io.github.RyanAlberts/agent-harnesses

Side-by-side comparison of 2-4 harnesses by github_id — for "should I use X or Y?". Answers the head-to-head question with the list's curation intelligence instead of a raw spec dump: each project's record aligned on the list's axes, an edge summary naming which project leads on stars / adoption simplicity / autonomy / failure recovery, a warning when a requested id is in the graveyard (archived or integrity-flagged), and the decision guide covering this matchup when one exists. github_ids: 2-4 ids, e.g. ["openclaw/openclaw", "NousResearch/hermes-agent"] — use search_harnesses to find an id. Returns JSON: {projects, edges, warnings, see_also, source, stars_captured}. Edge values are lists of names (more than one = a tie); an axis is omitted when fewer than two of the compared projects are rated on it.

Pick the top harnesses for a use case or task and compare them side by side. One call for "compare the best options for X": candidates are ranked the same way pick_harness ranks them, the top 2-4 are compared head-to-head (per-axis edge lists incl. the researched deep-dive axes — sandboxing, context memory, lifecycle hooks, prompt optimization — plus each project's build-vs-buy tier), with each pick's ranking reason and the decision guide covering the matchup when one exists. use_case: the task, e.g. "sandboxed code execution for generated code". limit: how many top candidates to compare (2-4, default 3). Returns JSON: {use_case, projects, why_picked, edges, warnings, see_also, source}.

Full markdown of one decision guide by slug (see list_comparisons). Guides cover architecture trade-offs, field reports, and the post-June-2026 billing reality — use them when a user is choosing between specific harnesses, not just browsing.

Full record for one project by github_id (e.g. "anomalyco/opencode").

Full markdown of one playbook by slug (see list_playbooks).

One skill by name (see list_skills): every file's content, each with the path to write it to. target picks the skills folder: "claude-code" writes to .claude/skills/<name>/, "agents" writes to .agents/skills/<name>/ (read by Codex, Gemini CLI, Cursor, and OpenCode). Write the files, then tell the user the skill is installed and how to ask for it.

One template by slug (see list_templates): its README plus the full content of every file, each with the path to write it to. Use this when a user asks you to set a harness up: write the files, then adapt them.

The list's categories and curated use-case intents, with project counts.

The list's head-to-head decision guides (e.g. "OpenClaw vs Hermes", "How to pick a harness") — slug, title, and summary for each. Fetch the full text of one with get_comparison(slug).

Step-by-step guides for one setup task each (build your own harness, one AGENTS.md for every harness, ...): slug, title, and summary. Fetch the full text with get_playbook(slug).

Agent skills that check a user's own harness setup (what the agent can reach, whether its 'tests pass' claims hold, where tokens go, which guardrails leak, ...): name, title, summary, description (when to use it), install command, and file list for each. Fetch one with get_skill(name).

Copy-paste setup files for agent harnesses (an AGENTS.md that works in every harness, safe Claude Code settings, a minimal harness in Python, ...): slug, title, summary, and file list for each. Fetch the files of one with get_template(slug).

Recommend agent harnesses for a use case, ranked from a hand-curated list of 100+. use_case: what you want to do, e.g. "terminal coding agent", "drop-in memory layer", "evaluate agents on coding benchmarks". max_complexity: cap on adoption surface — one of "super simple", "mostly simple", "slightly complex", "complex" (default: no cap). min_autonomy: require at least this designed autonomy regime — one of "step-gated", "checkpoint-gated", "bounded", "headless" (e.g. "bounded" means "must be able to run a whole task unattended"; excludes n/a entries). min_recovery: require at least this failure-recovery tier — one of "none", "retry", "resumable", "durable" (excludes n/a entries). open_source_only: drop projects with restricted or unknown licenses. Returns JSON: ranked picks with a one-line reason each.

Pick agent infrastructure at any level of the stack: curated list first, live web discovery second, so the answer is never limited to the list. Where pick_harness ranks only the curated list, this tool adds a live discovery pass for capabilities newer than the list: a GitHub repository search (recently pushed, star-ranked, already-listed and graveyard repos removed) and a Hacker News search (recent stories, 20+ points). Live results are labeled "unvetted": they have NOT passed the list's curation bar, so treat them as leads to evaluate, not recommendations. need: plain language, e.g. "somewhere safe to run agent-written code", "trace and score my agent in production", "hosted browsers for a scraping agent", "stop tool schemas from eating my context window". level: optional stack level; inferred from the need when omitted. One of: model-access, harness, orchestration, sandboxing, browser, memory, context, tools, evals, observability, security, skills. include_live_search: set False for a fully offline, curated-only answer. Live search sends only the need text to api.github.com and hn.algolia.com; an optional GH_TOKEN/GITHUB_TOKEN env var raises the GitHub rate limit. Returns JSON: {need, level, curated_picks, avoid, decision_guides, live_discovery, source, stars_captured}. Degrades gracefully: when live search is off or unreachable, curated results still return and live_discovery.status says why.

Opinionated single recommendation for a need — a decision, not a list. Where pick_harness returns a ranked shortlist, recommend commits: one top pick with the reason, up to two alternatives, any listed harnesses to AVOID for this need (archived, or flagged for star manipulation — with why), and the most relevant decision guide to read next. Use this when an agent or user asks "what should I actually use for X?". need: plain-language description of what you're building, e.g. "an always-on personal assistant in my chat apps" or "evaluate a coding agent on benchmarks". language: optional — restrict to a language/runtime tag (python, javascript, typescript, rust). must_run_unattended: require a harness designed to run a whole task with no human in the loop (autonomy bounded or headless). open_source_only: drop restricted or unknown-license projects. Returns JSON: {recommendation, alternatives, avoid, see_also, source}.

Keyword search across all 100+ projects (name, description, tags, category). Returns JSON: matching projects sorted by relevance then stars.

What changed in the list, week by week: harnesses added or removed, the biggest star gains, the fastest growers, projects archived upstream, and new guides, skills, and templates. Pass since=YYYY-MM-DD for every week after that date, or omit it for the latest `limit` weeks. Use it to answer "what's new in agent harnesses" or to check whether an earlier recommendation is out of date.