howp

Turn what you follow into measurable questions, bind them to Polymarket and Manifold markets, and write a forecast from their probabilities and the agent's judgement.

Install

Claude Code

/plugin marketplace add Akurganow/ai-plugins
/plugin install howp@ai-plugins

Source: Claude Code documentation, https://code.claude.com/docs/en/discover-plugins.

Codex

codex plugin marketplace add Akurganow/ai-plugins --ref main
codex plugin add howp@ai-plugins

Source: Codex documentation, https://developers.openai.com/plugins/build/plugins, for marketplace add; Codex source, plugin_cmd.rs, for plugin add.

Oh-My-Pi

omp plugin marketplace add Akurganow/ai-plugins
omp plugin install howp@ai-plugins

Source: Oh-My-Pi documentation, docs/marketplace.md.

Hermes

hermes plugins install Akurganow/ai-plugins/plugins/howp --no-enable
hermes plugins list
hermes plugins enable howp

Keep the plugins/howp suffix: without it Hermes copies the whole repository, and its two-level scan finds no package.

Source: Hermes documentation, developer-guide/plugins/index.md, for the commands; Hermes source, plugins_cmd.py (_resolve_git_url), for the subdirectory form and what happens without it.

Usage

Give the agent a first prompt:

Set up a forecast workspace in ~/howp. Ask me what I follow, then bind my questions to markets and write the first forecast.

With the interests skill, the agent interviews you one question at a time. It writes interests.yaml and questions/*.yaml only after you approve the list. With the forecast skill, it reads binaries.json, probes the hosts, downloads the hp archive for your machine and checks its sha256. It searches Polymarket and Manifold for a market per question and records each verdict with its reasoning. It fetches the market bodies, and hp records the probabilities. hp render writes the forecast page to data/dashboard.md in the workspace.

Later prompts refresh the forecast, explain a sharp move or write the weekly digest. Another forecasts one event from recorded results, the gated market and evidence, and scores that forecast after the outcome.

What's inside

Path	What it holds
`skills/interests/SKILL.md`	The `interests` skill: interviews you and writes the questions a forecast is built on.
`skills/interests/references/interview.md`	The interview, the formats of the two files you own, and what makes a good question.
`skills/forecast/SKILL.md`	The `forecast` skill: the platform rules, the workspace, the routine cycle and the four procedures.
`skills/forecast/references/install.md`	How the `forecast` skill gets `hp`: the platform gate, the preflight, the download, the sha256 check and the unpack.
`skills/forecast/references/procedures.md`	Binding a question to a market, explaining a sharp move, writing the weekly digest, and forecasting one event from recorded results, the gated market and evidence, then scoring it.
`skills/forecast/references/commands.md`	`hp --help` for every subcommand, written by the release job.
`binaries.json`	The released `hp` archives: version, targets, URLs and sha256 digests, written by the release job.
`plugin.json`	The manifest, with the host list under `extensions`.

Requirements and network

hp runs only on the targets binaries.json lists. On any other machine the forecast skill stops and names the listed targets. It never builds hp from source, because the source repository is private.

The forecast skill runs uname, curl, tar, mktemp, and shasum or sha256sum. Its procedure for a forecast from evidence and markets also runs jq. It needs no API key and no account.

Hosts

plugin.json lists every host the forecast skill's steps name, under extensions["io.github.akurganow.ai-plugins"].network.hosts:

github.com
release-assets.githubusercontent.com
raw.githubusercontent.com
gamma-api.polymarket.com
clob.polymarket.com
api.manifold.markets

The release archive comes from github.com and its redirect host, release-assets.githubusercontent.com, until the forecast skill caches a checked copy, as install.md Step 3 records. raw.githubusercontent.com serves binaries.json and plugin.json only when a client loads the forecast skill without its package. The Polymarket and Manifold hosts serve the market bodies. hp opens no socket: the agent fetches every URL hp names with its own tools, under your client's permission flow. The forecast skill probes each host a run needs and reports a blocked one instead of routing around it.

The list is a floor, not a fence. The agent searches for candidate markets, for the story behind a move, and for the results and evidence a forecast is built from, on its own. The forecast skill names no host for those searches, and the hosts they reach are yours to allow. The standard gives the list no meaning: the Agent Plugins manifest schema assigns "no semantics to namespace object contents" of extensions (schema, specification).

What it reads and writes

The forecast skill caches the archive at ${HOWP_CACHE:-$HOME/.cache/howp}/<version>. The workspace holds interests.yaml, questions/, matches/ and data/, and forecasts/ once you forecast from evidence and markets. HP_DATA_DIR moves data/ elsewhere.

When something fails

The archive's sha256 must equal the digest in binaries.json. On a mismatch the forecast skill deletes the archive, runs nothing and tells you.
A blocked host stops the step that needs it. The forecast skill names the host and where to allow it.
A market that does not answer costs that market, not the run.

Boundaries

It scores only a forecast recorded with hp forecast pool, against an outcome file written from the results. Whether any other forecast came true is yours to judge.
The package does not search for markets. The agent searches, and hp validates what the agent hands it.
It places no bets and needs no account on either venue.
The package calls no model. Every judgement comes from the agent you run it in, and every evidence effect from an estimator you run yourself.

License

MIT. See LICENSE.

Help

See SUPPORT.md.