# Autonomi

> Autonomi is a platform for autonomous AI agent systems. Our flagship product, Loki Mode, turns GitHub issues and free-text tasks into tested, verified pull requests with a signed receipt.

## Company

- Name: Autonomi
- Website: https://www.autonomi.dev
- Contact: founder@autonomi.dev
- GitHub: https://github.com/asklokesh
- Discord: https://discord.gg/k8NpBhc5KA
- LinkedIn: https://www.linkedin.com/company/autonomi-dev-agents

## Flagship Product: Loki Mode

Loki Mode turns an issue backlog into merge-ready pull requests: tested, with evidence attached, on your own infrastructure, with any supported model. It ends every run with a signed receipt that states what was proven and what was NOT PROVEN. Everything is free to use; there is no paid edition, pricing or license key.

- Repository: https://github.com/asklokesh/loki-mode
- Install: npm i -g loki-mode (the npm package installs the Bun runtime the engine needs as an optional dependency; upgrade with loki self-update)
- License: BUSL-1.1 (source-available, free to use)
- Current Stable Version: v11.3.2 (2026-10-08)

### Loki 10 workflow

1. loki doctor : check your setup
2. loki owner/repo#N : build and verify a fix for a GitHub issue (GitLab and Jira issue URLs also work), open a PR
3. loki "<task>" : the same for a quoted free-text task; loki quick "<task>" for a small task with no PR
4. loki backlog owner/repo --all | --label X | --issues 1,2,3 : one run per issue, each in its own worktree, with a concurrency cap (settings in loki.yaml)
5. loki verify : re-check a run's signed receipt (CI-gate exit codes); loki keys export prints the public signing key for third-party verification
6. loki (or loki dashboard) : open the Control Plane in your browser; loki control serve runs it as a service
7. loki queue add <issue...>, loki queue list, loki queue run [--no-pr] : run queued issues overnight and write a morning digest

Flags: --no-pr, --deep, --provider <name>, --max-cost <usd>, --attempts N. A run prints a short summary: PR, verdict (VERIFIED, PARTIAL, ALREADY_SATISFIED, SPEC_CONFLICT or FAILED), NOT PROVEN, cost and time. Scope control is advisory: edits outside the plan's scope are kept and listed under NOT PROVEN, the receipt and the PR body; it never reverts an edit. Cost cap: none on a Claude subscription login (info line only), $100 per run by default on an API key (--max-cost or loki.yaml budgets.per_run). With a sealed receipt the exit code follows its verdict; exit 3 (BUDGET_STOP) only when no receipt was sealed.

Providers: claude (default, cost measured); codex, cline, aider and opencode are experimental.

### Also shipped

- v11.3.2 (npm latest, 2026-10-08): loki start --attempts N can now open a PR (the 11.3.1 --no-pr requirement is gone); loki undo <run-id> --plan verifies the receipt and prints what undoing the run does (applying it needs LOKI_UNDO=1); loki verify-pr <owner/repo#N> runs the fail-to-pass check for a PR in a fail-closed sandbox; loki issues run triages open issues and runs them through the queue, splitting a too-large issue into stacked slices (--no-split opts out); loki export --sarif writes findings.sarif (SARIF 2.1.0) per run; the Reviewer brief gains a decision block; the intent card can carry an Out of scope line, and loki plan --spec writes an editable spec that loki start --spec runs against; an advisory quota forecast line before loki start and loki queue run; loki memory forget <id>; loki status shows the last run's sealed verdict; receipts report unmeasured usage as NOT RECORDED, never zero or a partial sum, and record total time, the stage split, cache tokens and per-model cost. Off by default: a per-change AI-BOM (CycloneDX 1.7 ML-BOM) with AI commit trailers and a PR marker (LOKI_AI_MARKING), cross-vendor review by default (LOKI_XVENDOR_DEFAULT=1), OpenTelemetry gen_ai spans (LOKI_OTEL_GENAI=1, with a collector config for Grafana, Honeycomb and Datadog), the MCP Tasks extension (LOKI_MCP_TASKS) and the MCP 2026-07-28 stateless profile (LOKI_MCP_2026_07).
- v11.3.1 (2026-10-08): ten capabilities. A cost preview on the start line from per-shape history ("estimate: NOT AVAILABLE" with no history) and estimate against actual on the receipt; a mutation proof line on the seal and in the PR body (LOKI_MUTATION_STRICT=1 downgrade is off by default); an intent card at plan and in the PR (LOKI_CONFIRM asks, LOKI_INTENT_CARD=0 turns it off); a Reviewer brief in the PR body; loki start --attempts N (1 to 5 worktree attempts, the most executed passing checks wins, BLOCKED when none sealed; required --no-pr in 11.3.1, opens a PR since 11.3.2); memory learned from merged PR review comments, fenced as untrusted data; an overnight loki queue add|list|run with a morning digest (unmetered cost reported as NOT RECORDED); a Before / After PR section with screenshots when a UI route changed; provider failover (one retry on the next provider after a classified rate limit or outage); a supply-chain guard (a made-up package on the default registry fails the seal). Also a blocking CI gate under 2 minutes with full suites nightly, a docs cleanup, and every git call Loki makes runs with a token-free environment.
- Model router: opt-in with LOKI_ROUTER=1, experimental and off by default (Opus plans, Sonnet executes, Haiku only when Opus assigns a unit).
- v11.0.0: loki legacy and LOKI_ENGINE are removed; task and issue runs use the Loki 10 engine. With a sealed receipt the CLI outcome and exit code follow its verdict. The Control Plane gets an Overview with a NEEDS YOU inbox, runs grouped by issue and a first-run checklist; opt-in cross-provider review before seal can only downgrade a verdict.
- v10.9.1: the browser UI is the Control Plane (loki ui, bare loki, loki dashboard, loki web; default port 47821) and the classic dashboard UI is deleted: dashboard-ui/ and dashboard/static/ are gone, / on the old server (port 57374) returns 410 naming the Control Plane, and loki dashboard open prints a Control Plane hint. The Control Plane now serves the routes the classic dashboard answered (/v1/audit, checkpoints, tasks, memory, context, focus, session control, Completion Council, metrics, cost and fleet); 34 more legacy routes answer 410 Gone naming their replacement and 21 with no /v1 equivalent yet answer 501. It includes everything from 10.8.0 (published to npm next only) and 10.9.0 (never published): the enterprise Control Plane UI with a new-run composer, run thread with live log, diff and receipt, run controls, runs table, cost and usage, receipts with in-place verify, plans and traceability, work board, merge queue and PR risk, workspaces, integrations, models and providers, settings that edit loki.yaml, a Cmd+K command palette, notifications and an audit log; Helm, docker-compose and the AWS ECS module run the Control Plane and require a token; Agent SDK 0.3.288. Engine: with no model configured the claude session runs on the provider default model with no forced sonnet (Engine Law L1, never below raw); the live status line no longer truncates to about 20 columns on a pty with no size; opt-in LOKI_E10_FIX_RESUME=1 (default off) resumes the previous session in fix rounds; Project Model discovery, a model-built and schema-checked map of the repo cached in .loki/project.json, is behind a flag (default off, LOKI_E10_PROJECT_MODEL=1 enables it) with no consumers wired yet. The Docker image build is fixed.
- v10.8.0 reached npm next only and never became latest (its Docker build failed); everything in it, the enterprise Control Plane UI as the default browser UI included, ships in v10.9.1. v10.9.0 was never published.
- v10.7.1: scope control is advisory and never reverts an edit (out-of-scope files are kept and listed under NOT PROVEN, the receipt and the PR body); the PR body is filled from the run's recorded intake, plan and verify data instead of "not recorded"; implement and fix run on the run's own model by default (the sonnet cascade is opt-in with LOKI_E10_CASCADE=1); Control Plane ingest verifies the sealed event log and shows a run that fails verification as TAMPERED; test suites no longer flood a live Control Plane with fixture runs; loki control prune --repo OWNER/NAME removes unwanted runs (add --dry-run to preview).
- v10.7.0: speed path on by default (LOKI_SPEED=0 opts out); no dollar cap on a Claude subscription login (info line only), $100 default cap on an API key; one live status line in quiet mode; the Control Plane imports runs with zero setup and shows them live; loki docker trusts the mounted /workspace for git; default-on listeners bind loopback only
- Control Plane (default UI since v10.6.10): loki control serve runs it as a service (a token is required when bound beyond localhost); a container image (Dockerfile.control-plane) and a Helm chart ship in the repo. The classic dashboard UI is deleted in v10.9.1; port 57374 answers legacy API paths through a shim (default Control Plane port: 47821).
- loki verify <run> --pubkey <file> checks a receipt against a named public key; loki workspace run (multi-repo) works from an npm install; Jira and Linear issue refs work as input; the root GitHub Action runs Loki 10
- loki-seal: a Claude Code Stop hook that runs your repo's real test suite when the agent tries to finish and refuses done if tests newly fail or tests or CI config were deleted, skipped or weakened. Install: /plugin marketplace add asklokesh/loki-mode, then /plugin install loki-seal@loki-mode

### Previous engine (PRD-file starts and local-model providers)

loki legacy and LOKI_ENGINE are removed (v11.0.0). Task and issue runs (loki start "<task>", loki owner/repo#N, loki queue, --attempts) use the Loki 10 engine; a PRD-file start (loki start ./prd.md) and local-model providers (opencode) still run on the previous engine until they move to the Loki 10 engine. The previous engine provides the PRD-first flow, the RARV loop and the completion council.

### Installation Methods

1. npm: npm i -g loki-mode
2. Bun: bun install -g loki-mode
3. Homebrew: brew install asklokesh/tap/loki-mode
4. Docker: docker pull asklokesh/loki-mode

## Pages

- Homepage: https://www.autonomi.dev/
- Product (Loki Mode): https://www.autonomi.dev/loki-mode
- Use Cases: https://www.autonomi.dev/use-cases
- Use Case - Bring Your Own LLM: https://www.autonomi.dev/use-cases/bring-your-own-llm
- Use Case - Enterprise Governance: https://www.autonomi.dev/use-cases/enterprise-governance
- Use Case - PRD to Production: https://www.autonomi.dev/use-cases/prd-to-production
- Use Case - Free to Use: https://www.autonomi.dev/use-cases/open-source-pricing
- Documentation: https://www.autonomi.dev/docs
- Blog: https://www.autonomi.dev/blog
- About: https://www.autonomi.dev/about
- Labs: https://www.autonomi.dev/labs

## Documentation

- Introduction: https://www.autonomi.dev/docs/introduction
- Quick Start: https://www.autonomi.dev/docs/quick-start
- Installation: https://www.autonomi.dev/docs/installation
- CLI Commands: https://www.autonomi.dev/docs/cli-commands
- Configuration: https://www.autonomi.dev/docs/configuration
- Security: https://www.autonomi.dev/docs/security
- Troubleshooting: https://www.autonomi.dev/docs/troubleshooting

## Optional

- llms-full.txt: https://www.autonomi.dev/llms-full.txt