Practonaut

Practonaut

A local practice gym for AI agents on real apps. Demonstrate a task once, let any MCP agent practise it, and get an independent, reproducible score — one npx command, no VM, no Docker.

Practonaut turns a real application — today Excalidraw — into a practice environment. A task is a start drawing plus a goal. You, or an agent over MCP, change the diagram one action at a time; each action is applied to a real, pinned Excalidraw engine and read back before it counts. When the attempt ends, a separate checker decides whether the goal is met. People and agents work on the same session at the same time, so you can watch an agent in your browser: the history marks each change as yours or as the agent's (*Agent · claude-code*).

Quick start in 60 seconds

Requires Node.js 22.18 or newer (macOS, Linux or Windows).

npx practonaut                 # opens the example task in your browser at once

npx practonaut setup # once: the Chromium headless shell for the real Excalidraw engine (about 105–130 MB)

Watch a demonstration in the Player solves the example task once: actions land on the board, the real Excalidraw confirms each one, and the result check judges the end. It is never your result; Stop or Escape ends it. Before setup everything works model-only: actions are not confirmed in the real app, and a banner says so and names the download.

Let an agent practise (Claude Code shown; any MCP client works):

claude mcp add practonaut -- npx -y practonaut mcp --task excalidraw/connect

On native Windows (not WSL) npx is a .cmd file that an MCP client cannot start directly; add cmd /c in front: claude mcp add practonaut -- cmd /c npx -y practonaut mcp --task excalidraw/connect.

Score it on a task, with a fresh session per run and an independent check:

npx practonaut bench --agent claude-code --tasks excalidraw/connect

More in getting started.

Create your own task

Draw the start and the goal in Excalidraw, then:

npx practonaut tasks from-scenes start.excalidraw goal.excalidraw --out my-task

npx practonaut play my-task/world

The difference becomes the goal (steps, labels, shapes, colours, frames, groups, arrows). Practonaut only writes the task if the goal passes its own check, the start fails, a generated reference solution passes and generated near misses fail. The same works in the Player under Create, where you can also record a demonstration in the real editor and have the rule confirmed by a second recording. See the authoring guide.

Benchmark your agent

npx practonaut bench --agent oracle --tasks all --seeds 2            # reference solutions: the score's ceiling

npx practonaut bench --agent random --tasks all --max-actions 30 # random actions: the floor npx practonaut bench --agent claude-code --tasks excalidraw/insert-between --repeats 3 npx practonaut bench --agent "my-agent --mcp {mcp_config} --prompt {prompt}" # any agent npx practonaut bench --agent claude-code --world my-task/world --oracle my-task/oracle.json # your own task npx practonaut report ~/.practonaut/bench/ /report.json --html report.html npx practonaut report ~/.practonaut/bench/ /report.json --verify # re-judge from the run files npx practonaut compare before/report.json after/report.json --fail-on-regression # did a change break an instance? npx practonaut report ~/.practonaut/bench/ /report.json --junit runs.xml --card card --badge badge.svg

Each run gets its own session and data folder; the agent reaches the session only through the Practonaut MCP tools and gets the same public task a person gets. The claude-code preset does not pass provider keys, tokens or billing switches from your environment to Claude Code (a set ANTHROPIC_API_KEY would silently switch it to API billing), so Claude Code uses the login it has — your subscription, if you logged in with one; --keep-env ANTHROPIC_API_KEY passes a key on purpose. Scores come from the host's history and a separate checker, never from what the agent says: pass@1, pass^k, actions, efficiency against the reference solution. The JSON report has a schema. report checks that every pass and every number follows from the verdicts in the file — consistency, not proof: a file edited consistently still passes, and report says "consistency-checked, not re-verified". report --verify re-runs the checker on the session and the real-app export kept in the bench folder, whose SHA-256 the report records, against the world and reference solutions stored in that folder (so a moved folder still verifies), re-derives each run's actions and reference length, and refuses any difference by name; host calls and times are marked as reported by the runner. Nothing is signed: whoever can write the bench folder can replace those files too.

A report says why a run failed, not only that it did: rates as fractions with a 95 % interval, failed runs first, each unmet goal condition with the action after which it was last lost (judged by the checker after every action), error classes with their explanation, the trajectory, the final board — the real Excalidraw's own export, or a schematic labelled as such — beside an example target, the call log and the tokens, turns and cost the client reports about itself. compare puts two reports side by side instance by instance: it refuses reports that are not comparable, names instance regressions (three repeats a side, all passed before, all failed after) and says "not distinguishable at this n" when an exact McNemar test cannot tell the two apart. Both reports must come from the same kind of agent: oracle and random mark the ends of the score and compare only with a built-in, a command template only with a command template; as the baseline for your agent, bench a configuration of it that you trust.

In CI, examples/ci/practonaut-bench.yml benches your agent with your own credentials and fails the job on an instance regression against a committed baseline; --junit gives CI test views one test case per run. --card and --badge draw the header line as a 1280 × 640 card (SVG and PNG, light and dark) and a badge, with n, the interval, the verification status and the isolation level on them.

The example report is a real run of Claude Code on three tasks, two repeats each (JSON, HTML) — one seed, six runs: it shows the format, it does not rank a model. Details: benchmarks.

Task library

Library, not a challenge tier. These tasks are an example set and a shared smoke test, not a measure of ability: current Claude models solve them, and five harder candidate families found no headroom on Practonaut's own actions in 0.2.0 (gate reports). A family becomes a challenge tier only after it passes a preregistered ladder gate.

16 tasks for Excalidraw 0.18.1, each in English and German with three label variants, a reference solution and near misses. List them with npx practonaut tasks list; open one with npx practonaut play ; address them as excalidraw/ in mcp and bench.

How scoring works

1. The host keeps the history. Every action goes through one session host. It records each action with a state hash and who made it; agents and people cannot write the state any other way.
2. The real engine confirms each action. With Chromium installed, each action is applied to a pinned Excalidraw running headless and read back; an action the real app disagrees with is rejected.
3. A separate checker decides. After the attempt, another process replays the history with its own interpreter, reads the real Excalidraw export with its own parser, and evaluates the goal. It first makes sure an empty, an unchanged and a wrong submission all fail.

An agent gets the same task and the same public goal conditions as a person in the Player — one sentence per condition plus its formal form, which is the goal predicate written out — and the same Check button. It acts through Practonaut's semantic actions over MCP, not through the GUI. The reference solution, the near misses and held-out checker data never reach it over MCP.

Limits

• One application. Excalidraw 0.18.1 is the only app so far. The adapter contract is meant for more (adapter guide); none exists yet.
• Semantic actions, not pixels. Agents act through Practonaut's actions (add a step, connect, rename, …), not by clicking a screen. This measures whether an agent can reach a goal in the app's model, not whether it can operate the GUI.
• Not sealed. Agents run as your OS user. An agent with its own file or shell tools could read the task library on disk, reference solutions included. The claude-code preset allows only Practonaut's tools; for other agents that is up to you.
• Small evidence so far. The example report has two runs per task. Whether practising here makes people better at the real tool has not been studied.
• Without Chromium Practonaut runs model-only: tasks work, but nothing is checked against the real app, and it says so.
• Screen-reader and 200 % zoom checks by a person: one run so far (VoiceOver on macOS with Chrome, 2026-10-17), no problems found; other screen readers and regular screen-reader users are not tested yet (manual checks).

Requirements

• Node.js ≥ 22.18 and npm.
• Optional: npx practonaut setup downloads Playwright's Chromium headless shell once for the real engine. Nothing is downloaded on install; the package has no install script.
• Local data lives in ~/.practonaut (or PRACTONAUT_HOME). No telemetry, no account. Practonaut's own process makes no network calls while practi