Evener

A non-interactive coding agent. Give it a prompt, it does the work.

Evener uses the LLM's native tool-calling to read files, write files, run commands, and search code in a loop until the work is complete. It supports OpenAI, Anthropic, and Google models.

For how the code is organized — modules, layout, and the build workspace — see docs/architecture.md. For the runtime contracts subagents, plugins, hooks, and helpers operate under (capability policy, lifecycle hooks, helper isolation, lineage), see docs/subagent-runtime-contracts.md; for background jobs and the job-control tools, see docs/job-control.md. To confine a session's file, process, and network access with --sandbox, see docs/sandboxing.md.

Build

make build

Build the standalone one-shot client (no agent loop):

make build-llmcall

Build the multi-session web orchestrator:

make build-hub

Install

Install the latest release on Linux x64 or macOS Apple silicon:

curl -fsSL https://raw.githubusercontent.com/prime-radiant-inc/evener/main/install.sh | sh

The release installer downloads the matching GitHub release archive, installs evener, evener-hub, evener-tui, evener-doctor, and evener-migrate under ~/.local/share/evener/bin, and symlinks them into ~/.local/bin.

Install a specific tagged release:

curl -fsSL https://raw.githubusercontent.com/prime-radiant-inc/evener/main/install.sh | env EVENER_INSTALL_VERSION=v1.2.3 sh

Install the latest successful build from main:

curl -fsSL https://raw.githubusercontent.com/prime-radiant-inc/evener/main/install.sh | env EVENER_INSTALL_VERSION=snapshot sh

Override the install prefix, using sudo for system-owned paths:

curl -fsSL https://raw.githubusercontent.com/prime-radiant-inc/evener/main/install.sh | sudo env PREFIX=/usr/local sh

From a source checkout:

make install

This builds evener, evener-hub, evener-tui, evener-doctor, and evener-migrate, installs the binaries under ~/.local/share/evener/bin, and symlinks them into ~/.local/bin. make install-home is an alias for the same layout.

System-style install:

sudo make install-system

This uses the same layout under /usr/local by default. Override PREFIX to stage elsewhere.

The installer only installs binaries and symlinks. Runtime/config directories are created by Evener when the relevant binary runs.

Verify the installed commands with:

evener --version
evener-tui --help
evener-doctor --help

Upgrade installed binaries manually:

evener upgrade

evener upgrade follows the binary's install channel: release builds upgrade to the latest release, and snapshot builds upgrade to the latest successful main build. You can override the target with evener upgrade release, evener upgrade snapshot, or a tagged version such as evener upgrade v1.2.3. The TUI and web UI also expose a manual /upgrade command that calls through the hub and uses the same channel tracking.

On first use, Evener creates:

• ~/.evener/run for live daemon rendezvous files.
• ~/.evener/auth-token for the local Hub/TUI bearer token.
• ${XDG_STATE_HOME:-~/.local/state}/evener/projects/<project-id>/ for saved per-project session state. The project ID is readable (derived from the canonical project path) and ends with a 10-character base62 suffix.
• ${XDG_CONFIG_HOME:-~/.config}/evener/skills for standalone user skills.
• ${XDG_CONFIG_HOME:-~/.config}/evener/plugins for user plugins.

The user skill and plugin directories are extension roots; installing Evener does not automatically enable their contents. Add standalone skill paths to skills_dirs and plugin paths to plugin_dirs in ~/.evener/launch.toml, or pass them with the corresponding CLI flags for a single run. Plugin-contained skills live under that plugin and become available through the plugin path.

Provider credentials are not created by install. Configure them through the Hub or TUI credentials UI, ~/.evener/credentials.toml, provider environment variables such as OPENAI_API_KEY, or OpenAI OAuth. See docs/environment.md for the complete environment variable reference.

Migrating from Serf

Evener was previously named Serf. If you have an existing Serf install, run evener-migrate once, before your first Evener launch:

evener-migrate

It moves ~/.serf to ~/.evener, ${XDG_CONFIG_HOME:-~/.config}/serf to .../evener, ${XDG_STATE_HOME:-~/.local/state}/serf to .../evener, and any per-project .serf directory (in the current directory or a Git ancestor) to .evener. It refuses to overwrite a destination that already exists, so it's safe to run more than once. Pass --dry-run to preview the moves first, or --verbose to see every path it checked.

Re-running it is also how you repair a machine that migrated before a fix landed: after moving each root (or finding it already moved), evener-migrate walks the destination and rewrites any leftover absolute references to the old serf path it finds inside text files there — for example a plugin marketplace registry that still points git pull at .../config/serf/plugins/marketplaces/<name>. It skips binaries and anything inside a Git working tree (a plugin marketplace clone, or a nested project checkout), so it's safe to run against a live install; files with nothing to rewrite are left untouched.

If you skip this, the first Evener binary you run creates a fresh, empty ~/.evener (and XDG config/state directories) — and once that empty directory exists, evener-migrate treats it as already migrated and skips it, leaving your old Serf data stranded with no further warning. To prevent that, Evener refuses to start when it finds a legacy Serf directory with no matching Evener directory yet, and its error names the path and points back to evener-migrate.

Usage

evener --model <provider/model> [flags] <prompt>

The prompt can be passed as arguments or piped via stdin:

# Prompt as arguments
evener --model openai/gpt-5.2 "add input validation to the signup handler"

# Prompt piped via stdin
echo "refactor auth to use JWT" | evener --model anthropic/claude-opus-4-6

For a shell command that must outlive the current Evener session, use detached mode. Detached commands are not jobs; redirect their own logs when needed:

{"command":"long-task > /tmp/long-task.log 2>&1","mode":"detached"}

llmcall (One-Shot LLM Client)

This repo also includes llmcall, a minimal CLI wrapper around the unified llm library for single “throwaway” calls.

Properties:

• Exactly one LLM call (no agent loop).
• Tool calls are forbidden (tool_choice=none). If the model returns tool calls, llmcall fails.
• No system prompt by default. You can optionally provide one, or force-disable with --no-system.

Build:

make build-llmcall

Examples:

./llmcall --provider openai --model gpt-5-mini-2025-08-07 "Write a haiku about build pipelines."

# JSON mode: parses and re-prints as JSON (fails if output isn't valid JSON)
echo 'Return JSON: {"ok": true}' | ./llmcall --provider openai --model gpt-5-mini-2025-08-07 --format json

# JSON Schema mode: enforces + validates structured output
./llmcall --provider openai --model gpt-5-mini-2025-08-07 --schema /path/to/schema.json "Return an object matching the schema."

llmcall resolves provider/model from env if omitted:

• LLM_PROVIDER or EVENER_PROVIDER
• LLM_MODEL or EVENER_MODEL

Provider and model

Evener takes a provider-qualified model in one value: --model <provider/model>. Providers: openai, anthropic, google, minimax, openrouter, openrouter-anthropic, kimi, glm, ollama.

Use --model or set EVENER_MODEL to the same provider/model format.

For local models via Ollama, see docs/ollama.md.

Environment variables

See docs/environment.md for the complete list. Common variables:

Flags

Structured output

Pass --output-schema <json> to replace the communicate tool's output field schema with your own. The flag takes an inline JSON string (file paths are not supported).

evener --model openai/gpt-5.2 \
  --output-schema '{"type":"object","properties":{"plan":{"type":"string"}},"required":["plan"],"additionalProperties":false}' \
  "Draft a one-paragraph plan for fixing the flaky test."

The supplied schema replaces output wholesale — the default