MindForge

A governance and orchestration layer for Claude Code (and Antigravity, Cursor, Copilot, Gemini, OpenCode).

Claude Code alone runs one agent in one context. MindForge adds the parts that don't fit in a single context window: skills that auto-load by trigger, personas you can call by name, a wave-based executor that fans work out to fresh-context subagents and commits per task, a tamper-evident audit chain, and cost-aware routing across providers. Install it once and get /mindforge:plan-phase → /mindforge:execute-phase → /mindforge:verify-phase → /mindforge:ship as your actual working loop, not a slogan.

221 commands · 354 skills · 216 personas · 164 subagents · 35 workflows

Jump to: Latest release · What you get · What is actually enforced · Install · Verify · Quick start (new) · Quick start (existing) · How it fits together · Documentation · Core workflow · Dynamic workflows · Updates · Token usage · License

What you get

Ships three ways: an npm package (npx mindforge-cc@latest), a Claude Code plugin marketplace entry, and an MCP server.

What is actually enforced

Read this before you rely on anything below blocking a bad command. MindForge ships a large corpus of agent instructions — commands, skills, personas, protocols — and those are advisory: they work by being in the model's context, and a model can decline them. The parts that would block an action are hooks. Through 11.9.2 no channel registered them. 11.9.3 added the registration code but it declined to run on almost every project, so in practice nothing was enforced there either. As of 11.9.4 both channels register and execute them on Claude Code, and nowhere else.

What that means, measured rather than asserted:

• The npx channel generates the config it never used to ship. files[] has 53 entries and none of them contains settings, so no settings file is published — instead bin/installer/hook-registration.js writes one at install time, merging append-only into any file you already have. Measured on a confined install: 8 hooks registered into .claude/settings.json, of which the installer's own preflight executed 7 and verified all 3 deny-class hooks returning exit 2 before keeping the file. A preflight failure rolls the registration back rather than leaving a config whose commands do not run.
• The plugin channel's dispatcher runs. It previously crashed on every fire — run-with-flags.js requires ./lib/hook-flags and plugins/mindforge/scripts/lib/ was not copied in. That directory now exists, all 14 path tokens in plugins/mindforge/hooks/hooks.json resolve under the plugin root, and driving the dispatcher by hand returns exit 2 for mindforge-block-no-verify and mindforge-config-protection.

[!WARNING] Still not enforced, deliberately and with a printed reason for each: any runtime other than Claude Code (Cursor, Copilot, Gemini/Antigravity, OpenCode), --global scope, a self-install inside a MindForge checkout, and Windows. Writing a Claude-schema config into .cursor/ without an execution-verified hook contract would be decorative. Every outcome, including "not registered", is printed by the installer and written to .mindforge/hook-registration.json.

Three things gate whether a registered hook is live, none of them in MindForge's control: the harness must be restarted (hooks are snapshotted at session start), the project must be trusted in the harness, and CLAUDE_PROJECT_DIR must be set with node on the hook PATH — if it is not, the commands exit 1 and the gate is simply absent, which is a deliberate trade against a fail-closed tail that was measured denying benign commands on a fresh clone. See Hooks are installed but nothing is blocked in docs/troubleshooting.md.

So: on Claude Code, treat MindForge as a policy enforcement point for the 8 registered hooks plus a tamper-evident audit log; on every other harness, as governance-by-convention plus that same audit log. Installing it also expands your repository's trust boundary by a large volume of agent instructions — review what you install. The audit chain is verifiable today (node bin/verify-audit.js).

Install

Pick whichever matches how you work — all of these are real, live channels.

On a typical connection, npx mindforge-cc@latest --claude --local finishes in well under 10 seconds (measured: ~5.4s locally) — reproduce with time npx mindforge-cc@latest --claude --local in an empty directory. Install makes zero model calls; you don't spend a token or a dollar until you run a command that dispatches to a subagent.

npx (recommended)

Writes .mindforge/ governance, memory, and planning into your project:

npx mindforge-cc@latest --claude --local      # Claude Code, this project only
npx mindforge-cc@latest --antigravity --local # Antigravity, this project only
npx mindforge-cc@latest                       # interactive wizard, pre-selects a detected runtime

The bare form only detects anything inside an interactive TTY wizard session, where it pre-selects — you still confirm — whichever runtime it finds. Run it non-interactively (CI, piped, scripted, or anywhere stdin isn't a TTY) and it skips the wizard entirely and installs --claude by default, regardless of what's actually on the machine.

Global (system-wide, for your primary AI coding runtime):

npx mindforge-cc@latest --claude --global

(npm install -g mindforge-cc@latest only puts the mindforge-cc/mindforge binaries on your PATH — it doesn't select a runtime or write anything. Run the command above, or the equivalent mindforge-cc --claude --global once installed, to actually scaffold a global setup.)

Other runtimes — same flag pattern, swap --global/--local:

Advanced: --runtime claude,cursor (combined runtimes) · --with-utils (installs local bin/ utilities) · --minimal (essential scaffolding only, no persona library) · --force (rewrite an existing .mindforge/MINDFORGE-SCHEMA.json with the current, stricter schema)

Claude Code plugin marketplace

No project files written — the plugin's hooks fire, see What is actually enforced for what that does and does not cover.

/plugin marketplace add sairam0424/MindForge
/plugin install mindforge@mindforge

Prefer just a slice (e.g. Python agents)? mindforge-lang@mindforge and 9 other focused packs exist — see docs/plugin-installation.md for all 10, token-budget guidance, and team setup.

Standalone MCP server

claude mcp add mindforge -- npx -y mindforge-mcp-server

Exposes 8 tools over stdio (6 read-only, 1 guarded write, 1 guarded browse proxy). Also listed on the MCP Registry as io.github.sairam0424/mindforge — that entry is republished manually and can lag; check what it actually serves before relying on it, or install mindforge-mcp-server from npm directly to pin a version.

Homebrew

brew install sairam0424/tap/mindforge

SDK

Build on MindForge programmatically:

npm i mindforge-sdk

[!NOTE] Upgrading from 11.9.x? The installer does not overwrite an existing .mindforge/MINDFORGE-SCHEMA.json, so 11.9.2's armed config validator keeps the older permissive schema on a plain upgrade — run with --force for the stricter gate. The daily cost cap declared as [COST_HARD_LIMIT_USD] in MINDFORGE.md was not enforced in 11.9.2 (11.9.3 arms it), and an upgrade never rewrites an existing MINDFORGE.md — add [COST_HARD_LIMIT_USD] = 25.00 yourself if yours predates the key.

Full install matrix, plugin packs, and team-setup guidance: docs/getting-started.md.

Verify

These /mindforge:* commands require the Claude Code plugin or an npx/Homebrew framework install — the standalone MCP server exposes MCP tools instead, and mindforge-sdk exposes a programmatic API; neither installs these slash commands.

/mindforge:health              # framework + installation health check
/mindforge:health --repair     # documented in the command spec, but NOT wired into the CLI
                                # backing path — silently ignored, output is byte-identical to plain health
/mindforge:status              # project status snapshot
/mindforge:next                # auto-discover your first task

Full verification walkthrough: docs/quick-verify.md.

Quick start (new project)

/mindforge:init-project
/mindforge:plan-phase 1
/mindforge:execute-phase 1
/mindforge:verify-phase 1
/mindforge:ship 1

Quick start (existing codebase)

/mindforge:map-codebase
/mindforge:do I want to plan the next phase
/mindforge:plan-phase 1

Latest release

v12.0.0 (2026-09-24) — First release aimed at real external users. The major-version bump marks that shift, not a breaking change — there isn't one; every item here is a fix. A second, independent 8-agent audit checked whether v11.9.9 actually cleared that bar (security/STRIDE, staff-engineer code review, deps+license, a live production dry-run across all 6 supported runtimes, docs accuracy, re-verification of the prior release's deferred backlog, test-coverage gaps, and a full trace of the release pipeline) and found 1 CRITICAL + 4 HIGH issues still standing in the way. The CRITICAL: --global installs printed a fabricated banner claiming 216 personas/122 skills were "active" while writing none of them — now prints an honest description of what a global install actually writes. The four HIGH: an unhedged "active" claim for a confirmed no-op feature plus unearned "Autonomous Enterprise/Sovereign" marketing language in the install banner, both reworded; a dynamic-workflow script's own null-guard commit missed one crash-causing edge case, now covered; and two real test-coverage gaps (nothing guarded the removed godmode skill or the --minimal persona fix against regressing) are closed. Eight more, lower severity: a security dropper-chain pattern gap, a latent prototype-pollution path, a CodeQL-flagged regex-escape bug, a non-LTS Node base image that slipped in via Dependabot, a docs table citing 16 nonexistent personas, a release-pipeline step that's failed cosmetically on the last 4 releases, a cross-runtime hook-registration parity gap, and CI/doc hygiene fixes. See RELEASENOTES.md or CHANGELOG.md.

How it fits together

                     /mindforge:plan-phase N
                              |
                              v
   Skill Loader (trigger-match, tier: Project > Org > Core)
                              |
                              v
        Context Injector (<=60K tokens)  --> Cost Router
                              |             (Haiku / Sonnet / Opus / Gemini,
                              v              by task difficulty)
      Fresh-context Subagent (implement -> self-verify -> commit)
                              |
                              v
   Verification (build / typecheck / lint / test / security / diff)
                     pass    |    fail
              +--------------+--------------+
              v                             v
   Handoff (.planning/HANDOFF.json    Temporal rollback -> sets status
              + AUDIT.jsonl)              "awaiting_regeneration"

The fail path is real but partial: bin/hindsight-injector.js rolls back .planning/ state and sets auto-state.json.status = "awaiting_regeneration" — verified, and hash-chained into the audit log like everything else. What is not currently true: nothing in bin/ reads that status back out to automatically re-trigger the wave (awaiting_regeneration has one writer, zero readers today) — regeneration after a rollback is a manual step, not a closed loop.

Four layers underlie this, top to bottom: Interface (.claude/, .agent/ — the 221 slash commands and hooks), Engine specs (.mindforge/ — 232 of the 354 skills plus 216 personas and config.json runtime knobs; the other 122 skills are extended-tier, under .agent/skills/), Execution (bin/, ~32K raw / ~25K stripped-of-comments LOC — the wave executor, governance, memory, and dashboard code that actually runs), and Persistence (.planning/ — STATE.md, the audit chain, resumable HANDOFF.json). Edit behavior in layer 2 where possible; layer 3 is the only place with real enforcement, per What is actually enforced above.

Documentation

Six categories, read in this order the first time:

Core workflow

Full command list: docs/commands-reference.md.

Dynamic Workflow Library

35 pre-built multi-agent workflow scripts that run via Claude Code's Workflow tool. Each fans out concurrent agents, synthesizes results, and returns structured output across 5 tiers (Research, Dev, Ops, Intelligence, Beast).

Discover: /mindforge:wf-catalog or node bin/mindforge-cli.js workflow list

Full, verified 35-workflow table by tier: docs/workflow-atlas.md.

Updates and migrations

Run /mindforge:update (add --apply to install) — see docs/upgrade.md for the full upgrade guide and fallback steps.

/mindforge:plugins list
/mindforge:plugins install mindforge-plugin-<name>
/mindforge:plugins validate

Token usage profiling

/mindforge:tokens --optimise

(--profile doesn't exist; real flags are --phase N, --session ID, --window short|medium|long, and --optimise.) See .mindforge/production/token-optimiser.md.

Installing and running /mindforge:plan-phase//mindforge:execute-phase costs real model calls once a subagent starts working — install itself does not (npx mindforge-cc@latest never calls a model). Spend is capped by [COST_HARD_LIMIT_USD] in MINDFORGE.md (default 25.00), enforced in code by bin/models/cost-tracker.js's preflight() before each call goes out — not a policy statement. The per-project token/cost reports from /mindforge:tokens and .mindforge/production/token-optimiser.md are heuristic estimates (file size / 4), explicitly logged with measured: false.

License

MIT © 2026 Sairam Ugge (GitHub: Sairam0000)