Hark is built in the open. Every meaningful change lands here, newest first.
RSSNote: Marketing plan, Nov 2 – Dec 1, 2026 (summary and link) 5 proposed decisions: audience ranking; positioning and headlines; founder-led X voice; marketing loop rules; launch week Show HN Nov 17 / Product Hunt Nov 18 7 growth-pipeline items: title tag, OG image and alt, social proof, hark-mod stars, retire hark@hark in X templates, directory listings (planned); /for/agent-teams page (idea) Doc link: marketing plan Page Open question: a free entry point vs the accepted 14-day trial decision
Recipients can rate a handoff and flag missing blocks. Owners add gaps to the project, repeated gaps warn before sending, new projects start with type-appropriate trust, and Connections shows which calls counted. Public pages now explain evidence, continuity, quiet review, approvals, repo rules, plan reconciliation, access, five-block records and agent teams.
Shipped from the Instinct agent test: plan reconciliation against the repo on connect, every push/PR/release, and nightly (a single "Looks done — mark complete?" card with evidence; V-085 produced 11 items; human-only items never matched to commits); decision-vs-task contradiction flags on task cards, in the brief, and in Needs you with Keep task / Replace decision / Edit task, rechecked on every load; stale-item prune as a review card with per-item Mark done / Still relevant / Remove; summary date-stamped ("auto · as of Oct 20" or "written by you · Oct 6") with Refresh from latest. Reads are side-effect free: GET brief returns session: none for read-only and Propose keys; sessions open on explicit start or first write; write-less sessions auto-close after 30 minutes idle (four stale ones closed, including "Instinct test 1" on V-085). New key scope "Propose" (default: can read and suggest, cannot accept); relay toggle only on Propose/Read+write. One-link agent connect at /connect/agent?code= with scope, projects, expiry, relay, Approve. Two read-only repo endpoints: repo activity and changes_since_accepted (REST + MCP). Recommended agent User-Agent documented. Not done: a Cloudflare rule change (not controllable from Lovable); live tests show python-requests and curl are not blocked. 82 tests pass.
Agent team mode live: per-key agent labels and trust, Tasks on all project types with seven task tools (52 tools total), Agents feed, task wizard with briefing, quiet Needs-you with Auto-accepted today + undo, Memory view with search, "I run AI agents" onboarding door, Agent teams docs page Import output-token cap added (4,000) Fixed in build: the "·" in agent labels broke an HTTP header and failed agent writes Daily digest email NOT built (Lovable email service disallows digest-type mail)
Shipped Agent team mode (the "fabric" for people who run several AI agents): per-key agent identity (label, avatar) and per-key trust (decisions, non-goals, phase never auto-accept); Tasks on every project type with seven tools (list_tasks, get_task, claim_task atomic, release_task, add_task_note, add_attempt, complete_task — 52 tools total); import output-token cap (4,000); Agent team mode navigation Projects · Agents · Needs you; Agents feed grouped by day with filters and "who's doing what" plus a not-heard-from flag; four-field task wizard producing a copyable agent briefing; quiet Needs-you (decisions and task confirmations only, "Confirm all completed tasks", "Auto-accepted today" with undo); Memory view (what's true / what happened / what's being worked on / who's doing what) with plain-language search; "I run AI agents" onboarding door and the Agent teams connect-docs page. Verified live with two labeled keys. Not built: daily digest email (Lovable's email service disallows digest-type mail; summaries stay in-app — plan: deliver the daily summary via the Telegram bot and Slack instead). Limits: task context does not yet select decisions by keyword; the tool half of attribution is self-reported. Bug found and fixed in build: the "·" separator in agent labels broke an HTTP header and failed agent writes. Audit before build confirmed the only model calls in the product are the three Bring-in imports (GPT-6 Astra; October gateway spend 0.63 credits); handoffs, drafts, briefs, digests and reviews are deterministic — a fabric user's marginal cost is storage and bandwidth.
PR https://github.com/xai-org/plugin-marketplace/pull/1252 opened from Hark-Studio:add-hark (replaces #1230) PR https://github.com/xai-org/plugin-marketplace/pull/1230 closed with comment "Superseded by #1252, opened from the Hark-Studio org fork."
github.com/Hark-Studio/hark-plugin: main pushed, tag v0.1.0 = d88a073f0e0ec724e24cc419b392b879f4e677d4 commit 3a04be8 Pin the xAI marketplace entry to v0.1.0 (listings/xai-marketplace-entry.json, README listings table) PR https://github.com/xai-org/plugin-marketplace/pull/1230 (branch add-hark on fork mschreiber89/plugin-marketplace): entry "hark", category development, remote source pinned to d88a073, domains harkstudio.io; plugin-index.json regenerated ship logged: hark-plugin 0.1.0 published and submitted to the Grok Build marketplace venture link added: hark-plugin repo
Public repo pushed (main) and tagged v0.1.0 at d88a073f0e0ec724e24cc419b392b879f4e677d4. PR opened against xai-org/plugin-marketplace adding the `hark` entry (category development, remote source pinned to that SHA, domains harkstudio.io): https://github.com/xai-org/plugin-marketplace/pull/1230. Not yet merged; Hark must not be described as available on Grok until it is. The OpenAI directory submission and the Grok Bot path are still open.
Shipped a REST façade at https://harkstudio.io/api/v1/ (13 endpoints) that mirrors the MCP tools 1:1 — same keys, limits, sessions (X-Hark-Conversation = _meta.conversation), proposal rules, and attribution ("by Muse", "API (curl)"). OpenAPI 3.1 at /api/v1/openapi.json generated from the tool definitions (can't drift; sync test fails if a tool lacks a route or an explicit exclusion), readable reference at /api. New tool create_handoff (MCP + REST; 45 tools total). New record kind "note" (agent-authored, always a proposal, never shown as Observed); "observation" removed from agent-selectable kinds. Read-only keys can open read-only sessions (by design) and are refused on every write. Connect-docs pages added for Meta Muse (custom connector via OpenAPI + key), Muse Code (MCP), and Personal agents (Instinct, Poke, OpenClaw). Unlocks Meta Muse custom connectors and any HTTP-only agent today; Meta Connector Platform submission is next.
new repo ~/code/hark-plugin (local git, main, commit d88a073), not on GitHub yet plugin.json: Agent Plugins 1.0.0 manifest, name "hark" 0.1.0, OpenAI interface block (displayName Hark, category Productivity, support/privacy/terms URLs) mcp.json: streamable-http server at https://harkstudio.io/mcp; .mcp.json: same server as type http .grok-plugin/plugin.json, .claude-plugin/plugin.json, .codex-plugin/plugin.json: duplicate manifests pointing at ./skills/ and ./.mcp.json .agents/plugins/marketplace.json: one-entry Codex marketplace for `codex plugin marketplace add Hark-Studio/hark-plugin` skills/hark-session/SKILL.md: find .hark/project, get_agent_brief first, work inside decisions, log as you go, end_session or close_session skills/hark-handoff/SKILL.md: the four end_session fields, style, draft handoff + confirm_handoff skills/hark-bot-notes/SKILL.md: always-on rules: one session per run, allowed reads/notes, forbidden writes, read-only token, plan cap README.md: install per platform, setup, skills, every outbound request, layout, listings status, Hark-side prerequisites listings/: xai-marketplace-entry.json, xai-pr-body.md (security block filled), openai-directory.md (prereqs, ZIP, 5+3 test cases), grok-bot.md (path unverified) scripts/check.py + .github/workflows/check.yml: manifest/skill/asset/secret checks assets/: icon.png and logo.png copied from hark-mod (1024px), hark-icon-light.svg
Shipped the Hark mod for Claude Code (github.com/Hark-Studio/hark-mod, v0.2.1): session-start brief injection, post-compaction recap, decision guard on edits (scoped by path/area), stop = handoff via end_session with session_stopped fallback, status line and /hark commands. 60 tests; strict validation clean. Live-tested headless and interactive against V-059: brief used unprompted, guard dialog fired on a billing edit, handoffs recorded. Key now supplied only via the plugin's sensitive setting (no env/keychain reads); no child processes; literal endpoint; README documents every request, command, and file. Directory validation went 16 → 9 holds (remaining holds are inherent: reads conversation, hooks events, icon). Submitted to the Claude directory as a Plugin bundle. Server-side fixes shipped: access-key auth race (one-time sign-in link per request → shared refresh-renewed login; 401s never 500s), one session per conversation via _meta.conversation, header-based attribution "via Claude Code (mod)", end_session fast acknowledgement. Docs page live at /how-it-works/connect/claude-code-mod (early access).
The Hark mod for Claude Code adds the brief at session start, restores the plan after compaction, asks before edits that touch an agreed decision, drafts the handoff on stop, and shows status plus /hark commands. Early access: install today from github.com/Hark-Studio/hark-mod with /plugin marketplace add Hark-Studio/hark-mod, then /plugin install hark@hark.
Accepting or rejecting a suggestion updates the project, workshop, card chips, and phone tab within seconds, with no reload.
Current state comes first with a "Continue" button; one Needs you queue with Accept, Edit & accept, Reject; three labels: Agreed, Observed, Proposed. Phone-first Connect and a one-screen Hand off with a live gap check.
Brief reads no longer count toward your plan; briefs show their token cost and skip what is already in your repo instructions. Record what you tried and what didn't work, and resume a project in Cursor, Claude Code, Codex, or VS Code.
Reads no longer count toward your agent-call cap — only writes and AI-processing calls do, and the usage graph shows the two apart. Every brief now ends with its approximate token count, what was left out, and how many items your repo's CLAUDE.md/AGENTS.md already covers.
Record what was tried and didn't work; failed attempts show as a "Don't retry" list in the brief and handoff, and decisions carry the alternatives you rejected. Answers to recipient questions can become decisions, and "Resume in…" gives one copyable block for Cursor, Claude Code, Codex or VS Code. Hooks now fail open within 2 seconds, with a new hark doctor check.
Standard, Relaxed, or a 24-hour Autopilot per project and tool; "Accept all from this session" in one click. Decisions always stay manual, and anything that conflicts with an agreed decision goes to review.
Every sync snapshots the previous file first, compares before overwriting, and lets you restore any earlier version.
Each project picks how much agents approve on their own (Standard, Relaxed, or a 24-hour Autopilot per agent tool), marked Auto-approved everywhere. Decisions are never auto-approved, and anything that clashes with one waits for you. "Accept all from this session" clears the rest in one click.
Before syncing, Hark shows both copies side by side and keeps a version history you can restore from. Repeated Current state items now merge into one.
When a Claude Code or Cursor session stops, Hark drafts the handoff from commits, PRs and releases. Recipients can mint a read-only tool key, and owners get one daily digest when suggestions pile up.
One vocabulary across the site, spaced-out desktop menu, project page reordered: documents, Current state, repos, links, activity. GitHub sign-in removed; sign in with Google or email.
GitHub sign-in is gone because the sign-in service doesn't support it. Linking repos, repo activity and the .hark/ export work exactly as before.
Paste a conversation, or tap the sample, and see a proposed Current state before signing up; it carries into your account. Anonymous AI use sits under a daily spend cap.
Add to Cursor and VS Code, one-step repo linking with export on by default, and a simple view showing just Current state, Decisions and Hand off. Paste to Current state now takes about 19 seconds.
Someone who signs in from a handoff gets access to that one project and can add proposals you review. No workshop membership, no seat used, revocable anytime.
Read-only personal access tokens can now fetch the brief and list projects. Their sessions can never write, and any write is refused naming the token's scope.
One onboarding path, no "Agreed" label without an approval, and "Accept all quoted" takes only items tied to an exact passage or commit. Checklist labels now say exactly what was verified.
Every record now carries its source, acceptance, applicability and scope; recover a project from a conversation, repo or existing project and review it as a diff. Handoffs are recipient-first, with a gap check and a questions loop.
Turn export on for a linked repository and Hark writes the brief into the code itself, so it survives even if Hark is unreachable.
Four changes that stop context leaking between projects and stop agent memory entering your project page unreviewed.
See agent calls per day against your plan allowance, and hear about it before you hit the ceiling.
Link a repository to a project and let scripts and headless agents in without a browser sign-in.
The living half of a project: where the build actually stands, and one call that hands it to any agent.
Publish a project, gather followers, and show the build off wherever your code lives.
Hark is now readable by search engines, answer engines and agent directories: server-rendered content pages, structured data on every public route, an llms.txt surface, Markdown mirrors of the docs and published MCP metadata.
Setting up the session flow in a repo is now a single command, and Hark measures whether a different agent actually picked the work back up.
Ideas and progress can now reach Hark, and leave Hark, without opening the app.
A two-minute weekly pass over every active project, and a quick way to tell people what changed.
Every new workshop now starts with a free 14-day trial and no payment details. Solo builders start on Builder, teams on Studio.
Two ways into Hark that both end in a verified handoff, instead of an empty workshop.