Coder Eval — evaluate and benchmark AI coding agents and their skills

Playwright for coding agents — one declarative test file, any agent runtime, a real sandbox, and a pass/fail gate in CI.

Coder Eval (pip install coder-eval / uv tool install coder-eval) is an open-source, agent-agnostic framework for evaluating and benchmarking AI coding agents and their skills — built for benchmark authors, CLI builders, and skill builders — with sandboxing, reproducibility, and data-driven analysis. It runs a real agent — Claude Code, OpenAI Codex, Google Antigravity (Gemini), or OpenCode — in a sandbox against declarative YAML tasks, then scores the files and commands it actually produced. Changing harness is one field (agent.type); the tasks, criteria, scoring, telemetry, and reports stay the same.

Reach for it when you want to benchmark agents on your own domain tasks, test whether a skill triggers in the agent you ship for, A/B-test Claude Code vs. Codex vs. Gemini vs. OpenCode (or model vs. model, prompt vs. prompt), or gate CI on coding-agent quality. It is not a fixed leaderboard: unlike SWE-bench or SkillsBench, which rank models on a shared task set, you bring the tasks and you bring the scoring — weighted 0.0–1.0 criteria, a skill_triggered activation check, an A/B experiment layer, and per-tool cost telemetry, over whatever work you care about. See How it compares. 📚 Full docs: coder-eval.com/docs.

• Declarative YAML tasks with pinned dependencies and clear success criteria
• Sandboxed execution in isolated environments with resource limits
• Weighted, continuous scoring (0.0–1.0) with fractional credit and thresholds
• Many criterion types — from file checks to code similarity and LLM-graded rubrics
• Agent-agnostic by design — Claude Code, OpenAI Codex, Antigravity (Gemini), and OpenCode today; add your own harness through the plugin SPI
• Experiment layer — A/B agent configs (models, tools, prompts) side-by-side
• Full telemetry — every tool call, token counts, and cost, with real-time streaming

Watch the intro

CoderEval-YouTube-V1.mp4

🔊 Turn the sound on — GitHub's inline player always starts muted.

▶ Also on YouTube: Coder Eval: UiPath open-source framework to test AI Coding Agents — what the framework does, and how a run works end to end.

What you can do with it

• Benchmark coding agents — score an agent across a suite of tasks with weighted scoring and pass/fail thresholds
• Compare models & configs — A/B-test Claude vs. Codex vs. Gemini vs. OpenCode, model vs. model, tool-on vs. tool-off, prompt vs. prompt
• Evaluate skills — verify an agent actually engages a target skill (skill_triggered) and score skill-driven suites (SkillsBench-style), on whichever harness your users run
• Keep skills up to date in CI — re-validate your skills on every change or on a schedule; catch silent regressions when models, prompts, or the skills themselves drift
• Gate CI on agent quality — run the suite in GitHub Actions and fail the build on regressions
• Bring your own dataset — fan one task out over many rows for larger benchmark suites

Keeping skills fresh? Run Coder Eval as a scheduled GitHub Actions job so your skills are continuously re-evaluated against the latest model — a skill that quietly stops triggering surfaces as a failing criterion before your users hit it. See Tutorial 02 — Running Coder Eval in CI.

Quick Start

Prerequisites: Python 3.13+, uv 0.8+, and the runtime of at least one coding agent — plus that agent's own model credentials. Pick the agent you want to evaluate; two of the four ship with a Coder Eval extra, the other two are separate CLIs you install yourself:

Agent	agent.type	Runtime	Guide
Claude Code (default)	claude-code	brew install claude — separate CLI	Claude Code
OpenAI Codex	codex	uv sync --extra codex — the extra ships the Codex SDK + CLI	Codex
Google Antigravity (Gemini)	antigravity	uv sync --extra antigravity — the extra ships the harness binary	Antigravity
OpenCode (open-weight models)	opencode	npm install -g opencode-ai — separate CLI	OpenCode

The examples below use the default claude-code agent. Developed on macOS; CI runs on Linux.

git clone https://github.com/UiPath/coder_eval.git
cd coder_eval

uv sync                      # install the framework (add --extra codex /
                             # --extra antigravity for those runtimes)
cp .env.example .env         # then set ANTHROPIC_API_KEY — or skip that: an
                             # existing Claude Code login (`claude login`) is
                             # picked up automatically

uv run coder-eval plan tasks/hello_date.yaml   # validate (no tokens spent)
uv run coder-eval run  tasks/hello_date.yaml   # run your first evaluation
uv run coder-eval report runs/latest           # view the result

New here? Follow Tutorial 01 — Your First Evaluation.

Two more extras you only need on purpose: --extra dev adds the contributor toolchain (pytest, ruff, pyright, pre-commit — see CONTRIBUTING.md), and --extra uipath adds the in-host uipath SDK for local sandbox parity (public PyPI, no credentials). Without either, the framework still runs end-to-end.

Just want the CLI, without cloning? Install the published package — this is also what a CI job or another repo does:

uv tool install coder-eval    # puts the `coder-eval` CLI on your PATH,
                              # in its own isolated environment

uv tool install "coder-eval[codex,antigravity]"   # same, with agent extras
coder-eval --version                              # verify the install

To add it as a project dependency instead: uv add coder-eval or pip install coder-eval. In a real CI gate, pin to a specific released version so a harness upgrade can't silently move your results. (The example tasks/ live in this repo — clone it or point the CLI at your own task files.) See Tutorial 02 — Running Coder Eval in CI for the full setup.

Use inside Claude Code

Coder Eval evaluates any of the supported agents, and it also ships an authoring front-end for one of them: this repo is a Claude Code plugin marketplace, so the whole loop — scaffold a suite, author a task, check whether a skill triggers, read the results — runs inside Claude Code. The suites you author this way run on every harness:

/plugin marketplace add UiPath/coder_eval
/plugin install coder-eval@coder-eval

That adds six slash commands: /coder-eval:init, /coder-eval:check-skill, /coder-eval:task, /coder-eval:lint-tasks, /coder-eval:analyze and /coder-eval:ci. They drive the coder-eval CLI, so install it too (uv tool install coder-eval). See Claude Code Plugin.

Use as a GitHub Action

A composite action — on the Marketplace as coder_eval — runs coder-eval as a CI gate. It installs the pinned CLI, runs your tasks, writes a JUnit XML report, reports where its artifacts landed, and fails the step on any task failure:

- uses: actions/setup-node@v4      # the claude-code agent needs the Claude CLI…
  with: { node-version: '20' }
- run: npm install -g @anthropic-ai/claude-code

- uses: UiPath/coder_eval@v0       # …then run the gate (@v1 once 1.0.0 ships; @vX.Y.Z pins exactly)
  id: eval
  with:
    args: |
      tests/tasks/**/*.yaml
      --model
      claude-sonnet-5
    env: |
      ANTHROPIC_API_KEY=${{ secrets.ANTHROPIC_API_KEY }}

Eight inputs, and none of them is a coder-eval run flag. The CLI has 21; GitHub silently ignores an input the referenced tag does not define, so a forwarding input that is mistyped or newer than your pin yields a run that measured something else and still exits 0. A wrong CLI flag is a hard error. So flags and task globs all go through args, and an input exists only where the action does something with the value besides pass it along.

Input	Default	Purpose
args	—	Task paths/globs and every flag for coder-eval run, one argument per line, verbatim
version	pinned release	PyPI version, or local to install from the checkout
extras	—	coder-eval extras, composed into the install requirement (codex, antigravity,litellm)
extra-packages	—	Extra requirements installed into coder-eval's environment (--with), one per line
install-flags	—	Flags for uv tool install, one per line (--prerelease=allow, --extra-index-url …)
env	—	Credentials/backend passthrough: newline-separated NAME=VALUE pairs, exported for the run step only
working-directory	.	Directory every step of the action runs in
run-dir	runs/ci	Run directory; also where the reports are written

Outputs: run-dir, junit-path (<run-dir>/junit.xml) and run-md-path (<run-dir>/run.md). The action writes nothing to the job summary — a consumer that has to redact the report first cannot undo a write that already happened:

- if: always()
  run: cat "${{ steps.eval.outputs.run-md-path }}" >> "$GITHUB_STEP_SUMMARY"
- uses: mikepenz/action-junit-report@v5
  if: always()
  with:
    report_paths: ${{ steps.eval.outputs.junit-path }}

Credentials and backend config are the sole responsibility of env — a passthrough exported for the run step only (never written to $GITHUB_ENV, so it can't leak into later steps). Set whatever the run needs, Anthropic or not:

- uses: UiPath/coder_eval@v0
  with:
    args: tests/tasks/**/*.yaml
    env: |
      API_BACKEND=bedrock
      AWS_BEARER_TOKEN_BEDROCK=${{ secrets.BEDROCK_TOKEN }}

The step's exit code is coder-eval's own: non-zero on any failed task.

Agent runtime is the caller's responsibility. The action is agent-agnostic — it installs coder-eval but no coding-agent runtime, which is why the example above starts with actions/setup-node + npm install -g @anthropic-ai/claude-code: the default claude-code agent needs the claude CLI on PATH before the action runs. Swap those steps for your own agent's runtime as needed.

Security. Evaluated tasks execute agent-generated code. Do not run this action under pull_request_target with secrets exposed to untrusted fork PRs — use pull_request and gate on the same-repo condition, as this repo's own dogfood job does.

Telemetry

📊 Usage telemetry is on by default. coder-eval sends anonymous usage telemetry (command names, outcomes, counts, durations, an anonymous install id, platform info) to help improve the tool. It never captures prompts, file contents, or repo paths, and prints a one-time notice on first run. To disable it, set TELEMETRY_ENABLED=false in your .env or environment. See Usage Telemetry for details and how to route it to your own resource.

Documentation

Guide	What's in it
Tutorials	Step-by-step walkthroughs — start here
User Guide	Full CLI, configuration, output, and environment-variable reference
Task Definition Guide	The task-file schema — all criterion types, scoring, templates
Claude Code	Configuring and running the default Claude Code agent
Codex	Running the OpenAI Codex agent
Antigravity (Gemini)	Running the Google Antigravity / Gemini agent
OpenCode	Running the OpenCode agent on open-weight models
Run-Limit Parity	What each run_limits field means on every harness
A/B Experiments	Compare models / tools / prompts across the same tasks
Bring Your Own Dataset	Fan a single task out over a dataset
Dialog Mode	Evaluate agents in multi-turn conversation via a simulated user
Docker Isolation	The container sandbox driver, with custom images
CI Gate & GitHub Action	Run Coder Eval as a CI gate — the Marketplace Action, JUnit output, run reports
Claude Code Plugin	Install the Claude Code plugin — author, run, and analyze suites from inside the agent
Extending Coder Eval	Author a custom agent, criterion, or model pricing via the plugin SPI
Report Schema	Field-level reference for run.json / variant.json / task.json
How It Compares	vs. SWE-bench, SkillsBench, Harbor, OpenAI Evals, hand-rolled scripts

Repo doc	What's in it
CLAUDE.md	Architecture, key patterns, and extension points
CONTRIBUTING.md	Dev setup, quality bar, and how to contribute

How it compares

• vs. fixed benchmarks (SWE-bench, SkillsBench) — they score a canonical dataset; Coder Eval scores your tasks with continuous 0.0–1.0 weighted criteria (and can still wrap a fixed dataset via Bring Your Own Dataset).
• vs. large-scale / RL harnesses (Harbor) — Harbor targets scale and RL rollouts; Coder Eval targets weighted, skill-aware suites gated in CI.
• vs. model-output eval tools (OpenAI Evals) — they grade model text; Coder Eval runs a full agent in a sandbox and scores the files and commands it produced.
• vs. hand-rolled scripts — reproducible sandboxes, weighted criteria, cost/token telemetry, A/B experiments, and CI-ready pass/fail gates out of the box.

See the full comparison — with sources.

Task Definition

A task is a YAML file: a prompt, the agent config, a sandbox, and success criteria.

task_id: "hello_world"
description: "Create a Python script that prints Hello, World!"
initial_prompt: "Create hello.py that prints 'Hello, World!'"

agent:
  type: "claude-code"
  permission_mode: "acceptEdits"
  allowed_tools: ["Read", "Write", "Bash"]

sandbox:
  driver: "tempdir"
  python: {}

success_criteria:
  - type: "file_exists"
    path: "hello.py"
    description: "hello.py must be created"
  - type: "run_command"
    command: "python hello.py"
    timeout: 10
    description: "Script must execute successfully"

agent.type is the only harness-specific line: swap it for codex, antigravity, or opencode — or override it per run with coder-eval run … -D agent.type=opencode — and the same criteria score the same way. Tasks can omit the agent section entirely — defaults resolve from the experiment layer (experiments/default.yaml). For the full schema and every criterion type, see the Task Definition Guide; for what each run_limits field means on each harness, see Run-Limit Parity.

Tip: With the Claude Code plugin installed, use /coder-eval:task to scaffold a task from a natural-language description, and /coder-eval:analyze runs/latest to get improvement suggestions from a completed run.

Development

make install    # package + dev + [uipath] deps + pre-commit hooks
make verify     # format + lint + typecheck + test + coverage (CI equivalent)

Run make verify before pushing — it mirrors CI (80% coverage threshold). See CONTRIBUTING.md for the full workflow, commit conventions, and extension points (new criteria, new agents).

Known limits & non-goals

• Not a fixed benchmark or leaderboard — Coder Eval scores your tasks and ships example tasks, not a canonical scored dataset.
• Tasks execute real code — run untrusted tasks only under the container driver (see Docker Isolation); the tempdir driver is not a security boundary.
• Bring your own agent runtime and model credentials — Coder Eval never supplies model access, and it only ships an agent runtime where an extra says so (codex, antigravity); Claude Code and OpenCode are separate CLIs. Supply the runtime (see Quick Start) and the keys it needs — Anthropic, Bedrock, OpenAI, Gemini, or an OpenRouter key for open-weight models via OpenCode.
• Python 3.13+ only.

Support & security

• Security vulnerabilities — report privately via SECURITY.md; never open a public issue.
• Bugs & questions — open a GitHub issue.
• Everything else — reach the maintainers privately at coder-eval@uipath.com.

License

© 2026 UiPath. Licensed under the Apache License, Version 2.0 — see LICENSE and NOTICE.

Acknowledgments

Built with Pydantic, Typer, and Rich, on top of the harnesses it drives — the Claude Agent SDK, the Codex SDK, Google Antigravity, and OpenCode.