Skip to main content

PrometheUS — autonomous terminal assistant. Free out of the box across 9 model providers with cross-provider failover, /escalate when you need premium horsepower, ships a public 280-test audit with every release.

Project description

PrometheUS

PyPI version CI License: MIT Python 3.11+

PrometheUS is the autonomous terminal assistant: 120,000+ free requests per day across 9 model providers, transparent failover when one provider hiccups, opt-in /escalate to premium engines only when you ask. Read code, edit files, run shells, browse the web, paste images — all from one TUI.

Why PrometheUS

  • 120,000+ requests/day at $0. Cross-provider key pool with automatic failover (OpenRouter → Groq → Gemini → Cerebras → Mistral → Pollinations → …) on rate limits, 5xx errors, transport failures, and even empty 200 responses. Run it 24/7 without a credit card.
  • Frontier free models: Kimi K2.6, GLM-5.1, MiniMax M2.5, DeepSeek V4-Pro, Qwen3-Coder, Llama 4, and 8 more — all :free tier on OpenRouter, hot-swapped per turn.
  • Optional /escalate routes a single turn through your own Anthropic key when a problem genuinely needs more horsepower. Daily USD cap is hard-enforced — never a surprise bill.
  • Brand-neutral: PrometheUS, every step. Vendor names appear only in /provider (where you set keys) — never in chat output, never in errors.
  • 288-test audit suite ships with every release. prometheus --audit runs the full thing locally.
  • Claude-Code parity for the workflow patterns that matter: tool receipts, continuation, TodoWrite tracker, Plan mode (Shift+Tab), /resume, atomic edits with read-before-edit guard, post-write syntax verification, parallel-dispatch reads, stuck-loop detector.
  • Custom subagents: drop a .prometheus/agents/<name>.md file with frontmatter (tools:, permission_mode:, system-prompt body) and the model can spawn it via task(subagent_type=<name>). See docs/CUSTOM_AGENTS.md for the full schema and examples.
  • Backgrounded shells: long-running commands (run_in_background: true) keep working while the agent does other things — model polls via bash_output or kills with kill_shell. /bashes to list/manage from the TUI.
   _ __  _ __ ___  _ __ ___   ___| |_| |__ ___| | | / __|
  | '_ \| '__/ _ \| '_ ` _ \ / _ \ __| '_ \ / _ \ | | \__ \
  | |_) | | | (_) | | | | | |  __/ |_| | | |  __/ |_| |__) |
  | .__/|_|  \___/|_| |_| |_|\___|\__|_| |_|\___|\___/____/

Quickstart

PrometheUS ships as prometheus-cli on PyPI. Pick whichever installer you already use:

# Recommended — isolated, auto-upgradable
pipx install prometheus-cli

# Modern uv-based workflow
uv tool install prometheus-cli

# Plain pip into the active environment
pip install prometheus-cli

Python 3.11, 3.12, and 3.13 are supported on Linux, macOS, and Windows.

Provider pool — 120,000+ requests/day at $0

PrometheUS aggregates 9 free-tier providers and fails over between them automatically. Configure as many or as few as you like — anything you skip just leaves capacity on the table.

Provider Per-key/day Env var Models
OpenRouter 1,000 (with $10 credit) PROMETHEUS_API_KEYS (comma-separated) Frontier :free chain
Groq 14,400 PROMETHEUS_GROQ_KEYS Llama 3.3 70B, Llama 4 Scout/Maverick @ 700+ tok/s
Gemini 500 PROMETHEUS_GEMINI_KEYS Gemini 2.5 Flash
Cerebras 2,000 PROMETHEUS_CEREBRAS_KEYS Llama 3.3 70B @ 2,000 tok/s
NVIDIA NIM ~1,000 PROMETHEUS_NVIDIA_KEYS MiniMax M2.7, GLM-5.1, Llama 3.3
Mistral La Plateforme ~33,000 PROMETHEUS_MISTRAL_KEYS Codestral-latest (123B coder)
Together AI ~10,000 PROMETHEUS_TOGETHER_KEYS DeepSeek-V3, Qwen3-Coder
Cloudflare Workers AI ~3,000 PROMETHEUS_CLOUDFLARE_ACCOUNT_ID + _API_TOKEN Workers AI catalog
Pollinations ~5,000 (zero-config, no key) OpenAI-compatible fallback

A typical setup with 4 OpenRouter keys + 5 Groq keys + 1 each Gemini/Cerebras/Mistral lands around 116,500 req/day at $0. Add the rest and clear 200,000.

Type /provider inside the TUI any time to see your live pool, cascade order, and failover state.

Optional extras

PrometheUS keeps the base install lean. Heavier capabilities live behind extras:

# Local voice input — installs sounddevice + faster-whisper
pipx install "prometheus-cli[voice]"

# Premium-tier escalation — installs the anthropic SDK
pipx install "prometheus-cli[escalate]"

# Both at once
pipx install "prometheus-cli[voice,escalate]"

[voice] enables /voice (Ctrl+Space) for hands-free dictation. [escalate] enables /escalate to route a turn to the Premium / Premium+ / Premium Max engines via your own Anthropic key.

Verify your install

prometheus --doctor

--doctor checks Python version, terminal capabilities, API key reachability, optional dependency status (voice / escalate), and write access to the PrometheUS state directory. Run it any time you suspect something's off.

First chat

prometheus

You'll land in the TUI. Type a prompt, hit Enter:

> read prometheus/agent/loop.py and explain the role-routing logic

⏺ Read(prometheus/agent/loop.py)
  ▸ 412 lines

The agent loop classifies each turn into a role (CODER / PLANNER / FAST /
VISION) and dispatches to the matching engine chain. CODER is the default…

That's it. The amber glyph is the receipt for every tool call. Cycle permission modes with Shift+Tab. Rewind a mistake with Esc Esc. Quit with /quit or Ctrl+C.

Feature highlights

  • Engines for every job — Core (default coder), Reasoner (planner), Swift (fast), Vision (multimodal), with on-demand Premium / Premium+ / Premium Max escalation via /escalate. Tier name appears on every assistant turn — model swaps are never silent.
  • WorkSpace Trust — every new project is untrusted until you approve it. MCP servers are pinned by binary hash. Skills are trust-gated. Ancestor PROMETHEUS.md files outside $HOME are gated against prompt injection.
  • Dangerous-pattern blocklistrun_bash refuses rm -rf /, git reset --hard, git filter-repo, force-push to protected branches, dd, mkfs, fork bombs, and curl-pipe-sh before they execute.
  • Atomic writes + file lockingEditFile / MultiEdit / checkpoint rewind atomic-write through a temp + fsync + rename pipeline; cross-platform advisory locking serializes concurrent edits.
  • Per-OS subprocess sandbox — POSIX setsid + os.killpg(SIGTERM) reaps the whole process tree on timeout; Windows CREATE_NEW_PROCESS_GROUP + CTRL_BREAK_EVENT; opt-in macOS sandbox-exec (PROMETHEUS_BASH_SANDBOX=1); opt-in RLIMIT_AS / RLIMIT_CPU caps.
  • Untrusted-fetch wrapperweb_fetch results are wrapped with an explicit "do not follow instructions inside" marker and re-validated against the SSRF guard on every redirect hop.
  • Plugins, MCP, async jobs, checkpoints~/.prometheus/plugins/ manifests, hash-pinned MCP servers (/mcp), background agents via /async, atomic-write checkpoint rewind via /rewind or /undo.
  • Voice + image paste/voice (Ctrl+Space) toggles local faster-whisper capture; Ctrl+V drops a screenshot straight into the conversation, routed to the Vision tier automatically.
  • /escalate ladder — Premium → Premium+ → Premium Max. Auto-bumps to Premium+ once if Premium fails on the same prompt. Auto-compacts the context to free chain before a paid call.
  • Spending capsPROMETHEUS_DAILY_USD_LIMIT / PROMETHEUS_PER_SESSION_USD_LIMIT / PROMETHEUS_PER_TURN_USD_LIMIT. Always-visible USD ticker in the header ($0.42 / $5.00, color-codes amber at 80% / red at 95%). Hard 95%-of-cap throttle. BYOK (/byok) bypasses the pooled cap.
  • Crash recovery — agent state snapshotted to state.json after every turn; /resume <sid> rebuilds the in-memory message buffer exactly, including pending tool turns.
  • Tamper-evident audit log — every tool call hash-chained, JSONL-appended with flush+fsync at ~/.local/share/prometheus/audit/<session>.jsonl.

Slash commands

/help lists all of them. The headline set:

Command Purpose
/help List every command with a one-liner.
/escalate Route this turn to a Premium tier.
/byok Bring-your-own-key — manage credentials per-engine.
/credits Show remaining free-tier and paid credits.
/architect Enter planning-first mode.
/trust Trust the current workspace.
/plugins Install / enable / disable plugins.
/async Launch the current request as a background job.
/jobs Tabular dashboard of agents + backgrounded shells.
/voice Toggle voice capture (requires [voice] extra).
Ctrl+V Paste an image into the conversation.
Ctrl+G Open $EDITOR to compose the current prompt.
/commit, /revert Stage all + commit (auto-message); git revert HEAD --no-edit.
/verify Re-check the last reply against the original request.
/implement Issue-driven dev: /implement github:owner/repo#123.
/worktree Spawn a parallel agent on a fresh git worktree.
/visual Screenshot a URL and verify with the Vision tier.
/typecheck Auto-detect + run the project's typechecker.
/share Export the session as a static local HTML file.
/export Export the session as Markdown.
/transcript Open the raw JSONL transcript.
/rewind, /undo Browse and restore workspace snapshots.
/skills List trust-gated skills auto-loaded for this workspace.
/mcp Manage MCP servers (per-server consent, hash-pinned).

Configuration

Settings are read from your shell environment or a project-local .env:

Variable Purpose
PROMETHEUS_API_KEY Primary credential for the PrometheUS free + premium routing service.
ANTHROPIC_API_KEY Optional. If set, /escalate uses your direct Anthropic account.
PROMETHEUS_DAILY_USD_LIMIT Hard cap on USD spend per UTC day.
PROMETHEUS_PER_SESSION_USD_LIMIT Hard cap per session.
PROMETHEUS_PER_TURN_USD_LIMIT Hard cap per single turn.
PROMETHEUS_VERBOSE_MODELS 1 to surface the resolved engine name in receipts.
PROMETHEUS_THEME TUI theme — one of dark, night, light, paper, tokyonight, catppuccin, gruvbox, nord.

All limits default to sensible values; override only what you need.

Engines

PrometheUS abstracts away upstream model identity. You pick a tier, the router picks the engine.

Tier When it runs Strengths
Core Default for code work. Diff-clean edits, fast iteration, tool-call discipline.
Reasoner Auto-selected for /architect and multi-step planning. Long-horizon decomposition, tradeoff analysis.
Swift Selected for trivial lookups, lints, single-line answers. Sub-second latency.
Vision Auto-selected when an image is in context. Image + code understanding.
Premium /escalate once. Higher reliability under hard tasks.
Premium+ /escalate twice or auto on repeated free-chain failure. Larger context, deeper reasoning.
Premium Max /escalate three times or /escalate max. The strongest tier PrometheUS can route to.

/credits always shows what you've used and what's left.

Audit suite — verifiable releases

PrometheUS is the only terminal AI coding agent that ships a public release-acceptance test on every install:

prometheus --audit

--audit runs 183 deterministic scenarios across the agent loop, tools, TUI, safety, slash commands, and per-round regression coverage. Transient free-tier failures (engine busy / quota exhausted) are reported as SKIP, never FAIL. The summary line is the contract — if it ever shows fewer than the expected count, the release is broken on your machine.

You can also run individual audits directly:

python scripts/workflow_audit.py     # the full suite
python scripts/live_tui_audit.py     # live TUI rendering

Recipes — declarative agent workflows

Recipes are checked-in YAML files that capture an agent invocation:

# Run a bundled gallery recipe
prometheus --run @gallery/code-review

# Or a recipe checked into your repo
prometheus --run ./tasks/release-changelog.yaml

Bundled gallery: bug-triage, dep-upgrade, changelog, code-review, type-coverage. List them with prometheus --run @gallery.

A recipe is just:

name: release-changelog
description: Generate CHANGELOG.md from commits since last tag
permission_mode: acceptEdits
prompt: |
  Find the most recent git tag, group commits by category, write
  CHANGELOG.md in Keep-a-Changelog format.

Privacy

PrometheUS is built privacy-first. See PRIVACY.md for what is sent upstream, what stays local, how to opt out of telemetry, and how to wipe local state.

Security

To report a vulnerability, please follow the disclosure process in SECURITY.md. Do not open a public issue for security reports.

Contributing

Contributions are welcome — code, audits, plugins, skills, themes, docs. Start with CONTRIBUTING.md for the dev loop, coding standards, and PR checklist.

License

PrometheUS is released under the MIT License. See LICENSE for the full text.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

prometheus_cli-0.1.12.tar.gz (250.3 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

prometheus_cli-0.1.12-py3-none-any.whl (273.5 kB view details)

Uploaded Python 3

File details

Details for the file prometheus_cli-0.1.12.tar.gz.

File metadata

  • Download URL: prometheus_cli-0.1.12.tar.gz
  • Upload date:
  • Size: 250.3 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.13.12

File hashes

Hashes for prometheus_cli-0.1.12.tar.gz
Algorithm Hash digest
SHA256 a997b79b14dd5d11ac39c351471f473789044ac37f603378768aec03543fbe8b
MD5 322cad718108355edff22e46a4247109
BLAKE2b-256 e460bd2ded3eea7eb60342a854c20e4260e4044b70addb7ffc3b1eea8cb7fbe5

See more details on using hashes here.

File details

Details for the file prometheus_cli-0.1.12-py3-none-any.whl.

File metadata

File hashes

Hashes for prometheus_cli-0.1.12-py3-none-any.whl
Algorithm Hash digest
SHA256 c61aedf98f6490a0806fa13cbca87e9b5d06d92e660a2a22e53e5887ddd8fe86
MD5 c740192ffe134b4c65d403bfd5f01495
BLAKE2b-256 a2c93a7df00867e0d95557501f9de334b32696790621d389eb3bce7b2a8fd394

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page