enigma-cli

v1.36.8

Published

3 hours ago

Everything you need to work with a coding agent: install shared policy skills for Claude Code, OpenAI Codex and opencode, and set up portable git security hooks.

0High
0Medium
0Low

fjrg2007

ai agent claude-code codex opencode skills agents.md git-hooks secrets cli

enigma-cli

Everything you need to work with a coding agent, in one command. enigma installs a shared set of engineering policy skills into the agents you actually use (Claude Code, OpenAI Codex, opencode, Kimi Code) and sets up portable git security hooks that block secrets, .env files, and dependency dirs from being committed.

Install

One command. It installs or updates the enigma command to the latest version, then runs enigma install interactively so you choose what to set up. Re-run it anytime to update:

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/FJRG2007/enigma/main/scripts/install.sh | sh

# Windows (PowerShell)
irm https://raw.githubusercontent.com/FJRG2007/enigma/main/scripts/install.ps1 | iex

Or one-shot, no global install, no prompts - deploy the skills to every supported agent at user level:

npx enigma-cli@latest install --all --yes

That first install is the only one you ever need to run by hand: afterwards, launching a tool through enigma (e.g. enigma claude) auto-syncs the deployed skills and memory with the installed package version (see Auto-sync).

Requirements

Minimum: Node.js >= 18 (with npm), Git and at least one coding agent. Recommended: add Claude Code, the GitHub CLI, Bun and Warp.

Minimum

Commands

enigma                 Interactive menu: choose features to set up
enigma install         Install/update agent skills
enigma security        Set up git security hooks in the current repo
enigma guard [--all]   Run the commit guard: staged files, --all for all tracked, or
                       --range <base>..<head> for what a commit range touched (pre-push,
                       CI). --json emits one document; exits 1 on findings, 2 if it
                       could not run at all
enigma config [k v]    Show or set runtime toggles (e.g. config commit-emoji off)
enigma <tool> [acct]   Launch claude | codex | opencode | kimi with an account
                       (resolution: explicit > active profile > tool active);
                       auto-syncs deployed skills first
enigma account ...     Manage per-tool accounts (list/add/use/login/remove)
enigma profile ...     Group one account per tool (list/add/use/set/unset/remove)
enigma skills ...      List skills and manage discards (list/discard/restore)
enigma compress [file] Compress JSON/logs/text to fewer tokens (reversible via CCR);
                       --retrieve <hash> restores, --stats shows total savings,
                       --clear wipes all dashboard data (stats/history/cache)
enigma verify          Check that work reported as finished actually is: scans the change for
                       unfinished work and runs your verification command. parity <src> <dst>
                       compares a codebase against a port of it. --json emits one document;
                       exits 1 on findings, 2 when the check could not run. Also runs at turn
                       end, where it additionally checks the reply itself against your
                       output-style level
enigma mcp             Run the context-compression MCP server over stdio
enigma dashboard|dash  Open the local dashboard (manage enigma; see savings) in your browser (http://enigma,
                       or http://localhost:24282 if :80/hosts is unavailable)
enigma ssh [alias]     SSH connection manager: connect by alias, or list | add | edit | remove |
                       info; tunnels are standalone: tunnel add <name> <server> <spec>,
                       tunnel start|stop <name>, tunnels (list with live status)
                       (encrypted passwords, saved key/jump/port-forwards; e.g. 9090:db:5432)
enigma seal            Maintenance: (re)compute skill content hashes
enigma check           Integrity gate: verify skills are well-formed and sealed
enigma help | version

Every command answers enigma <command> --help with its own usage, flags and exit codes.

Running enigma inside a container or CI runner

enigma install --hooks <cls>    Which hooks to wire: post-edit, stop | all | none
enigma install --no-hooks       Wire none of them        --no-statusline  Leave statusLine alone
enigma install --ref <tag|sha>  Pin the skills ref (the resolved commit is printed)
enigma install --assets-from <dir>  Install from a staged assets tree (implies --offline)
enigma install --offline        Make no network call at all

ENIGMA_AGENT_CLAUDE=<abs path>   Point the gate at an agent installed off PATH
                                 (also _CODEX, _OPENCODE, _ROVODEV, _PI). The daemon runs
                                 the pipeline, so set it before the daemon starts
ENIGMA_BIN_PATH=<abs path>       Use this enigma binary instead of downloading the release
ENIGMA_SKILLS_REF=<tag|sha>      Default skills ref (install --ref overrides it)
ENIGMA_OFFLINE=1                 No enigma command reaches the network - the skill check,
                                 the update notice and the background linter/dashboard
                                 installs all stand down (install --offline sets it)

gh is needed only by the gate's push, pr and ci steps: enigma gate axi run --skip push,pr,ci and enigma gate init work in an image without it.

Agent skills

Skills are authored once and deployed to every selected agent (no per-agent duplication). enigma install auto-detects which agents are installed (CLI on PATH or a config dir like ~/.claude, ~/.codex, ~/.config/opencode, ~/.kimi-code) and preselects them; --all targets every supported agent.

| Agent | Scope | Skills | Memory file | | ----------- | ------ | ----------------------------- | ------------------------------ | | Claude Code | global | ~/.claude/skills/ | ~/.claude/CLAUDE.md | | OpenAI Codex| global | ~/.agents/skills/ | ~/.codex/AGENTS.md | | opencode | global | ~/.config/opencode/skills/ | ~/.config/opencode/AGENTS.md |

(--local installs into the current project instead.)

Don't want one of the skills? Discard it from the hub's install panel (the SKILLS section lists every skill; unchecking one discards it) or with enigma skills discard <name>: it is removed from every agent and skipped by future installs, updates and auto-syncs until you restore it with enigma skills restore <name>.

Slash commands

enigma install also deploys reusable slash commands to every selected agent (a full install only - --skills-only / --memory-only skip them). Each command is authored once and copied verbatim to the agent's command directory, where the file name becomes the command.

| Agent | Scope | Command directory | | ----------- | ------ | ------------------------------- | | Claude Code | global | ~/.claude/commands/ | | opencode | global | ~/.config/opencode/command/ | | OpenAI Codex| global | ~/.codex/prompts/ |

(--local installs into the project's .claude/commands/ and .opencode/command/. Codex has no project-local prompt directory, so it only receives commands at global scope.)

Commands are enigma-managed: if a same-named command already exists and is not enigma's, it is replaced so enigma's command always wins the name. A command you have not changed is left untouched; auto-sync keeps it current on every launch.

`/improve`

Two modes in one command. Implement mode edits a focused area directly; Advisor mode is strictly read-only and writes self-contained implementation plans into plans/ for another (cheaper) agent to execute.

The mode is resolved from the arguments: an advisor keyword (audit, quick, deep, branch, next, plan, review-plan, execute, reconcile) selects Advisor mode; otherwise an area token selects Implement mode. A bare security or performance runs Implement mode (edits code) for backward compatibility - to audit those read-only instead, prefix an advisor keyword (audit security, quick perf). With no argument (or an unknown one) it prints both usages and stops.

Implement mode (/improve <area>):

| Invocation | What it does | | -------------------------------- | --------------------------------------------------- | | /improve ui or /improve frontend | Visual design, components, accessibility, responsiveness (same workflow) | | /improve security | Secrets, authz, input validation, OWASP, dependency audit | | /improve performance | Profile hot paths, queries/indexes, caching, bundle/render | | /improve seo | Metadata, semantic HTML, structured data, crawlability, Core Web Vitals | | /improve refactor or /improve refactorize | Dedup and consistency: consolidate duplicate code/components and divergent implementations of one concept into a single source of truth, apply ciphera-style and minimal-code, remove dead code - without changing behavior |

It detects the project stack first, reuses existing code, applies the smallest change, follows any matching policy skill, and verifies with the project's build/lint/test before reporting.

Advisor mode (read-only; writes only to plans/):

| Invocation | What it does | | -------------------------------- | --------------------------------------------------- | | /improve audit | Full audit -> prioritized findings table -> plans you select | | /improve quick / /improve deep | Effort level for the audit (hotspots only / whole repo) | | /improve audit <focus> | Focused audit (e.g. security, perf, tests, bugs) | | /improve branch | Audit only what the current branch changes | | /improve next | Grounded feature/direction suggestions (also features, roadmap) | | /improve plan <description> | Skip the audit, spec one thing as a single plan | | /improve review-plan <file> | Critique and tighten an existing plan | | /improve execute <plan> | Dispatch a cheaper executor in an isolated worktree, review its diff | | /improve reconcile | Refresh the backlog: verify DONE, unblock, retire dead findings | | ... --issues | Also publish each written plan as a GitHub issue via gh |

Advisor mode never modifies source code, never mutates the working tree, and never reproduces secret values. Each plan is self-contained, stamps the commit it was written against, and carries machine-checkable done criteria and STOP conditions so a weaker executor can run it without this session's context.

Auto-sync on launch

After the first enigma install, you never need to run it again: whenever you launch a tool through enigma (enigma claude, enigma account run work), enigma first compares the deployed skills/memory against the installed package version and silently refreshes anything that changed (new skills, updated versions, removed skills, memory-file edits). On by default; opt out with:

enigma config auto-sync off

Auto-sync is deliberately conservative:

It only touches agents/scopes that already have a deployment - it never performs a first install (that stays your explicit enigma install).
Skills you modified locally are never overwritten (same rule as --keep-modified).
The memory file (CLAUDE.md / AGENTS.md) is only rewritten when it is byte-identical to what enigma last wrote (tracked in ~/.enigma/state.json) - a file you authored or edited is never touched.
A sync failure never blocks the launch; the tool starts anyway.

Auto-lint on edit

Opt-in. When on, enigma autonomously installs @enigmax/linter into a managed dir (~/.enigma/linter, in the background) and wires a post-write hook into each agent, so every file an agent writes is linted the moment the edit finishes:

enigma config auto-lint on    # installs the linter and wires the hooks
enigma config auto-lint off   # removes the hooks

The hook is designed for minimum token cost: it auto-fixes the safe formatting rules in place, and surfaces only the unfixable findings (hardcoded secrets, URL imports, style issues that need judgement) back to the agent. A clean file produces no output at all - zero added tokens.

It rewrites the whole edited file, so it is worth knowing exactly what it changes: trailing whitespace, blank lines (collapsed, leading/trailing removed), the final newline, the order of a contiguous import group (shortest declaration first), the semicolon terminating an interface or type-literal member, and the trailing comma in a named import/export list. Judgement-dependent rules and the security audits are only reported. The last two - the punctuation fixes - are skipped in a project configured with Prettier (a .prettierrc* / prettier.config.* file, or a prettier key in a package.json, looked up as far as the project root), so the two formatters never fight over the same lines; their two rules stay quiet there too, so nothing is reported that the hook has refused to fix.

Claude Code: a PostToolUse hook in settings.json (matcher Edit|Write|MultiEdit|NotebookEdit). Unfixable findings come back via the hook's stderr, which Claude feeds to the model.
opencode: an auto-loaded plugin in ~/.config/opencode/plugins/ whose tool.execute.after runs the same check and appends the findings to the tool output the model sees.

Both invoke one shared runner (~/.enigma/hooks/lint-hook.mjs) that resolves the managed linter and no-ops cleanly until the background install lands (self-healing). The wiring is mirrored into managed account config dirs on sync, like the other managed settings. Default: off (it changes agent behavior and installs a package, so it stays an explicit choice).

Git security hooks

enigma security drops a portable, dependency-free commit guard into any repo: it copies guard.mjs into the repo's .githooks/, writes a cross-platform pre-commit shim and a toggle config, and points core.hooksPath at it. Commit .githooks/ so the team inherits it. Because it runs on git commit, it also covers commits made through the GitHub CLI (gh).

On every commit the guard, OS-agnostically:

Blocks committed secrets (API keys, tokens, private keys).
Blocks .env / .env.local (allows *.example / *.sample / *.template).
Blocks dependency/cache dirs (node_modules, __pycache__, virtualenvs).
Warns on generated dirs (dist, build, .next, coverage), log/OS-junk files, and files over 5 MB.

Each protection is individually toggleable (saved to .githooks/enigma-guard.json). Bypass once with git commit --no-verify.

enigma guard also runs standalone, outside the pre-commit hook, for a pre-push hook or CI job where nothing is staged: enigma guard --range <base>..<head> scans exactly what a commit range touched (a lone ref means <ref>..HEAD), reading each file at that range's head rather than the working tree, so a secret committed and then edited out is still caught. --json emits one machine-readable document; exit codes split 0 clean, 1 blocking findings, 2 the guard could not run at all (no repository, an unresolvable range).

Multiple accounts and profiles

Work across separate workflows with different accounts per tool - for example your company Claude Code account and your personal one - keeping each fully isolated and switching between them without ever logging out. Each account has its own credentials, session, and history, so client work never mixes with personal projects. Supported tools: Claude Code (CLAUDE_CONFIG_DIR), OpenAI Codex (CODEX_HOME), OpenCode (a private XDG_DATA_HOME / XDG_CONFIG_HOME pair per managed account; its default account keeps your real environment untouched) and Kimi Code (KIMI_CODE_HOME, which relocates its whole data root - credentials, sessions, skills and AGENTS.md).

This is for legitimate, professional account separation (one account per employer/context, as many organizations require). It is not a way to evade usage limits or Anthropic's terms - each account still authenticates as itself and is subject to its own limits. Use the account that each piece of work belongs to.

Claude Code reads its credentials and session from the directory in CLAUDE_CONFIG_DIR (default ~/.claude), so each profile just needs its own directory. Rather than hand-editing per-shell aliases, enigma launches Claude for you with that variable set - the same command on macOS, Linux and Windows.

enigma account add work --login   # create 'work' and run /login to authenticate
enigma account add personal       # create 'personal' (log in later)
enigma account add acme -t codex  # create a Codex account
enigma account list               # show all accounts (active one marked *)
enigma claude work                # run Claude Code as 'work'
enigma codex acme                 # run Codex as 'acme'
enigma account use personal       # make 'personal' the active account
enigma claude                     # run the resolved account (profile > active)
enigma claude work --resume <id>  # flags after the tool go to the tool
enigma account rename work corp   # rename an account (its config dir moves)
enigma account remove work        # delete an account and its config dir

Your existing ~/.claude / ~/.codex / ~/.kimi-code / opencode setup is always available as each tool's built-in default account (never deleted). New accounts live under ~/.enigma/<tool>/<name>/. Bare claude / codex / opencode / kimi commands keep using your real environment as before.

Managed accounts inherit your enigma setup automatically: because the tool reads everything from the account's config dir, enigma deploys the skills and memory file into it (seeded on account add and on launch) and mirrors the enigma-managed native settings from your default account on every launch - Claude's permission bypass, attribution overrides and statusline, Codex's approval_policy/sandbox_mode, opencode's "*": "allow" permission. Turning a knob off on the default account propagates too; per-account manual edits to those specific knobs are overwritten on the next launch, while every other account setting (theme, custom statusline, extra permissions) is left untouched.

Profiles (one account per tool)

A profile pins one account per tool under a single name - e.g. profile work = Claude Code work + Codex acme. While a profile is active, enigma <tool> launches that profile's account for the tool (explicit account arguments still win; unmapped tools fall back to their own active account):

enigma profile add work                 # create the profile
enigma profile set work claude work     # pin claude account 'work'
enigma profile set work codex acme      # pin codex account 'acme'
enigma profile use work                 # activate (enigma claude/codex now use it)
enigma profile list                     # profiles + mappings (* = active)
enigma profile rename work corp         # rename a profile (mappings stay)
enigma profile use none                 # deactivate

From the hub TUI (enigma), the Accounts panel lists every tool's accounts - with the signed-in identity (email for Claude/Codex, connected providers for OpenCode and Kimi Code) - and lets you add (a), set active (enter), connect/log in (c), rename (r), or remove (d). Adding first asks which tool with a searchable selector (type to filter, like opencode's model picker), then the account name, then offers to connect right away. The Profiles panel manages profiles end-to-end: enter switches the active one ((none) deactivates), a creates one, e edits its mappings (searchable tool selector, then a searchable account selector including (unpin)), r renames it, and d removes it (accounts are kept).

GitHub CLI telemetry (default off)

If the GitHub CLI (gh) is installed, enigma install disables its usage telemetry (gh config set telemetry disabled). This is pure privacy upside - telemetry is usage analytics only (command, flags, OS/version, device ids) and no gh feature depends on it - and it also avoids a known Windows bug where the detached gh send-telemetry subprocess spawns tzutil.exe without hiding its window, flashing a terminal on gh invocations (cli/cli#13354). Re-enable any time:

enigma config gh-telemetry on     # restore gh's default
enigma config gh-telemetry off    # disable again

Claude feedback survey (default off)

When Claude Code is a target, enigma install disables the periodic "How is Claude doing?" session-quality survey by setting env.CLAUDE_CODE_DISABLE_FEEDBACK_SURVEY to "1" in Claude's settings.json. It is a recurring prompt with no functional value, so it is off by default. Re-enable any time:

enigma config claude-survey on        # restore Claude's survey
enigma config claude-survey off       # disable again
enigma config claude-survey off -g    # global (~/.claude/settings.json)

Claude workspace trust (default on)

Claude Code asks "Is this a project you trust?" the first time it starts in a folder, and in your home directory it asks every time - the answer is never persisted there. enigma answers it for you, for any path, by recording the trust in Claude's own .claude.json: the filesystem root (which Claude inherits down every directory) plus each workspace you launch in.

This is a security trade-off, and the reason it is a toggle: that prompt is what makes you look at unfamiliar code before an agent runs in it.

enigma config claude-trust off        # keep Claude's trust prompt
enigma config claude-trust on         # answer it once and for all (default)

Turning it off stops the blanket, and leaves the folders already trusted alone - it never re-asks for projects you accepted yourself.

Commit emojis

By default the policy skills make commit subjects carry a leading type emoji (one per subject, e.g. for feat/fix); code, prose, and PR text stay emoji-free. The convention and its type-to-emoji map live in the git-policy skill, and the commits the gate makes for you follow the same setting. Opt out per repo or globally:

enigma config commit-emoji off       # disable (writes .enigma.json)
enigma config commit-emoji on        # re-enable
enigma config commit-emoji off -g    # global (~/.enigma.json)

Precedence: built-in default (on) -> ~/.enigma.json -> repo .enigma.json.

Parallel sub-agents

The memory file always tells agents to break long or complex tasks into smaller subtasks and complete them incrementally. The parallel part - delegating independent subtasks to sub-agents that run at the same time to finish faster - is opt-in, because spawning sub-agents multiplies token cost:

enigma config parallel-subagents on        # add the parallel section to the memory file
enigma config parallel-subagents off       # remove it (default)
enigma config parallel-subagents on -g      # global (~/.enigma.json)

This toggle edits the deployed agent memory file (adds or removes the section), so restart Claude Code / Codex / OpenCode after changing it for the new session to pick it up. Subtask decomposition itself is always on and is never removed.

Token-efficient output

Optionally compress the agent's chat prose to cut output tokens while keeping full technical accuracy. Chosen at install or via config; off by default:

enigma config output-style lite     # professional terse (drop filler, keep grammar)
enigma config output-style full     # shorter, drops articles and uses fragments
enigma config output-style ultra    # telegraphic, maximum compression
enigma config output-style off      # back to full prose (default)
enigma install --output-style lite  # set it during install

on/off also work (on = full). Like the toggle above it edits the memory file, so restart your agent after changing it. Code, comments, commits, and PRs stay normal, and the agent reverts to full prose for security warnings and other safety-critical replies. Only the selected level's rules are deployed, so asking for "normal mode" lifts the compression for the rest of the session while changing level takes enigma config output-style <level> and an agent restart.

Set to anything but off, it is also enforced: the enigma verify turn-end hook scans the reply itself for filler, an announcing-the-work preamble, and padded table rows, and denies the stop on the ones it can catch without false positives (enigma guardrails stats reports how often a reply broke the level).

Claude Code's status bar shows an [ENIGMA] badge at all times; while the mode is active it appends the level ([ENIGMA:FULL], [ENIGMA:LITE], [ENIGMA:ULTRA]). enigma wires this into settings.json during enigma install, only when you have no status line configured (it never replaces your own). If you upgraded the package, re-run enigma install once to wire it.

Minimal code (anti-overengineering)

The companion to token-efficient output: that compresses how the agent talks, this governs how it builds. It pushes the agent toward the laziest solution that works - YAGNI, the standard library and native platform features before custom code, one line before fifty. On by default at full:

enigma config minimal-code lite     # build what's asked, name the lazier alternative
enigma config minimal-code full     # YAGNI ladder enforced, shortest working diff (default)
enigma config minimal-code ultra    # YAGNI extremist, deletion before addition
enigma config minimal-code off      # opt out, no extra pressure
enigma install --minimal-code lite  # set a different level during install

on/off also work (on = full). Like token-efficient output it edits the memory file, so restart your agent after changing it. Security, input validation at trust boundaries, data-loss error handling, and accessibility are never simplified away. The full discipline lives in the anti-overengineering-policy skill, and the level is switchable mid-session by asking ("be more lazy", "full", "stop minimal-code").

For the on-demand passes, the anti-overengineering-review skill reviews a diff, audits the whole repo, or harvests the enigma: shortcut markers into a debt ledger - ask to "review for over-engineering", "audit the codebase for bloat", "what can we delete", or "list the deferred shortcuts". It emits a tagged list of cuts (stdlib/native/yagni/delete/shrink) with a net: -N lines score and applies nothing; correctness and security stay with the normal review.

Side-by-side "with vs without enigma" comparisons live in docs/examples/.

Context compression (opt-in)

A native, dependency-free engine that shrinks large content before it reaches the model - same information, far fewer tokens. It detects the content type and routes it:

JSON arrays of records are compressed statistically: a representative, schema-preserving sample is kept (head + stride samples + tail), and error rows, numeric anomalies and structural outliers are always preserved - unique entities with no signal are left intact so real data is never silently dropped.
Logs collapse by template (volatile timestamps/numbers/hashes are masked; runs of similar lines fold into one + a count), keeping error/warning lines.
Plain text is truncated head+tail. Code, diffs and markdown pass through untouched (no safe lossy transform without a parser).

Compression is reversible (CCR - Compress-Cache-Retrieve): when rows or spans are dropped, the full original is cached under ~/.enigma/ccr and the output carries a <<enigma:ccr:HASH ...>> marker. Pass the hash back to get the original.

cat tool-output.json | enigma compress     # compressed to stdout, savings to stderr
enigma compress big.log                     # compress a file
enigma compress --retrieve <hash>           # restore a cached original
enigma compress --stats                     # cumulative token savings
enigma compress --clear                     # wipe all dashboard data (stats/history/cache)

Dashboard (manage enigma + see savings)

enigma dashboard (alias dash) serves a local, loopback-only browser dashboard to manage all of enigma - switch accounts and profiles, enable or edit skills per app, change any setting, and free up your machine (kill a port, shut down WSL, quit Docker). It also surfaces real Claude usage and the compression savings enigma can measure: totals, estimated money saved, a per-day graph (tokens/$ and daily/cumulative), breakdowns by the app that requested each compression and by content type, a savings history, a recent-compressions table and reversible-cache (CCR) stats. It runs only while open; enigma config dashboard always keeps a lightweight background daemon, and enigma compress --clear resets the savings data.

The dashboard is fully modular - if you don't want it (e.g. on a server), you never need it, and you can add or remove it at any time. Its browser UI (the page plus a ~196 KB chart library) is not bundled in enigma-cli: it ships as a separate package, @enigmax/dashboard, that enigma fetches on demand the first time you open or enable the dashboard, into ~/.enigma/dashboard. enigma keeps it current on enigma update, so it is enigma's dependency to maintain, not yours. If you never use the dashboard, that bundle is never downloaded. enigma config dashboard off is a complete teardown: it stops any background daemon and removes the enigma hosts-file entry, so nothing is left behind. Re-enable it whenever you like with on-demand or always.

The Savings page also shows an Enigma Systems overview (which systems are active and how they're configured - only Context Compression and the opt-in usage/proxy stats are measured; the rest is shown as state, never invented savings), a Check & update button (refreshes skills and the dashboard UI in place; a CLI bump still needs enigma update in a terminal), and the Skills subpage to view/edit/enable/disable/remove skills and check for updates.

Proxy (experimental, opt-in, Claude Code only)

enigma config proxy on makes enigma claude route Claude Code through a local loopback proxy that forwards every request verbatim to Anthropic and streams the response straight back, reading only the token usage to record real measurements (no estimates). It is off by default and deliberately conservative: a faithful pass-through (no content or cache rewriting), it never stores auth headers or message content, binds 127.0.0.1 only, and falls back to a direct launch if it can't start. It applies only to launches via enigma claude (it injects ANTHROPIC_BASE_URL for that process only) and only to Claude Code. Leave it off until you've confirmed it works for your setup.

Secret protection (commit guard + prompt guard)

enigma's commit guard already blocks committed secrets, .env files and dependency dirs. You can now tailor it granularly - and not just for git:

Granular file rules. On top of the built-in protections, add your own blocked paths (extra globs to refuse, e.g. secrets/*.json), excluded paths (an allowlist the guard never flags, e.g. tests/fixtures/**), and custom secret patterns (extra regexes). Edit them visually in the dashboard Enigma Settings panel or the terminal UI (enigma config), or from the CLI: enigma config guard-block-paths add "secrets/*.json", enigma config guard-allow-paths add "tests/fixtures/**", enigma config guard-secret-patterns add "mycorp_[a-z0-9]{32}" (and ... remove <entry>, or the key alone to list). Defaults are unchanged; these only refine them. They live in ~/.enigma-guard.json and a repo can override per-.githooks/enigma-guard.json.
Prompt secret guard (opt-in, off by default, Claude Code only). Turn on enigma config prompt-secret-guard on and enigma claude routes through the local proxy, scanning each outgoing chat message for credentials before they reach the model. On a hit it either redacts the secret (default - replaces it with [REDACTED: ...] so the key never leaves your machine but the turn still works) or rejects the whole request (enigma config prompt-secret-mode reject - nothing reaches Claude). It uses the same patterns as the commit guard plus your custom ones. The dashboard's Enigma Systems panel shows the guard's state and how many prompts it has actually redacted/rejected.

You can also configure enigma from the dashboard itself: the Enigma Settings panel exposes the same options as the terminal UI (enigma config), editable in the browser. Writes apply immediately at global scope; memory-affecting toggles prompt for an agent restart. The settings write endpoint is loopback-only and origin-guarded (cross-site and DNS-rebinding requests are refused), and the server never binds anything but 127.0.0.1.

The money figure is an estimate: enigma isn't a proxy, so it can't see the model - it prices saved tokens per source with sensible defaults. Override the rate with enigma config token-price <usd-per-1M-input-tokens> (0 restores the defaults).

The time figure is likewise an estimate: it converts saved input tokens into saved prefill time at a default model speed. Override it with enigma config token-speed <tokens-per-second> (0 restores the default rate).

Claude usage (opt-in)

enigma config usage-stats on unlocks a full Usage view - in the dashboard (its own tab) and the terminal UI (a "Claude usage" entry) - that reads your own Claude Code session transcripts (~/.claude/projects/.../*.jsonl) and reports:

Estimated cost in USD from a per-model price table (Opus/Sonnet/Haiku, incl. cache read/write rates), plus measured input/output/cache tokens.
Breakdowns by model, by project, by account (every Claude login - default and managed accounts, not just ~/.claude), and a recent sessions table.
A provider coverage line: only Claude Code keeps a readable local usage store, so Codex, OpenCode and Kimi Code are shown as unavailable (no local token store) rather than faked. The session/weekly windows are Claude-specific (they come from Anthropic's rate-limit headers).
A current 5-hour block computed locally from transcript timestamps: tokens + cost used in the open window, the burn rate (tokens/min) and a projected end-of-window total. (No Anthropic API is called - this is reconstructed from your local logs.)
The same usage windows Claude Code shows as gauges: Current session, Weekly - All models, and Weekly - Sonnet only, each with its reset time and % used.
- Live (recommended): with the proxy on (enigma config proxy on) and launching via enigma claude, enigma reads Anthropic's real anthropic-ratelimit-unified-* response headers from your traffic and shows the exact %/reset Claude's own UI shows - no plan limit needed. Cards backed by this are marked live. (This is how usage trackers get the numbers; enigma gets them for free from traffic you already make, with no extra API call or token handling.)
- Fallback: before the proxy has seen a request, the card shows tokens used, or a % against a plan limit you set on the card or with enigma config plan-weekly-limit <tokens> / plan-session-limit / plan-weekly-sonnet-limit (weekly reset via enigma config plan-weekly-reset "mon 11:00"). enigma never fabricates a percentage.

It is read-only and loopback-only; nothing is sent anywhere. Off by default because it reads your session logs, which are broader than enigma's own compression data.

Honesty note: cost is an estimate (Anthropic does not record per-message cost in the transcript; real spend is billed by Anthropic). Token counts and prompt-cache reads are measured facts. enigma deliberately does not attribute savings to skills or to token-efficient output - a transcript has no counterfactual baseline, so any such figure would be invented. Only Claude Code is read today; Codex/OpenCode/Kimi Code use absent or undocumented local session stores and are not guessed.

Manage accounts & profiles from the dashboard

The dashboard Accounts tab does everything the terminal UI does: switch the active account, add / rename / remove managed accounts, and create profiles that pin one account per tool. Logging an account in still happens in a terminal (the browser cannot host a tool's interactive login), so the panel shows the exact command to run, e.g. enigma claude work.

Export / import your config

The Settings tab has Export config and Import config. Export downloads a single JSON bundle - your .enigma.json, the commit-guard config, and your account/profile structure (names + tool->account mappings). It is secret-free: no auth tokens or credentials are ever included, so the file is safe to move between machines or commit. Importing recreates the structure (seeding fresh account dirs); you then log each account in per machine. Both use the browser's native file dialogs, so you always choose where to save or which file to import.

As an MCP server

enigma mcp runs a stdio MCP server exposing three tools - enigma_compress, enigma_retrieve, enigma_stats - so an agent can compress large tool outputs itself and retrieve originals on demand. Enable deployment and it is registered in each managed agent's own config (Claude Code mcpServers, Codex [mcp_servers.enigma], opencode mcp), preserving your other servers:

enigma config compress on     # register the MCP server on install/sync
enigma config compress off    # remove it again

Off by default - adding an MCP server to your agents is an explicit choice. Everything runs locally; no data leaves your machine.

License

Apache-2.0.

Published

Vulnerabilities

Links

Maintainers

Keywords

Readme

enigma-cli

Install

Requirements

Minimum

Recommended

Commands

Running enigma inside a container or CI runner

Agent skills

Slash commands

/improve

Auto-sync on launch

Auto-lint on edit

Git security hooks

Multiple accounts and profiles

Profiles (one account per tool)

GitHub CLI telemetry (default off)

Claude feedback survey (default off)

Claude workspace trust (default on)

Commit emojis

Parallel sub-agents

Token-efficient output

Minimal code (anti-overengineering)

Context compression (opt-in)

Dashboard (manage enigma + see savings)

Proxy (experimental, opt-in, Claude Code only)

Secret protection (commit guard + prompt guard)

Claude usage (opt-in)

Manage accounts & profiles from the dashboard

Export / import your config

As an MCP server

License

`/improve`