessions, Cursor's missing project paths).

Every tool also declares an outputSchema and returns structuredContent that matches it, which is what Claude Code and Codex hand the model: a list sits under a key ({"threads": [...]}, {"facts": [...]}), markdown under markdown, and an answer that is guidance instead of results (no index yet, a version skew, a busy write, no embedding model) is {"notice": "..."}, which every tool's schema allows. The text block is unchanged for every client except check_decision's, which lists each prior decision with its status, date and [thread N] (and what replaced a superseded one), so a client that reads only the text cannot take a reversed decision for a settled one; clients older than MCP 2025-06-18 get no structured content. The output schemas are checked in under schemas/mcp/ (served at https://callimachus.app/schemas/mcp/<tool>.json) and a test fails when one drifts. get_thread and load_snapshot carry the transcript in both the text and markdown, so their response is about twice budget.

The server also offers three MCP prompts, recall (a topic), check_decision (a proposal) and prior_work (a task), which spell out the tool order of the /recall skill; Claude Code lists them as slash commands such as /mcp__callimachus__recall. Project memory is also a resource: callimachus://memory for the repo the server runs in, callimachus://memory/{project} for any other (a path, a project key, or current).

CLI: cal, pipe-friendly. The desktop app installs it with Settings → Agents → Enable for Claude Code: on macOS and Linux it is linked at ~/.local/bin/cal (if ~/.local/bin isn't on your shell's PATH, as on stock macOS, Settings shows the export PATH=... line to add), and on Windows it is placed as cal.exe next to the app. If another program already owns ~/.local/bin/cal, or a per-machine Windows install can't write cal.exe, Callimachus leaves cal out and says so in Settings; the hooks, MCP server and /recall skill work without it. You can also run npx callimachus-mcp --install-cal (macOS and Linux), which downloads the release's cal and links it at ~/.local/bin/cal (never over another program's cal), download the standalone cal-<target> binary from Releases, or build from a checkout with cargo install --locked --path apps/desktop/src-tauri/standalone --bin cal (add --force over a source install from an earlier version).

cal search "vector index migration" -y    # -y = hybrid (semantic + keyword)
cal recent -n 10
cal cat 42 | pbcopy                        # packed transcript → clipboard
cal stats                                  # index totals, per-source breakdown, times helped this week
cal export 42 --vault ~/Obsidian           # write a thread as an Obsidian note
cal ask "how did we set up releases?"      # cited RAG answer over your history
cal files embed/mod.rs                     # threads that touched a file path
cal memory                                 # this repo's distilled memory (decisions/gotchas/TODOs)
cal memory publish                         # write its pinned + recorded facts to .callimachus/memory/ to commit (--all: distilled too)
cal memory import --review                 # what the repo's committed memory would change here (drop --review to import now)
cal memory review                          # what agents did to memory that nobody reviewed: open flags and disputes, recent replacements
cal done 17                                # mark an open TODO done (id from `cal todos`)
cal remember decision "use sqlite-vec for KNN" --because "no cloud, KNN in SQL"  # record a fact (+ rationale) into the repo's memory
cal check "switch to pgvector"             # surface settled decisions before re-litigating one
cal commits                                # infer the thread→commit timeline for this repo (--json; or pass a path)
cal why src/auth.rs:42                     # why a line is the way it is: blame → commits → the sessions behind them + their decisions
cal trace export --dry-run                 # Agent Trace records for this repo's session-produced commits (drop --dry-run to write .agent-trace/traces.jsonl)
cal issues                                 # recurring errors you keep hitting across sessions (last 180 days), with the commit that fixed each (--all: generic ones too)
cal cost                                   # estimated $ spend by model + your priciest threads
cal cost --by commit --since 30            # what the sessions behind each recent commit cost (also thread, project, day)
cal report                                 # your agents, so far: sessions, hours, spend, recurring errors, memory (markdown)
cal audit-pr . --changed-files a,b --shas s1,s2  # one JSON bundle for an external PR auditor
cal snapshot 42 -l "pre-refactor"          # checkpoint a thread for handoff (transcript + project memory)
cal snapshots                              # list saved session snapshots (optionally for a project)
cal resume 7 -a claude                     # resume a snapshot in an agent CLI
cal agents                                 # write the repo's memory into AGENTS.md (any agent reads it)
cal hook                                   # print the repo's memory (use as a Claude Code SessionStart hook)
cal hook --explain                         # what a new session in this repo would get, or why nothing
cal hook --explain --for codex             # the same for a Codex session (also gemini, copilot, cursor)
cal doctor                                 # what decides whether it works here, to paste into a bug report (read-only)

cal doctor (also Settings → General → Copy diagnostics in the app) reports the index (path, size, free space, its schema against this binary, the newest session in it, the last full pass), who holds the locks next to it, each agent's hooks and MCP entry and whether the binaries they run exist, the embedding model and vector table, the snapshots next to the index with the app's last weekly integrity check, Claude Code's transcript retention and how many sessions exist only in the index, the hooks that gave up in the last seven days by class (schema_behind, schema_ahead, busy, payload, panic, error: hooks still exit silently, and record why in ledger.skips.jsonl next to the index), and each source's parse health (records read, malformed, of an unknown type, files that yielded no message). --quick-check adds SQLite's quick_check over the whole index (about a minute per 5 GB). It is read-only, shows your home folder as ~, and prints no project text, prompts or file contents. When a source's indexing pass suddenly finds far more unknown or broken records than it usually does (it compares against that source's own last passes), Settings → Index says the agent may have changed its transcript format, so messages are not lost silently before the agent deletes its originals.

star, tag, tags, todos, knowledge, distill, decisions, gotchas, related, and similar (the "have I done this before?" check) also exist. Run cal help for the full list.

--json-version 1 is the versioned contract for scripts and tools (the VS Code extension asks for it). With it (or CAL_JSON_VERSION=1 in the environment beside --json), the outputs of search, recent, related, cat, stats, memory, decisions, gotchas, commits, why, trace, cost, report, issues, audit-pr and doctor open with schema_version (now 1); a list comes as {"schema_version": 1, "results": [...]} and an object keeps its fields next to the version. When decisions or gotchas can't run (no embedding model on the machine), the list is empty and a notice says why. Each versioned output has a JSON Schema at https://callimachus.app/schemas/cal/<command>.v1.json (in the repo, schemas/cal/), generated from the Rust types: a test fails when the two drift. A new field can appear within a version, so ignore fields you don't know; removing or renaming one, or changing its type, bumps the version, and a reader should refuse a version newer than it was written for. Plain --json still prints the bare arrays and objects it always has, so existing scripts keep working, but that shape is unversioned and may change in a future major release.

Without the app (a server, a CI box, with the standalone cal above): cal runs the app's background jobs itself, against the same index and with the same rules.

cal index                  # one pass over every agent's history (creates the index; exit 1 on errors)
cal index --if-stale       # that pass in the background, only when no full pass finished in the last hour (a hook)
cal watch                  # keep it fresh as agents write: the app's watcher, one line per batch
cal embed                  # build the semantic index (downloads the ~128 MB model once)
cal embed --no-download    # the same, only with a model already on this machine
cal distill --auto         # one drain of the auto-distill worklist under the daily cap
cal backup                 # a copy of the index in its snapshots folder (free on APFS, else a full copy)
cal backup --to ~/Backups  # the same, into a folder you choose

cal index and cal watch build and refresh the index with no app at all. cal distill --auto (and the drain cal watch runs when auto-distill is on) needs memory turned on with an engine, which only the app's Settings → AI engine does today. Once that is done it drains with the app closed: a logged-in claude or codex CLI and local Ollama need nothing more, and an API key is read from the OS keychain where the app saved it. On a machine that never ran the app, cal distill --auto says memory is off.

Only one process watches an index at a time: the app holds index.db.gui.lock next to the index while it runs and cal watch holds index.db.watch.lock. cal watch refuses to start while the app runs, and the app keeps its own watcher off while cal watch runs (Settings → Index says so) and takes over within a minute after it stops (a second app open on the same index waits for the first the same way). cal distill --auto refuses while either runs, since both distill on their own, and the app starts no auto-distill round while cal watch or cal distill --auto runs. cal index and cal embed are safe alongside either; cal index holds index.db.index.lock, so two passes never overlap, and a semantic build holds index.db.embed.lock, so two never embed the same backlog: cal embed refuses while the app or cal watch is building (try again when it finishes), and their builds skip while it runs. Shrink index and Reclaim space in the app hold index.db.rewrite.lock for the whole rewrite: cal index, cal watch, cal embed and cal distill --auto refuse to start until it is done, and the app refuses to start one while any of them is writing. cal index --if-stale, the Claude Code plugin's SessionStart hook, starts a pass only when no app, cal watch or other pass runs, the index exists, is not empty and needs no upgrade, and no full pass finished in the last hour; it appends the pass's output to index.db.index.log and, like every hook, prints nothing and never creates the index and upgrades it only with the launcher's cal while no app runs the index. These four may create and upgrade the index; cal embed without --no-download is the only cal command that downloads the model, and hooks and the MCP server still never do. cal distill --auto holds index.db.distill.lock and needs auto-distill on in Settings → AI engine. cal index and cal embed do not yield to the app's Shrink index or Reclaim space, and the app refuses those two while cal watch, cal distill --auto or a background cal index runs; rerun the command afterwards if one interrupted it.

VS Code / Cursor: the extension adds a "Callimachus History" sidebar, a status-bar search button, commands to search / insert / copy threads, and Why This Line (editor context menu) for the sessions behind the line under the cursor (it shells out to cal). Install from the VS Code Marketplace or Open VSX (the registry Cursor and VSCodium use), or grab the .vsix from Releases. See apps/vscode/README.md.

Notes / limitations

• Cross-platform (macOS / Windows / Linux), including the "Open in CLI" / "Resume" launchers. They open Terminal on macOS, a new console window on Windows, and on Linux the first terminal they find ($TERMINAL, then x-terminal-emulator, GNOME Terminal, Konsole, xfce4-terminal, kitty, Alacritty, WezTerm, xterm). On a Linux machine with none of these, the buttons are hidden.

• Cline / Roo Code / Kilo Code are editor extensions with no CLI, so they are index-only (searchable, but not relaunchable via "Resume").

• Claude Code subagent transcripts are indexed but hidden behind a "subagents" toggle by default.

• Moved your Claude Code, Codex or Copilot CLI data (CLAUDE_CONFIG_DIR / CODEX_HOME / COPILOT_HOME)? An app opened from the Dock, Finder or Start menu doesn't see shell variables, so set those folders under Settings → Sources (changing one reindexes that source).

• Large first index is a one-time cost (the Claude corpus here was ~90k messages in ~25s); subsequent passes skip unchanged files.

• The index (index.db), the embedding model (models/) and the app's diagnostics log (callimachus.log) live together in the per-user app data folder dev.shaller.callimachus (~/Library/Application Support on macOS, %LOCALAPPDATA% on Windows, ~/.local/share on Linux). For a bug report, paste the output of cal doctor (or the app's Settings → General → Copy diagnostics) and attach callimachus.log. On macOS and Linux the folder and the index files are readable only by you.

• The index is often the only copy of your history: Claude Code deletes its own transcripts after 30 days by default. On macOS the app keeps snapshots of it in snapshots/ next to it: one before each schema upgrade (kept 7 days) and one a week, taken only after PRAGMA quick_check passed. They are APFS clones, free when taken and growing as the index changes; Settings → Index & storage and cal doctor show what they cost. Nothing is deleted to make room: while the disk has less free space than the index's own size no snapshot is taken, except a weekly one that replaces the last weekly one. Shrink index / Reclaim space remove the automatic ones (the rewrite would make each a full copy; the weekly one stays while the last integrity check had failed). Each index has its own identity, so a new index built after a damaged one was moved away never removes the old one's snapshots: they stay, holding the old history, until you delete them. If the index is ever damaged, the app's startup screen offers to restore the newest one (checked first; an older one is used when it fails the check), keeping the damaged files beside it. On a disk that cannot clone (exFAT, HFS+) Settings and cal doctor say snapshots are not automatic. Linux and Windows get no automatic snapshots yet; cal backup (or cal backup --to DIR) makes a full copy there, refused when the disk has less free space than the index plus 10%.

• More sources (Charm Crush, Factory Droid) are scoped but not yet integrated. See the indexer guide.

Contributing

Issues and PRs welcome. CONTRIBUTING.md covers local setup, conventions, and the release flow. Adding support for another agent is a small, documented contract, usually one indexer module + a migration + a few wiring points.

Security & privacy

Callimachus is local-first. Your index, transcripts and memory never leave your machine, except conversation content you explicitly send to an LLM provider you choose (the in-app chat, Knowledge distillation, Ask your history, project briefs and conflict review, and AI-summarized Obsidian exports; local Ollama keeps even that on-device) and the excerpts that agents you connect over MCP or hooks pull into their own sessions. API keys live in the OS keychain, never on disk.

The desktop app can send anonymous, opt-in usage analytics to PostHog (US): event names, app version, OS, CPU architecture and language, and a few properties picked from fixed lists (such as which agent source, AI provider or setting an event concerns, an error category, or a count rounded into a range), tied to a random install ID. It never sends file paths, thread titles, prompts, transcripts, project names or API keys. It is off until you choose Allow, and you can turn it off any time in Settings → General or with CALLIMACHUS_TELEMETRY=0 / DO_NOT_TRACK=1. The CLI and MCP server never send analytics. The only other network traffic is the update check against GitHub Releases and the embedding-model download from Hugging Face (desktop app only, the first time you build the semantic index). The website uses PostHog in cookieless mode (no cookies, no cross-site tracking).

The full breakdown is in SECURITY.md and the privacy policy. To report a vulnerability, see SECURITY.md.

License

Dual-licensed. Free and open source under AGPL-3.0-or-later: use, modify, and share it; if you distribute it or run a modified version as a network service, you must release your source under the AGPL.

For closed-source, proprietary, or for-profit/SaaS use that the AGPL doesn't permit, a commercial license is required. See COMMERCIAL.md (contact ari@shaller.dev).

© 2026 Ari Shaller. See NOTICE for attributions. Contributions are accepted under the terms in CONTRIBUTING.md.

Acknowledgements

Built on Tauri, fastembed-rs, sqlite-vec, and genai. Named for Callimachus of Cyrene, who catalogued the Library of Alexandria.