Loki Mode

The spec-driven autonomous builder with verified completion.

The free, source-available autonomous coding agent by Autonomi. Same Loki CLI, SDK, and MCP for everyone; the commercial editions for teams and enterprises are sold under the Autonomi brand (Autonomi Cloud, Autonomi Enterprise).

Hand it a spec. It does not accept "done" on an empty diff or failing tests.

Website | Documentation | Installation | Changelog | Purple Lab -- deprecated v7.44.0

Current release: v7.121.5

How it works: Drop a spec -- a PRD, GitHub issue, OpenAPI/JSON/YAML, or one-line brief. Loki Mode classifies complexity (run.sh:detect_complexity()), assembles an agent team from 41 specialized agent roles across 8 domains - prompt-defined specifications the orchestrator adopts per phase, with parallel review (blind council) and optional worktree streams on Claude Code, sequential on other providers - and runs autonomous RARV cycles (Reason - Act - Reflect - Verify, see run.sh:run_autonomous()) with 8 quality gates (see skills/quality-gates.md). Code is not "done" until it passes automated verification. Output is a Git repo with source, tests, configs, and audit logs.

Why Loki Mode?

• Spec-driven, autonomous, with a built-in trust layer -- Hand Loki a spec, walk away, come back to working code with tests. The full RARV-C closure loop (Reason - Act - Reflect - Verify - Close) runs until the work is actually done, not just attempted. The verified-completion evidence gate (skills/quality-gates.md) refuses any "done" claim on an empty git diff against the run-start commit, and blocks completion when tests run red, so "complete" means proven, not promised.
• A checklist verifier that is honest, not brittle -- Each completion checklist item is checked deterministically before the completion council will accept "done". The verifier speaks extended regex (grep -E) so real LLM-emitted patterns match instead of erroring, and it is runner-agnostic: it runs the project's own declared test command rather than assuming a fixed runner. Crucially, a check that cannot be established is reported as inconclusive (pending), never as a false pass and never as a false failure. rc == 0 alone is not a pass; a test check goes green only on a real "N passed" signal from the runner (v7.121.x).
• Production quality built in -- 8 quality gates (skills/quality-gates.md), blind 3-reviewer code review (run.sh:run_code_review()), anti-sycophancy checks
• Standalone verification: loki verify -- Run Loki's deterministic gates (build, tests, static analysis, secret scan, dependency audit) against any branch or PR diff, including code written by other agents or humans. CI-ready exit codes (0 VERIFIED, 1 CONCERNS, 2 BLOCKED), machine-readable evidence at .loki/verify/evidence.json. Inconclusive evidence is never reported as VERIFIED (v7.27.0).
• Living spec and pre-build interrogation -- loki spec locks a spec and detects drift deterministically (spec.lock, drift-report.json, and a SPEC_DRIFT finding in loki verify with CI exit codes), so you can tell when the build diverges from what was agreed. loki grill runs a Devil's-Advocate interrogation of the spec before you build, surfacing gaps and contradictions early (v7.28.0).
• Mid-flight model switching -- switch the model a live run uses from the dashboard (applies at the next iteration, current run only). A Fable tier lever exists in the CLI, dashboard, and override paths, but Claude Fable 5 is not yet available at the API, so selecting Fable currently collapses to Opus at every dispatch chokepoint and the loki plan quote reflects Opus accordingly. For every model lever (session pin, mid-flight override, architect pass) and every LOKI_MAX_TIER path, the loki plan quote, the dashboard's reported model, and the actual dispatched model agree, with the ceiling enforced (v7.31.0; Fable-to-Opus collapse v7.39.1).
• A calmer CLI -- the help surface is ~20 grouped workflow entries instead of a 70-command wall; merged commands live on as aliases that forward byte-identically with a one-line stderr pointer, so no script breaks (v7.31.0).
• Guided first build: loki quickstart -- four quick questions (setup check, one-line idea, template pick, plan review) and your build starts; pressing Enter through every step builds the sample Todo app. The plan step quotes the real cost/time estimate before anything is spent, and loki demo now confirms its estimate the same way. If no AI provider CLI is installed, Loki offers to install Claude Code (consent-gated, interactive terminals only) (v7.29.0).
• Live App Preview -- The dashboard embeds the locally-running app in an iframe so you can interact with it immediately during a build. Use loki preview (alias loki open) to print the URL and open it in your browser. Local-first: no hosted service, no vendor lock (v7.24.0).
• Compose-first fullstack -- When a spec needs more than one service (web + database + cache) Loki generates a 12-factor docker-compose.yml with healthchecks, depends_on wiring, env-var config, and a .env.example. The Live App Preview surfaces the web service URL (not a database port), and health reflects the web service's Docker healthcheck so a crashed app shows as crashed even when the database stays up. Single-service apps stay on a plain run command. All local-first, no hosted service (v7.26.0).
• Intelligent loki start -- For interactive foreground runs the dashboard auto-opens in the browser (cross-platform; skipped in CI, SSH-without-TTY, and piped runs; opt out with LOKI_NO_AUTO_OPEN=1). The completion summary shows "Your app is live at <url>" so you know exactly where to try what Loki just built. The autonomous loop passes Claude Code's --effort, --max-budget-usd, and --fallback-model on every iteration (each gated on CLI support and individual opt-out env vars) for better long-run unattended execution (v7.25.0).
• Cross-project memory -- Episodic/semantic/procedural memory with vector search; knowledge learned on one project surfaces on the next (v5.15.0+, see memory/engine.py)
• Self-hosted and private -- Your keys, your infrastructure, no data leaves your network
• Legacy system healing -- loki modernize heal archaeology/stabilize/isolate/modernize/validate phases (v6.67.0, see skills/healing.md)
• MCP server -- 34 tools (including ChromaDB code search) plus 3 resources and 2 prompts (mcp/server.py, with magic tools registered from mcp/magic_tools.py and the managed-memory tool from mcp/managed_tools.py). Of the 34, 33 are always available; loki_memory_redact is registered but only succeeds when LOKI_MANAGED_AGENTS=true and LOKI_MANAGED_MEMORY=true. Launch with loki mcp (bootstraps the Python MCP SDK on first run).
• Full-stack output -- Source code, tests, Docker Compose stacks (multi-service with healthchecks), CI/CD pipelines, audit logs
• Provider-agnostic -- runs on Claude, Codex, Cline, or Aider with automatic failover (loki-ts/src/runner/providers.ts); no vendor lock-in. Gemini CLI deprecated v7.5.18.
• Source-available (BUSL-1.1) -- Free for personal, internal, and academic use.

Loki does not lie about "done"

Most coding agents declare a task done by telling you so in a transcript. The transcript is the agent's own narration; there is nothing to check. Loki Mode takes a different stance: it does not call work done until the work is verified, and every build produces an Evidence Receipt you can re-verify yourself.

The receipt separates two things most tools blur together:

• Facts -- deterministic, non-LLM, and re-derivable by anyone: the git diff (base/head SHAs, file/insertion/deletion counts, a diff_sha256), the test command that ran with its exit code, the build command with its exit code, and each quality-gate verdict. A skeptic can recompute every one of these from the same repo state.
• Assessments -- AI judgments such as the review council's verdict. These are labeled explicitly as judgment, not proof, and never make the headline green on their own.

The receipt's headline is computed only from the facts:

• VERIFIED -- tests recorded a real command, ran, and exited 0; the diff is non-empty; nothing was skipped.
• VERIFIED WITH GAPS -- some facts checked out, but something was not run or was inconclusive. Every gap is listed by name, so silence never reads as a pass.
• NOT VERIFIED -- a test, build, or gate ran and failed (or there was nothing to verify).

This is honesty-of-done, not a claim of perfection. The receipt proves the completion claim is backed by deterministic evidence and is independently re-checkable; it does not claim the generated code is bug-free.

Verify it yourself

Receipts are written to .loki/proofs/<run_id>/ automatically at run completion (opt out with LOKI_PROOF=0). Inspect and re-check them with loki proof (aliased as loki receipt):

loki proof list              # every receipt: run id, time, council verdict, cost, files
loki proof show <id>         # the full proof.json (facts, assessments, honesty)
loki proof verify <id>       # re-check the receipt against the repo (exit 0 clean, 1 tamper/drift)

loki proof verify does two independent checks and prints the result as JSON:

• Tamper check -- recomputes the receipt's integrity hash and compares it to the recorded one. If anyone edited the receipt after it was written, hash_ok is false.
• Drift check -- re-runs the diff from the recorded base SHA against the current repo and compares the file/insertion/deletion counts and diff_sha256 to what the receipt recorded. If the repo no longer matches, diff_drift is true.

A clean receipt prints "ok": true and exits 0. A tampered or drifted receipt exits 1. When a check cannot run (for example a receipt with no recorded base SHA), the verifier reports it as unverifiable rather than passing it silently.

{
  "hash_ok": true,
  "diff_drift": false,
  "gpg_ok": "n/a",
  "degraded": [],
  "reason": "",
  "ok": true
}

You can share a receipt as a self-contained HTML page (loki proof open <id>), or publish it as a GitHub Gist with loki proof share <id> (opt-in; the page is redacted before it leaves your machine). An optional, off-by-default GPG detached signature (LOKI_PROOF_GPG_KEY) lets a third party confirm the receipt came from you.

Proven PR

When Loki opens a pull request, the PR body includes the Evidence Receipt summary, so a reviewer does not have to take the agent on faith. It shows the honest verdict (VERIFIED / VERIFIED WITH GAPS / NOT VERIFIED), the key facts (diff hash, tests, secure-gate, cost), and a "verify this yourself" line: loki proof verify <id> against the recorded base SHA. A green claim appears only when the receipt's own headline is VERIFIED. This is on by default whenever Loki opens or advises a PR; opt out with LOKI_PROVEN_PR=0.

An optional advisory status check (loki: verified-completion) maps the verdict to a GitHub check-run. It is opt-in (LOKI_PROVEN_PR_CHECK=1) and can never block a merge on its own. To make verified-completion blocking, add it as a required status check in your repository's branch-protection settings.

Get Started in 30 Seconds

Zero install, zero key, zero spend. See a real Evidence Receipt right now:

npx loki-mode tour                             # replay a real sample receipt (no install, no key, no spend)

Ready to build? Install and run the guided first build:

bun install -g loki-mode                       # install (npm/brew/Docker also work, see below)
loki quickstart                                # one guided command: your first real build

loki quickstart is the recommended way to start. It asks a few quick questions (setup check, one-line idea, template pick, plan review), quotes the real cost/time estimate before anything is spent, and then runs the build. Pressing Enter through every step builds the sample Todo app.

Want a taste with zero key and zero spend first? Run:

loki demo --offline                            # replay a real sample Evidence Receipt (no key, no spend)

It replays a real past build's Evidence Receipt so you can see Loki's honest verdict (VERIFIED / VERIFIED WITH GAPS / NOT VERIFIED) in seconds, with no provider, no API key, no spend, and no network. It is a sample (a replay), not a verdict on your own code.

Prefer the explicit, scriptable path? Scaffold a PRD and run the build yourself:

loki init my-app --template simple-todo-app    # scaffold a starter PRD
cd my-app && loki start prd.md                 # autonomous build from the spec

One thing to know first: Loki drives a separate coding-agent CLI (Claude Code is the recommended one) and needs it plus a couple of common tools on your PATH. Run loki doctor any time and it tells you exactly what is present and what is missing, with a copy-pasteable install command for each gap.

loki doctor                                    # check your setup before the first build

Required:

• An agent provider CLI: Claude Code (claude, Tier 1, recommended and E2E-verified - the provider Loki Mode is built for). Codex, Cline, and Aider are supported as experimental providers (wiring in place; not yet E2E-verified by us). Loki cannot run a build without one of these installed and authenticated.
• Python 3.10+ (python3) for the dashboard, memory system, and orchestration helpers.
• Git 2.x (git) for checkpoints and worktrees.
• curl for installation and network calls.

Recommended:

• Bun 1.3.0+ (bun) for the fast runtime (the recommended install path above installs it).
• Node.js 18+ and npm if you install via npm instead of Bun.
• jq for nicer JSON handling in shell flows.
• Docker if you want Loki's App Runner to run containerized projects, or to run Loki itself from the published image.

You also need credentials for whichever provider you use (for Claude Code, an authenticated claude login or ANTHROPIC_API_KEY). loki doctor flags a missing or unauthenticated provider as the first thing to fix.

If you do not have Bun yet:

curl -fsSL https://bun.sh/install | bash       # macOS / Linux (or: brew install oven-sh/bun/bun)

Other spec sources work the same way:

loki start owner/repo#123                       # a GitHub issue
loki start ./openapi.yaml                        # an OpenAPI/YAML spec

Or skip scaffolding and go straight to a quick task:

loki quick "build a landing page with a signup form"

Other install methods (all work, all keep working):

Upgrading:

loki self-update                  # upgrade in place via current manager
loki self-update --to bun         # switch from npm/brew to Bun
loki self-update --check          # show current install path + manager

loki self-update auto-detects which package manager installed loki and runs the right upgrade. If you installed via npm and want to switch to Bun (recommended for v8.0.0 forward-compat), loki self-update --to bun does the migration in one command (installs via Bun first, then uninstalls the npm copy).

See the Installation Guide for the long form.

Loki Mode runs a dual runtime by deliberate design: the battle-tested Bash engine is the stable core (the autonomous loop, quality gates, and completion council stay on it; it receives bug fixes and hardening), and new product surfaces are built TypeScript/Bun-first as modules that wrap the engine rather than reimplement it. An earlier plan to make v8 Bun-only has been superseded by this stable-engine approach: rewriting the verified trust layer would risk the exact guarantees this product exists to provide, for no capability gain. Bash support is not going away.

What ships today:

• Commands routed to the Bun runtime when bun is on PATH (the router lives in bin/loki): version, --version, -v, status, stats, doctor, provider (covers provider show and provider list), memory (covers memory list and memory index), rollback, kpis, and internal.
• Every other command continues to execute on the existing Bash CLI (autonomy/loki), including the autonomous loki start / loki run loop which remains the Bash orchestrator (autonomy/run.sh).
• If bun is not on PATH, the shim falls through to Bash silently. Existing users without Bun installed see no behavior change.

Rollback flag:

Force every command to take the legacy Bash path:

LOKI_LEGACY_BASH=1 loki <cmd>

This is the documented escape hatch for any user who hits a regression on the Bun route. The Bash path remains the source of truth through Phase 5.

Phase 6 (planned, calendar TBD):

The next major release sunsets the Bash runtime entirely. There is no firm calendar date. Users who need to stay on the Bash route should pin the last v7.x release.

Cost:

• Adds a Bun runtime dependency (Bun 1.3.0 or newer recommended; the shim works as long as bun resolves).
• Adds a Bun toolchain to the system (Bun itself is roughly 50 MB installed via brew install or the official curl installer). The published loki-ts/dist/loki.js bundle inside the npm tarball is approximately 152 KB.
• Speedup on the ported commands is measured in .loki/metrics/migration_bench_soak.jsonl and analysed in ADR-001. Recorded soak results show roughly 3x to 5x faster execution on the ported commands (per-command range 2.9x to 5.0x); treat as indicative, not contractual.

More:

• UPGRADING.md -- per-version upgrade and rollback guidance.
• ADR-001: Runtime Migration -- design rationale and phase definitions.

See the full Installation Guide.

A "spec" is whatever you hand loki start. Loki auto-detects the format and normalises it before the RARV loop. A Markdown PRD is one form of spec; the table below lists every input the CLI accepts.

All formats land in the same RARV pipeline and pass the same 8 quality gates (skills/quality-gates.md).

What You Can Build

What To Expect

RARV Cycle

Every iteration: Reason (read state) - Act (execute, commit) - Reflect (update context) - Verify (run tests, check spec). Failures trigger self-correction.

Core Workflow

41 Agent Roles

8 domains: engineering, operations, business, data, product, growth, review, orchestration. These are prompt-defined role specifications the orchestrator adopts per phase, auto-composed by PRD complexity; parallelism comes from the blind review council, the adversarial reviewer, and optional git-worktree streams on Claude Code, sequential on other providers.

Agent Types

8 Quality Gates

Static analysis, test suite (pass/fail), blind 3-reviewer code review with severity blocking, anti-sycophancy Devil's Advocate, mock-integrity detection, test-mutation detection, documentation coverage, and Magic Modules debate. Backward-compatibility is a conditional healing-mode auditor (not one of the 8). Code does not ship until all gates pass.

Quality Gates

Memory System

3-tier architecture: episodic (interaction traces), semantic (generalized patterns), procedural (learned skills). Vector search optional.

Memory Architecture

Dashboard

Real-time monitoring, agent status, task queue, WebSocket streaming, and Live App Preview (embedded iframe of the running app with Refresh/Open/Restart toolbar). Auto-starts at localhost:57374.

Dashboard Guide

Enterprise Layer

TLS, OIDC bearer-token validation (the foundation for SSO; browser SAML login is roadmap), scoped RBAC, OTEL tracing, policy engine, audit trails. Activated via env vars. See Enterprise Identity Roadmap.

Enterprise Guide

Purple Lab

[DEPRECATED in v7.44.0] Purple Lab (loki web, port 57375) is deprecated. The local build monitor and project dashboard are now the dashboard (auto-launched by loki start, http://localhost:57374). For the hosted/commercial platform, see Autonomi Cloud.

The historical feature set (platform pages, Monaco IDE workspace, AI chat panel) lives on in the dashboard and in Autonomi Cloud. loki web still invokes the old binary for backward compatibility but will be removed in a future major version.

Loki Mode vs. Alternatives

Loki Mode is the only platform that is fully self-hosted, source-available (BUSL-1.1), and includes automated quality verification. Your code, your keys, your infrastructure.

Loki's autonomy and quality loop are the product; the underlying coding CLI is swappable. Loki runs on any of the providers below so you are never locked to one vendor.

Status legend: "E2E-verified" means we run real spec-to-code builds on it ourselves. Claude Code is the primary, fully supported provider and the one Loki Mode is built for; it gets full features (subagents, parallelization, MCP, Task tool). "Experimental" means the wiring is in place but we have not produced an end-to-end verified build ourselves; treat as community-tested. Experimental providers run sequentially. Auto-failover switches providers when rate-limited. See Provider Guide.

CLI Reference

Run loki --help for all options. Full reference: CLI Reference | Config: config.example.yaml

Loki Mode's accuracy and autonomy behaviors are default-on. Each is an opt-out escape hatch, not a setting you have to discover. The most relevant knobs from the v7.41.x accuracy/autonomy hardening:

This is a subset. See the wiki for the full env-var reference and the RARV-C closure knobs (LOKI_INJECT_FINDINGS, LOKI_OVERRIDE_COUNCIL, LOKI_AUTO_LEARNINGS, LOKI_HANDOFF_MD).

Loki Mode integrates with the BMAD Method, a structured AI-driven agile methodology. If your project uses BMAD for requirements elicitation, Loki Mode can consume those artifacts directly:

loki start --bmad-project ./my-project

The adapter handles BMAD's frontmatter conventions, FR-format functional requirements, Given/When/Then acceptance criteria, and artifact chain validation. Non-BMAD projects are unaffected -- the integration is opt-in via --bmad-project.

See BMAD Integration Validation.

Enterprise features are included but require env var activation. Self-audit: 35/45 capabilities working, 0 broken, 1,314 tests passing.

export LOKI_TLS_ENABLED=true
export LOKI_OIDC_PROVIDER=google
export LOKI_AUDIT_ENABLED=true
loki enterprise status

Enterprise Architecture | Security | Authentication | Authorization | Metrics | Audit Logging

Self-reported results from the included test harness. Verification scripts included for reproduction.

See benchmarks/ for methodology.

11 slides: Problem, Solution, 41 Agents, RARV Cycle, 8 Quality Gates (HumanEval 98.78%), Multi-Provider, Enterprise Hardening (Live App Preview), Full Lifecycle

Download PPTX

Limitations

What "autonomous" means: The system runs RARV cycles without prompting. It does NOT access your cloud accounts, payment systems, or external services unless you provide credentials. Human oversight is expected for deployment, API keys, and critical decisions.

Research Foundation

Practitioner insights: Boris Cherny, Simon Willison, HN Community

Full Acknowledgements -- 50+ papers and resources

Contributing

git clone https://github.com/asklokesh/loki-mode.git && cd loki-mode
npm install && npm test              # CLI + Node test suites
python3 -m pytest                    # Python test suite

See CONTRIBUTING.md for guidelines.

License

Business Source License 1.1 -- Free for personal, internal, academic, and non-commercial use. Converts to Apache 2.0 on March 19, 2030. Contact founder@autonomi.dev for commercial licensing.

Autonomi | Documentation | Changelog | Comparisons

Star History