Claude Goal

Autonomous goal loops for Claude Code. Say "set up a goal" or use /goal-start; add a budget only when you want a limiter.

Docs  ·  Install  ·  Commands  ·  How it works  ·  Roadmap

---

> [!NOTE] > Public beta. The happy path is real-Claude-validated end-to-end across five acceptance scenarios; outside-user miles are still accumulating. File issues against anything surprising.

What is this

claude-goal is a Claude Code plugin that drives the agent through an autonomous loop until a goal is provably met. Type /goal-start "objective" or explicitly tell Claude "set up a goal for this and continue" and the agent self-iterates — each turn ends in a Stop-hook continuation that feeds the next prompt — until one of the following: the model passes its own completion audit, the work is genuinely blocked, a budget profile or raw token budget exhausts, a profile turn/time envelope exhausts, or you stop it.

It's a production-grade companion to Claude Code 2.1.139+'s built-in /goal. The native command is great for casual conditions; this plugin adds deterministic budgets, lifecycle controls, /compact recovery, persistence across restarts, and a tool-equipped evaluator subagent that verifies completion by running tests / reading files / checking exit codes — not by trusting the worker's self-narrative.

vs Claude Code's built-in /goal

| | Built-in /goal (2.1.139+) | claude-goal | |---|---|---| | Autonomous continuation | ✓ | ✓ | | Budget control | phrase in the condition ("or stop after 20 turns"), judged by the evaluator | deterministic profiles (quick, standard, deep, overnight, auto) or raw token caps, enforced by hooks | | Turn / time caps | prompt-level only | practical-unlimited by default; profile-sized hard caps, /goal-extend to raise | | Pause · resume · abandon | /goal clear only | /goal-pause, /goal-resume, /goal-abandon | | Mid-run objective update | new condition replaces the old goal | /goal-update keeps budget, accounting, and history | | Blocked-state recovery | evaluator can fail the goal as impossible | status=blocked, visible status, /goal-resume, resume_goal MCP; evaluator impossible verdict maps to blocked | | /compact recovery | n/a | accounting_uncertain flag + /goal-reconcile | | Restart persistence | condition restored on --resume; turn/token counters reset | SQLite preserves everything | | Background-task awareness | defers evaluation while background shells/subagents run | continuation fires regardless | | Completion judgment | small fast model (default Haiku), transcript-only — it cannot run tools | a plugin subagent that verifies with tools; its verdict is recorded and enforced by the MCP server before completed_by:"evaluator" is accepted |

Run them side-by-side. They don't collide.

Quickstart

# Install from the Nuko Nova Tools marketplace
/plugin marketplace add nuko-nova-dynamics/claude-marketplace
/plugin install claude-goal@nuko-nova-tools

# Start a goal
/goal-start "list all .ts files under src/ and print a line count for each"

# Natural-language goal starts work too. Claude derives the objective; no budget is set unless you ask.
set up a goal and refactor the auth module to use async/await, then keep going

# Or pick a run profile. Profiles set token, continuation, and wall-clock caps.
/goal-start "refactor the auth module to use async/await" --budget standard

Claude confirms the goal, begins working, and continues across turns without further prompting. When it decides the objective is met, it dispatches the evaluator subagent for verification, then calls update_goal and stops.

> [!TIP] > Omit --budget for unbounded token, turn, and wall-clock limits. Add a limiter only when you want one: quick (2M tokens, 50 turns, 2h), standard (10M, 200 turns, 8h), deep (100M, 1,000 turns, 24h), overnight (1B, 5,000 turns, 72h), auto for deterministic objective-based selection, or a raw token number for advanced tuning. Explicit natural-language goal starts follow the same rule: no explicit budget means unbounded. See docs/concepts/budgets for details.

How it works

sequenceDiagram
    autonumber
    participant User
    participant CC as Claude Code
    participant Hook as Stop hook
    participant DB as SQLite
    participant Eval as goal-evaluator subagent

    User->>CC: /goal-start "objective" or explicit goal request
    CC->>DB: create_goal (status=active)
    loop until done / blocked / paused / capped
        CC->>CC: assistant turn (tools, edits, reasoning)
        CC->>Hook: Stop event
        Hook->>DB: account worker + subagent tokens
        alt budget / cap exhausted
            Hook->>DB: status=budget_limited / paused
            Hook-->>CC: emit one-shot reason, stop
        else still going
            Hook-->>CC: {"decision":"block","reason":""}
        end
    end
    CC->>Eval: dispatch claude-goal:goal-evaluator
    Eval->>Eval: run tests, read files, check exit codes
    Eval-->>CC: {"verdict":"complete"|"incomplete"|"unverifiable"}
    CC->>DB: update_goal status=complete completed_by=evaluator
    Hook->>Hook: F5 — bounded retry to catch late completion-turn tokens

Mechanics in a paragraph each

Stop hook — continuation injection. After every turn scripts/stop.sh fires. If a goal is active, the hook emits {"decision":"block","reason":""} and Claude Code feeds that prompt back to the model. The continuation prompt — adapted from Codex's continuation.md — reminds the model of the objective and asks it to either keep working or call update_goal if done.

Token accounting. scripts/post-tool-batch.sh reads the session transcript JSONL after every tool batch, finds assistant messages past the last cursor, and sums input_tokens + cache_creation_input_tokens + output_tokens. Cache-read tokens are excluded. Parent-worker counts go to tokens_used; subagent counts go to subagent_tokens through per-agent cursors stored in subagent_token_cursors.

Completion — dual path. The worker can self-audit and call update_goal status:complete (completed_by: "self_update"). The continuation prompt also instructs the worker to dispatch the claude-goal:goal-evaluator subagent before declaring done. The evaluator runs in a fresh context with Bash + Read + jq + sqlite3 — it reads the objective from the DB, queries real state with tools, records its verdict through the record_verdict MCP tool, and returns {"verdict":"complete"|"incomplete"|"unverifiable"|"impossible"}. On complete, the worker calls update_goal completed_by:"evaluator" — which the MCP server accepts only when a recent recorded complete verdict exists for the active goal, so evaluator completion cannot be claimed without an evaluator run. On impossible, the worker marks the goal blocked with the evaluator's reason. Evaluator completion can also close a goal paused by accounting_error after /compact or a budget_limited goal when the token cap races with a verified-complete verdict. Self-update completion cannot bypass those guarded states. The two paths coexist; evaluator is preferred, self-audit is the fallback.

Blocked state. If the same blocker repeats across at least three consecutive continuation turns and no meaningful progress is possible without user input or an external-state change, the worker can call update_goal status:blocked blocked_reason:"...". Blocked goals stop auto-continuing, stay visible in status/history, and can be restarted with /goal-resume or the resume_goal MCP tool.

Budget / cap enforcement. At goal creation, a named profile expands into a token budget, continuation cap, and wall-clock cap; raw token budgets remain token-only. At the start of each Stop hook run, the hook checks (tokens_used + subagent_tokens) >= token_budget, `continuations_remaining

Commands

| Command | What it does | |---|---| | /goal-start "objective" [--budget quick\|standard\|deep\|overnight\|auto\|N] or explicit prose like "set up a goal for this" | Start a new goal. Omit --budget or omit any prose budget request for unbounded. Replaces any prior completed/abandoned goal for this session. | | /goal-status | Current goal, status, selected budget profile/source, worker + subagent tokens, continuations remaining, warnings. | | /goal-update "new objective" | Replace the objective of the active goal mid-run. Budget, token accounting, and history carry over. | | /goal-pause · /goal-resume | User pause/resume, or resume a blocked goal. | | /goal-abandon (/goal-stop) | Abandon permanently. Stops the auto-continuation loop. | | /goal-extend --add-continuations N | Raise turn cap on a continuation_cap-paused goal and resume. | | /goal-extend --add-hours N | Raise wall-clock cap on a wall_clock_cap-paused goal and resume. | | /goal-extend --add-tokens N | Raise token budget and resume a budget_limited goal. | | /goal-reconcile --accept-reset | Clear the accounting_uncertain flag set after /compact. Resets accounting to the current transcript cursor. | | /goal-doctor | Preflight self-test: SQLite, MCP connectivity, hook registration, shell deps. | | /goal-history [--all] [--format=json] | Tracked goals for this session, or --all for cross-session, sorted newest first. | | /goal-cleanup --list · --delete [--older-than HOURS] | Surface and reap orphaned goals left by /clear. |

Install

From the Nuko Nova Tools marketplace (recommended)

/plugin marketplace add nuko-nova-dynamics/claude-marketplace
/plugin install claude-goal@nuko-nova-tools

Updates land via /plugin update claude-goal once new versions are tagged.

From the release tarball (offline / air-gapped)

mkdir -p ~/.claude/plugins/local/claude-goal
tar -xzf claude-goal-v0.2.9.tar.gz -C ~/.claude/plugins/local/claude-goal
claude --plugin-dir ~/.claude/plugins/local/claude-goal

The tarball ships the prebuilt MCP dist/, pruned production dependencies under mcp/goal-server/node_modules/, LICENSE, RELEASE_NOTES.md, and ROADMAP.md. No package-manager install needed on the target machine beyond a Node 22 runtime.

From git (development / contributors)

git clone https://github.com/nuko-nova-dynamics/claude-goal.git
cd claude-goal
(cd mcp/goal-server && npm ci && npm run build)
claude --plugin-dir "$PWD"

Optional statusline

{
  "statusLine": {
    "type": "command",
    "command": "/path/to/claude-goal/statusline/status.sh"
  }
}

Renders one of:

• ◎ goal active (1.2M / 5M) — active with budget
• ◎ goal active (800K) — active, no budget
• ◎ goal paused (user) — user paused
• ◎ goal blocked — blocked until user input or external state changes
• ◎ goal unmet (budget exhausted) — budget hit
• ◎ goal achieved — complete

Known limitations

• macOS · Linux · WSL only. Native Windows triggers a fail-fast guard in stop.sh. WSL is best-effort and untested.
• Claude Code's Stop-hook block cap. Since CC 2.1.143 the harness force-ends a turn after 8 consecutive work-less Stop-hook blocks (CLAUDE_CODE_STOP_HOOK_BLOCK_CAP, default 8; 0 disables). Normal continuations reset the counter by doing tool work, so long runs are unaffected — but a goal whose turns produce prose only (e.g. update_goal keeps failing while the model keeps asserting done-ness) is cut short with the goal still active; any new user message resumes the loop. /goal-doctor reports the current cap. For long unattended runs, consider raising it in settings.json env.
• Auto-mode classifier may block update_goal. Use concrete, achievable objectives ("add a --verbose flag to the CLI" — not "do the task").
• /clear orphans the active goal. Use /goal-abandon before clearing, or /goal-cleanup to reap orphan rows later.
• Final-turn accounting is bounded. F5 catches late completion-turn transcript flushes with five 100ms retries. If Claude Code flushes completion usage after that window, a tiny residual undercount is still possible — the retry is intentionally bounded so Stop hooks do not hang.
• Plugin upgrade lifecycle is undocumented upstream. After updating plugin files, run claude restart to reload the MCP server. Hot-reload is not guaranteed.

Platform support

| Platform | Status | |---|---| | macOS | Tested | | Linux | Tested | | WSL | Best-effort, untested | | Native Windows | Fail-fast guard — not supported |

Develop

# Bats — hooks, skills, integration loops, release packaging, regression probes
bats $(find tests -name '*.bats' | sort)

# Vitest — MCP server, migrations, tools, token math, fixtures
npm --prefix mcp/goal-server test

Roadmap and design notes live in [ROADMAP.md](ROADMAP.md). Full docs at nuko-nova-dynamics.github.io/claude-goal.

Acknowledgments

The autonomous loop mechanism — Stop hook injecting a continuation prompt, model self-driving until a completion signal — is ported from OpenAI Codex's /goal feature (codex-rs). The continuation.md and budget-limit.md prompt templates in prompts/ are adapted from Codex's core/templates/goals/ with minor modifications for Claude Code's hook API.

License

MIT © Nuko Nova Dynamics

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: nuko-nova-dynamics
• Source: nuko-nova-dynamics/claude-goal
• License: MIT
• Homepage: https://nuko-nova-dynamics.github.io/claude-goal/

Install and usage instructions live in the source repository linked above.