meris-agent 0.0.2

Model-agnostic coding agent with first-class Harness engineering

Meris Agent

Meris Cloud (hosted SaaS) is proprietary and not in this repo — see CLOUD.md.

Model-agnostic terminal coding agent — Harness-first, inspired by AtomCode, Claude Code, Codex, Cursor, and Kiro.

Agent = Model + Harness

North star: a self-evolving coding agent — the more you use it in a project, the better it fits your repo and habits. Evolution lives in the Harness (rules, skills, sessions), not in hoping the model remembers.

Meris ships a lean CLI (meris) plus a Harness layer you control: AGENTS.md, permissions, hooks, and Definition-of-Done sensors. Bring your own LLM (OpenAI-compatible or Anthropic).

Install

From PyPI:

pip install meris-agent
pip install "meris-agent[tui]"        # Textual TUI
pip install "meris-agent[anthropic]"  # native Anthropic provider
pip install "meris-agent[mcp]"        # MCP client
pip install "meris-agent[full]"       # all optional deps

From source:

git clone https://github.com/meris-agent/meris.git
cd meris
pip install -e .

Requires Python 3.11+. Meris does not bundle a model — set an API key (see below).

Quick start

cd your-project

meris version

# Pick a provider preset (OpenAI, DeepSeek, Gemini, GLM, Qwen, Groq, Anthropic, Ollama, …)
export MERIS_PROVIDER=deepseek   # or: openai | gemini | glm | moonshot | qwen | anthropic | ollama
export DEEPSEEK_API_KEY=sk-...   # env name depends on provider — run: meris models show deepseek

meris models list                # all presets + current env
meris init-harness .
meris doctor

meris ask "where is auth handled?"
meris plan "add rate limiting to /api/users"
meris run --approve "fix the failing test in tests/test_auth.py"

On Windows, use set VAR=value instead of export. Copy .env.example to .env for local overrides. Full vendor table: docs/MODELS.md. Hand this to new users: docs/USER_SETUP.md.

Commands

Command	What it does
meris ask "…"	Read-only Q&A over your repo
meris plan "…"	Write a task list to .meris/plan/tasks.md (no code changes)
meris run "…"	Full agent: read / edit / bash / git tools
meris run --approve "…"	Same as run, but confirm each write/bash
meris run --from-plan "…"	Implement the saved plan
meris run --no-sensor "…"	Skip post-edit DoD sensors
meris run --event-stream FILE	Append JSONL loop events (- = stdout)
meris exec "…" --json	Headless run; JSON result on stdout (CI/scripts)
meris review [--staged]	Read-only git diff review (markdown checklist)
meris harness check	Static harness lint (paths, imports, DoD helpers)
meris benchmark run	Run benchmark task suite
meris benchmark run --local-only	Offline local tasks (no API)
meris release check	Pre-tag readiness (pytest, mock benchmark, harness)
meris ratchet scan	Passive harness evolution from failures
meris ratchet digest	Active habit mining from sessions
meris models list	Show all provider presets + current env
meris models show <provider>	Show model & env details for a provider
meris doctor	Check API key, model, and harness files
meris init-harness .	Scaffold AGENTS.md, .meris/settings.yaml, PROGRESS.md

Harness (your project)

After meris init-harness, these files steer the agent in your repo:

File	Role
AGENTS.md	Project rules, layout, DoD
.meris/settings.yaml	Tool permissions, sensors, hooks, MCP, models
PROGRESS.md	Cross-session progress notes
.meris/skills/*.md	Optional domain skills
.meris/spec/*.md	Optional spec workflow (Kiro-style)

Templates live in templates/. Example vault harness: docs/examples/ainote-vault/.

Configuration

End-user onboarding (install → API key → init-harness → doctor): docs/USER_SETUP.md.

Environment variables (common):

Variable	Purpose
OPENAI_API_KEY / LLM_API_KEY	API key (OpenAI-compatible)
MERIS_BASE_URL	API base URL
MERIS_MODEL	Model name
MERIS_PROVIDER=anthropic	Use native Anthropic (ANTHROPIC_API_KEY)
MERIS_NATIVE	1/0 force on/off; unset + built meris-rs → auto native
MERIS_NATIVE_PROVIDER	0 to keep Python HTTP for chat while native harness on

Context budget in .meris/settings.yaml:

"context": { "maxMessages": 48, "maxTokens": 32000, "maxToolTokens": 2000 }

Permissions — allow/deny tool patterns (bash globs, path blocks). See templates/settings.example.yaml.

Workflows

Plan → run

meris plan "add session prune command"
meris run --from-plan "implement the plan" --approve

Spec checklist

meris spec init "rate limiting"
meris spec status
meris spec next --note "REST API only"
meris run --from-spec "implement checklist"

Benchmark (harness regression)

meris benchmark list
meris benchmark run

Ratchet (passive: failures → harness; active: habits from history)

meris init-harness .
meris ratchet learn --init             # project layout → proposals

# Passive (keep using these)
meris run --ratchet "your task"        # after run: profile + scan proposals
meris benchmark run                    # failures → events.jsonl; may auto-scan
meris ratchet scan
meris ratchet analyze --last-fail      # LLM proposals (needs API)
meris ratchet review                   # apply or reject interactively

# Active — repeated user themes → ask before writing harness
meris ratchet digest                   # rule-based; --llm for API pass
meris ratchet digest --dry-run
meris ratchet insights review          # confirm → proposal → apply
meris ratchet profile --show           # habits from approve_denied events

TUI

pip install "meris-agent[tui]"
meris tui --cwd . --mode run --approve

Interactive log + task input. Sessions panel on the left — Enter to resume, Ctrl+R latest, Ctrl+S refresh, Ctrl+L clear log. Enter or Ctrl+Enter sends a message.

MCP (external tools)

Add servers in .meris/settings.json:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/dir"]
    },
    "remote": {
      "transport": "sse",
      "url": "http://localhost:8080/sse"
    }
  }
}

pip install "meris-agent[mcp]"
meris mcp list

Tools register as mcp_{server}_{tool}; resources and prompts as mcp_{server}_read_resource / mcp_{server}_get_prompt.

Sessions & parallel

Every run saves to .meris/sessions/{id}.json:

meris session list
meris session show abc123
meris session resume abc123
meris run "task" --session-id my-id
meris session prune --keep 10

Run multiple tasks:

meris parallel "explain auth" "explain db layer" --mode ask -j 2
meris parallel "fix test A" "fix test B" --mode run --isolate

In run mode, the agent can call subagent_run for read-only sub-tasks with isolated context.

Hooks & skills

Tool hooks (Claude Code–style) in settings.json:

{
  "hooks": {
    "preToolUse": [
      { "matcher": "bash", "command": "echo pre $MERIS_TOOL_NAME" }
    ],
    "postToolUse": [
      { "matcher": "write_file|edit_file", "command": "echo post" }
    ],
    "onSave": [
      { "matcher": "*.py", "command": "python -m ruff check $MERIS_SAVED_PATH" }
    ],
    "onCommit": [
      { "command": "pytest tests/ -q" }
    ]
  }
}

Env: MERIS_TOOL_NAME, MERIS_TOOL_ARGS, MERIS_TOOL_RESULT, MERIS_HOOK_PHASE.

Skills — add .meris/skills/{name}.md; the agent loads them via load_skill.

Optional: IDE & Rust

Component	Docs
VS Code / Cursor extension	extensions/vscode-meris/ · LOCAL_SETUP
Rust core (meris-rs)	meris native build · NATIVE_BINARY · auto native when binary present

Windows one-shot dev setup: powershell -ExecutionPolicy Bypass -File scripts\setup-local.ps1

Contributing & development

Topic	Link
Documentation index	docs/README.md
Ratchet (meris ratchet scan)	docs/RATCHET_DESIGN.md
Harness architecture	docs/harness/architecture.md
Publish to PyPI	scripts/publish-pypi.ps1
Meris Cloud (hosted SaaS)	Proprietary — see CLOUD.md

Naming: Meris · CLI meris · PyPI meris-agent · env MERIS_* · Harness dir .meris/

Repo layout:

meris/
├── meris/              # Python agent package
├── meris-rs/           # Optional Rust harness core
├── extensions/vscode-meris/
├── templates/
└── tests/

License

MIT — see LICENSE.