Worca

Worca turns a task into reviewed, merge-ready work. It is a deterministic multi-agent pipeline: you point it at a project, describe the work, and a state machine runs the agents of a workflow in sequence — by default Clarify → Plan → Refine → Implement → Review — looping until the work clears quality gates, pausing to ask you the questions that matter, and keeping every run's changes apart in its own git worktree and branch — or, for a plain folder, on a copy you apply back with Apply to folder; with No project, in a folder of its own. Pick Auto and Worca chooses the workflow for the task, draw your own in Workflows (or ask its chat to build it), or start the shipped Presentation workflow to turn a brief into a slide deck, Architecture (C4) to keep a diagram of your code that draws itself from the code, or Promo video to film your product's real UI. Agents run on Claude Code by default, or (beta) on Codex, the GitHub Copilot CLI or Cursor.

It ships as a web UI, a CLI, and an installable /worca skill for Claude Code — all running the same engine — and chat plugins (Telegram, Slack, Discord, Teams) let you drive runs from a chat. See the architecture in one picture.

How a run works

Pick Auto — first in the New pipeline Workflow picker and selected by default (--workflow auto on the CLI) — and Worca picks the workflow for you. One model call reads the task text, the attached files and a small offline fingerprint of the repository (and may take a brief read-only look at the checkout), then classifies the task: a free-form prompt, a partial plan or a complete plan, how large the change is, and signals such as a big web UI feature. It knows every agent only by its metadata and the front matter of its agent file (never the agent's full instructions) and assembles a matching workflow from those agents. A saved workflow with exactly that shape is reused; otherwise the proposal is saved as a new workflow once accepted. With Human in the loop on (the default), you see the proposal first: Accept & run, Revise it in plain text (Advanced and up), or Cancel run. Turn the switch off (--no-human) and the proposal is accepted as is, with no clarify step and no agent questions; a loop that runs out of cycles, a recovery prompt, a cost cap or an error still pause the run. Pick any saved workflow instead to skip all of this. Auto works for workspace runs too.

The built-in Default workflow hands your task to five agents:

1. Clarify — instead of assuming, a Clarify agent turns hidden decisions into up to 8 multiple-choice questions (2–4 options plus free text, optionally with a recommended pick), and asks none when the task is clear. The planner records your answers in the plan's ## Clarifications (Q&A) section, so every later step sees them.
2. Plan — the planner explores the codebase and writes an implementation plan with concrete code snippets.
3. Refine Plan — the refiner reviews and rewrites the plan (-v2, -v3, …) until no critical/major issues remain.
4. Implementation — the implementer follows the approved plan with no deviation (recording any it is forced into), using TDD (red-green-refactor).
5. Review Implementation — the code reviewer reviews the git diff against the plan and hands blocking findings back to the implementer, looping Implement → Review until clean.

Each loop has a cycle cap (default 3, adjustable per loop). A loop that hits it with blocking findings left pauses the run and lists the open issues: I approve another cycle, or Don't approve another cycle and continue with what you have. Only critical/major findings block; minor findings and suggestions never do. A finished run's work is committed to its own worca-cc/… branch, ready for Create PR.

Every finished run keeps its full record, one tab each in Runs: Overview (the review verdict and its findings, time, cost, files changed), Diff (every file with its +/− and patch), Artifacts (the plans, reviews and other files the steps wrote), Actions, Workflow (the graph as it ran, with per-step time and cost at Expert), Q&A (the clarify answers and any mid-run questions), and at Expert Logs and Agents. Diff, Artifacts and Actions start at Advanced (see UI levels). A run in a Folder project works on a copy: in place of Diff and Actions it has Changes (every file it added, updated or deleted, at every level), and you land it with Apply to folder (files changed in the folder since the run started are skipped and listed in the dialog) or Discard. A run with No project works in its own folder: it opens on Artifacts at every level and has no Diff, Changes or Actions.

Features

Pipeline

• Deterministic engine — a scheduler walks the workflow graph and decides what runs next; agents do the creative work, never the control flow. Every step, verdict, cost and artifact is recorded (architecture).
• Pause & resume, even across restarts — Pause parks a run, and so does a cost cap or a failing step; only Stop ends one. Resume re-attaches the paused Claude sessions (claude --resume) in the same worktree, uncommitted agent work included, or Resume at… schedules it — at a time, or after another run ends, so paused runs can resume one after another. Paused runs survive a server restart, and a run cut off mid-step comes back as Interrupted, resumable from its last clean point. A step that loses the network waits for it and retries by itself; a run parked by an outage resumes when the network is back.
• Isolated worktrees — each run works in its own git worktree on its own worca-cc/… branch, so agents never edit your checkout and parallel runs don't collide. When a run ends, its work is committed to that branch and the worktree is removed; the branch stays. A worktree isolates changes, not the host: natively, agents run as your user and can reach whatever you can (see Guardrails).
• Run on the source branch — switch it on under the feature branch (or --on-source-branch) and the run still works on a hidden branch, but when it finishes done its source branch is fast-forwarded onto that work and the hidden branch is deleted; the run's page then offers Push instead of a pull request. Only ever a fast-forward: a source that moved meanwhile, a dirty checkout of it, or a run that did not finish leaves the work on its own branch, with the commands to merge it by hand.
• Live cockpit — Runs in the sidebar lists every run, live and finished; its View menu shows All, Needs you, Running, Waiting, Finished, Ready to ship (finished, no PR yet) or Archived, and groups by project or date. A live run's page shows its state and current step, ticking time, cost and changed files, and a Live view of the running step inside its workflow graph; Diff shows the worktree's changes so far, and Logs streams every step with search (filters by source, level, step and cycle at Expert). The top bar's Activity pill counts what needs you and what is running; Settings › General › Alerts adds desktop notifications and a tab badge when a run waits for you. Archived runs (Advanced) lists deleted runs and restores them.
• One-click PRs — a finished run's page offers Create PR: Create PR lets you edit the title, write the description or Generate it (Settings › Runs › Pull requests can draft it as soon as the dialog opens), open it as a draft and watch it, then pushes the branch and opens the PR with the GitHub CLI (gh), or on Azure DevOps with a personal access token, and Bitbucket Cloud with an access or API token. A run started from a GitHub issue closes it on merge, and a workspace run opens one linked PR per changed repo. The button then becomes View PR. Publish branch (Advanced) pushes the branch without a PR.
• Watch PR and conflicts — with Watch PR on, Worca reads the open PR about once a minute; a failing check, a merge conflict or a requested change starts a fix run on the PR's branch and pushes the fix. After a run Worca also checks whether the branch still merges into its base: Update branch when it does, Resolve in a pipeline or Resolve in a terminal when it conflicts.
• Mock mode — the Mock mode switch on New pipeline (--mock on the CLI, WORCA_MOCK=1 for a whole server) runs the pipeline against a deterministic offline mock — no claude, no tokens — for demos, development and CI.

Agents

• 32 built-in agents, all data — each is a markdown prompt plus a metadata sidecar declaring its typed input/output ports, so a new agent drops in without engine changes; your own and plugin agents load the same way.
  • Coding — Clarify, Plan, Refine Plan, Plan Review, Decompose (splits a plan into vertical-slice tasks, one implementer each), Implementation, Review Implementation, Manual Tests Checklist, and Manual web UI testing (runs that checklist in a browser via Playwright).
  • Workspaces — Workspace Review (one review per changed project, one cross-project verdict) and the workspace scan's Workspace Survey, Workspace Usage and Workspace Synthesis, which only the scan can place.
  • Presentations — Deck Clarify, Narrative Spine, Visual System, Deck Builder, Deck Audit, Deck Review, Deck Outputs, Deck Export.
  • Architecture — C4: setup (the which-file, which-bands form) and C4: architect (reads the code, writes the LikeC4 model).
  • Video — Promo: setup, explore, script, voice, capture, mascot, scenes and review.
  • Memory defragment — restructures one memory scope.
• Agents in the Library — Workflows › Library › Agents lists every agent with its ports; built-ins are read-only, so Duplicate one to get an editable copy (docs/ui-levels.md).
• AI-assisted agent creation — + New agent in the Library (or New agent… in the canvas + menu): give a name, a purpose and a description and Worca drafts both the system prompt and the metadata, or paste your own prompt and let it infer just the wiring; edit the draft, Regenerate, Save agent.
• Ask forms — an agent that needs a human can declare the form its question is asked with: JSON in its sidecar, drawn by Worca from a fixed widget catalog (galleries, rankings, per-item review lists, previews of the files it just produced) and validated by the server on the way back. The CLI and chat get a text projection (chat answers with /answer field=value); a finished run's page replays the form read-only with its answer. Review Implementation, Narrative Spine, Visual System and Deck Outputs ask this way; plugins can ship forms too.
• Per-agent model & effort — pick a model and reasoning effort (medium, high, xhigh, max) per agent node: in the toolbar over a selected agent card in Workflows, or in New pipeline › Advanced › Agents for one project, kept for its later runs (Save as workflow defaults moves them into the workflow). A project pick beats a team-policy default, which beats the workflow's own; --model only fills nodes with none. Fan-out agents also take a sub-agent model.

Scripts

• Cards that run your own program — a script card keeps the whole outside of an agent card (typed ports, verdict routing, loops) and replaces the inside with your own Node.js, Shell or Python program (Python when the host has an interpreter), so a test gate or a transform costs no model call. See docs/scripts.md.
• Scripts in the Library — Workflows › Library › Scripts lists every built-in, own and plugin script with its runtime and ports. + New script picks a runtime, then reads the ports and params from the code as you type; Duplicate copies any script into your own folder; Delete refuses while a saved workflow still places it. Built-in and plugin scripts open read-only.
• Test bench — under the editor, runs one script by itself, even an unsaved draft, with hand-filled params and inputs through the very same runner a pipeline run uses. Save as case keeps the setup with an optional expectation (verdict, fired ports, summary text); Run all re-runs every case.
• Shipped scripts — Shell, JavaScript and Python cards, Git diff, and the steps behind the presentation, architecture and workspace-map workflows come built in. Plugins can ship scripts and test cases of their own.

Workflows

• A full-screen editor — Build › Workflows (every interface mode) is a canvas of frosted cards: drag a card from the Library, or add one with the black +, and wire its typed ports (md, json, void) with smooth wires coloured by type. Each card declares what it consumes and produces, a wire only connects compatible ports, and Save stays off until the graph is valid. A blocking output wired back to an earlier card becomes a loop (drawn amber, swooping under the cards with a ≤N pill): a reviewer keeps sending work back until it passes or the loop hits its cycle cap — 3 by default, set per loop wire in its toolbar. Flow cards — Task (the run's request), End (the result), AND, OR, Combine — express joins, choices and merges without any code. A selected card's toolbar sets its model, effort and await-all; More holds every other setting. The top bar keeps only Back, Workflows ▾ (New canvas, Open…, Import…, Export…), the name, the Library toggle and Save; Auto-layout and the zoom menu sit bottom right. See docs/ui-levels.md.
• The Library — the card beside the canvas switches between Agents, Scripts and Workflows. Every agent and script action lives there (view, edit, create, duplicate, delete, the script test bench); their editors open in a sheet inside the view. The old #composer, #agents, #scripts and #agent-create addresses land in the view.
• The composer chat — the pill beside + grows into a chat on Chat's engine that only works on the open canvas: it adds, removes and rewires agent and script cards and sets model, effort, max cycles and await-all (each change lands at once with an inline Undo), builds whole workflows (Apply to canvas opens one as a new unsaved workflow), and drafts new agents (metadata and system prompt) and scripts (code, interface and test cases, which it can test) that you Save from their card. It never runs anything: save the workflow, then start a run from New run. New chat starts over.
• From Chat too — the main chat drafts agents and scripts and proposes whole workflows (scripts, gates and new agents included) with the same cards, without a canvas: Save a draft or open it in its editor, and Save as workflow, Open in Workflows (a new unsaved workflow on the canvas) or Decline a proposed graph.
• Saved pipelines — the Library's Workflows tab lists them with domain chips; click a row to open it. The built-in Default and Memory defragment open read-only (Save makes a copy). Saved pipelines are offered in New pipeline's Workflow picker, and so in schedules.
• Share a pipeline — Export… on a Library row (or in Workflows ▾) offers three formats. A JSON file carries the graph only; the recipient picks it up with Workflows ▾ › Import… (a taken name gets a (2) suffix; nothing is ever overwritten, and any script commands are shown for approval first). In Expert, a Claude Code skill writes a SKILL.md plus the agents it dispatches under .claude/, so the pipeline runs inside Claude Code without Worca (not yet for pipelines with script cards). A Worca plugin folder bundles the pipeline with your own agents and scripts it uses and the skills they need — built-in agents are never copied — so the recipient runs worca plugin link <folder> and picks up edits with worca plugin reimport <name> — or pick Marketplace to open a pull request that adds it to a registered team marketplace (one branch worca/plugin/<name>, validated with the marketplace's own discovery; from a fork when you cannot push). The CLI does the same with worca workflow list, worca workflow export <id> --format json|claude|plugin and worca workflow import <file>.

Presentations

• From a brief to a slide deck — the built-in Presentation workflow turns a request into a deck: pick it in New pipeline's Workflow picker (Advanced; Simple lists only Auto and Default), have Chat propose one, or run worca --workflow wf_presentation --prompt "…". Deck Clarify asks what shapes the talk (live or read alone, audience, duration, the ask, sources, brand); Narrative Spine writes one hallway sentence, a 3–5 section spine and a takeaway title per slide; Visual System fixes grounds, type scale, compositions and the accent, with rendered preview sheets. You approve the spine and the system on ask forms or send changes back, and Deck Outputs asks what to deliver (PDF + standalone HTML, PDF only, standalone HTML only, or the deck files only) and whether to add narration audio.
• Built, measured, reviewed — Deck Builder writes a 1920×1080 HTML deck on the vendored OpenDeck kit, one section per slide. Deck Audit renders it headless, screenshots every slide and reports mechanical facts (boxes escaping the slide, clipped or overlapping text, contrast, text size, words per slide); once that is clean, Deck Review judges the rendered slides (focal points, takeaway titles, one hero moment). Findings go back to the builder, up to 3 cycles each.
• Files you can send — after a clean review, script cards print deck.pdf with headless Chrome (page count checked against the slide count), optionally record a per-slide ElevenLabs narration from the speaker notes (the API key is used for that run only, never written to disk), and inline everything into one self-contained deck.standalone.html; Deck Export then checks the result against what the request asked for. Their findings loop back to the builder too, up to 2 cycles. The files live in the run's folder, not your repository — open them on the run's Artifacts tab (Advanced). The audit and the PDF need Chrome, Chromium or Edge; the script cards need Python 3.8+. The narrated deck linked from docs/why-worca.md runs on the same kit.

Architecture (C4)

• A model of your code, kept by your code — the built-in Architecture (C4) workflow (worca --workflow wf_architecture_c4 --prompt "…", on a project or a workspace) finds the project's LikeC4 files, asks on one card which to update — or where to create one and how to group the deployment sheet into bands, each with prose criteria for what belongs in it — and which images to export. The architect reads the code and writes the model; a check card parses it with the viewer's own parser and sends blocking problems back, up to 2 cycles.
• Drawn the same everywhere — open the .c4 from a run's Artifacts tab and it draws in Worca's viewer, with a view picker, an info rail, hover and selection highlighting, a JPEG button and Full screen. The render card writes SVG, JPEG or PNG of every view with the same renderer (JPEG and PNG need Chrome, Chromium or Edge; in a container, the -full image). Worca's own model is docs/architecture/architecture.c4; the design is docs/architecture-c4.md.

Promo video

• A product film from your real UI — the built-in Promo video workflow (worca --workflow wf_promo_video --prompt "…") explores the project, writes a script you approve, captures the product's real screens in a sandboxed copy of the app, animates them with an optional mascot crew, reviews the film frame by frame and renders a 1080p60 MP4 with captions and YouTube metadata. A setup form asks for tone, length, pace, mascots and audio: narrated with ElevenLabs (pick a voice from playable previews; the key is never stored) or silent. Capture and render need Chrome or Edge and ffmpeg on the host; a run without them is refused up front.

Guardrails

• Named policy sets, picked per run — the Guardrails picker under New pipeline › Advanced offers the built-in Permissive / Normal / Strict tiers plus your own sets. A new install defaults to Normal; an install from before that keeps Permissive, which adds no rules. Normal protects credential files (.env*, keys, cert stores, container secrets, Worca's own DB and settings) and blocks publication commands (git push, npm/yarn/pnpm publish); Strict adds an environment scrub on agent spawn, denies network-egress tools (including git clone/fetch/pull, DNS and download CLIs), gh, docker push, cloud CLIs and WebFetch/WebSearch, and protects home-dir credential stores. A team policy can preset the picker and warn when a run's set ranks below its minimum tier.
• Managed in Settings › Guardrails — the tab (Advanced) lists every set with its details; Create guardrails (Expert) starts from any set or blank and edits five fields: honor project settings, environment scrub, environment allowlist, protected paths and deny rules. Built-ins are read-only; editing a set of your own applies to every later run that picks it.
• Enforced via Claude Code permissions — a set compiles to permissions.deny rules on every pipeline agent spawn, one set for every member of a workspace run; repo settings can't undo them. No preset denies MCP registry tools, but a deny rule you write on a server reaches every copy of it. See docs/guardrails.md for the full model and its honest limitations.
• Not a sandbox — native mode runs agents as your user. Normal and Strict restrict tools, not the filesystem: an agent with Bash can still reach your files, HOME and network through an interpreter. For untrusted work, run Worca in a container (docs/docker.md), where agents see only the projects you mount and can run as their own user.

MCP servers & credential broker

• Worca's own MCP registry — servers installed for Worca only, never written into your Claude Code config, that pipeline agents and Chat can call. The Connectors page (Advanced) has three views: Servers, the catalog (servers shipped by plugins, added by hand with Add MCP server as stdio, http or sse, or required by a team policy), and Sets, where each server gets its values and secrets, so one server can sit in two sets with different credentials, and Skills, the skill catalog (Skills): skills that plugins ship or that you import, added to sets like servers. The built-in marketplace ships MCP servers as plugins (see Plugins & chat).
• Sets per project, choices per run — Chat always uses the built-in General set; a project picks its sets on its Sets tab (Include General in runs is on by default), New pipeline › Advanced › Sets switches servers and skills off for one run, and the Ask composer's Sets · N button picks them per chat. Test starts a server exactly as a run would and lists its tools.
• Secrets out of config files — values live only in ~/.worca-cc/mcp/secrets.json (mode 0600) and reach a server as spawn environment; the run's MCP config holds ${MCPSECRET_…} references, and a stdio server starts with a short list of variables plus what its definition declares. The agents of a run that uses a set can still read its secrets; MCP servers lists what else can.
• Credential broker for container and hosted installs — model keys (Anthropic, OpenAI, OpenRouter, Copilot) live only in a separate broker container. Each claude process gets a short-lived wbt_ token that works only on the broker's private port and is revoked when the process exits; the broker adds the real key on the way to the provider its slot is pinned to. With the broker on, Worca refuses to start while any model credential is within agents' reach. In multi mode each teammate saves their own keys to the broker's encrypted vault (Settings › General › My credentials › Manage keys), and Statistics shows spend By person. worca broker serve|secrets|revoke --person <email>; see Credential broker.

Workspaces

• Multi-project workspaces — group 2–40 related projects — Git repositories or Folder projects — under Workspaces (Advanced level): Create workspace, pick the projects — or bring repos straight in with Add… or a pasted list (a repo, a folder of repos, a VS Code .code-workspace file; a ticked repo that is not a project yet becomes one) — Scan interconnections; the workspace is saved when that scan ends. Add projects / Remove on its Overview tab re-scan the new set on the workspace's saved scan model — view or change it with the header Scan model control (a re-scan with no usable pick asks you to choose one first).
• A map built from code, with evidence — the scan is itself a pipeline run that never changes a member. Code detectors find REST, gRPC and GraphQL APIs, message queues, build dependencies, shared databases and service calls and do the matching; agents fill what code cannot see and confirm uses. Every scanned edge carries file:line evidence and a confidence (exact, verified, heuristic, inferred), every member a coverage level.
• Review it on the Map tab — the workspace's Map tab draws the graph and lists every edge with its evidence; filter by project, kind, confidence or state, Confirm or Reject an edge, and Add edge for one the scan missed. Reviews survive Re-scan, and the description rendered from the map reaches every agent of a workspace run.
• Runs across all members — in New pipeline set Target to Workspace: one plan tags each task with the projects it touches, implementers edit only those, every Git member gets its own branch and worktree, every Folder member its own copy that you apply to its folder (or discard) on its own, and Workspace Review reviews each changed member and merges the findings into one verdict per review cycle. See docs/workspace-map.md.

Actions

• Try the result before you ship it — Check out on a finished run's Actions tab puts its branch in a worktree of its own and runs the project's setup once; then the project's own commands start with one click: a dev server with an automatic port, a ready check and an Open link, the tests, a workspace's whole stack, each with a live log. Built-in buttons open the checkout in your editor, terminal or file manager, Copy command gives the git lines for your own clone, and Discard saves uncommitted changes as a patch first.
• Configure once — commands live on the project's Actions tab, stacks on the workspace's; Settings › Runs › Actions sets the keep policy, port range, editor and terminal. All Actions tabs and that card are Advanced level; the sidebar's Running actions rows show at every level while a service runs, each with Stop (on the collapsed rail, one tile opens them).
• People start commands, agents don't — every start goes by action id from stored config; Chat can read and propose actions but never runs one; a hosted Worca runs nothing unless WORCA_ACTIONS_REMOTE=1. See docs/actions.md.
• A terminal for runs and projects — a pane on every page (Ctrl+`) with a shell in the run's or the project's folder, run by the Worca server; each command is recorded with output, exit code, duration and who ran it. A hosted Worca runs none unless WORCA_TERMINAL_REMOTE=1. See docs/terminal.md.

Plugins & chat

• Plugin system with marketplaces — a plugin is a git repo or folder with a worca-cc-plugin.json. It can contribute task sources (e.g. GitHub Issues, Azure Boards), agents (with their ask forms), scripts, skills, workflow templates, models, chat channels, and MCP servers for the registry. Install from the Marketplace page's Browse tab (Advanced) or worca plugin install; Add marketplace takes a GitHub URL, owner/repo or a local path. On a machine without git, marketplaces and installs work for GitHub repositories over HTTPS, a local plugin folder is linked, and other sources say Needs git. Installing from the UI is an explicit consent ceremony pinned to the commit you saw: what's installed, which ask forms an agent can show and which file types they may display, which secrets each part requests, where models route, which setup commands run. Update all reviews every update in one sheet — the plugin folder's commits, a diffstat and what changes; an update that asks for new access (a new secret, changed setup or model env, a new source, channel, agent or model, an MCP server or skill change) waits until you tick it as reviewed, and each applies exactly the commit you reviewed (worca plugin update --diff adds the full diff; worca plugin purge <name> deletes a removed plugin's leftover settings and secrets). Package your own with Export… on a saved workflow or Share as plugin… on the Models page; see the authoring guide.
• Built-in marketplace — Worca CC Official, Worca's GitHub repository on its dev branch, is registered on every install (nothing is installed until you pick it): the GitHub Issues and Azure Boards task sources, the four chat channels below, and MCP plugins for Atlassian, Cloudflare, Firebase, Notion and Railway. It works without git too (it lives on GitHub). Set WORCA_BUILTIN_MARKETPLACE to another repo URL or a local checkout to use that instead (followed at its HEAD; only committed files count).
• Tasks from your tracker — an installed task source adds its button to New pipeline › Task source (Advanced): search issues or work items, or paste a GitHub issue URL or owner/repo#123, and the ticket becomes the run's task. Each source can close its issue or work item when the run succeeds; Ask Worca can find a ticket and propose a run for it. See GitHub Issues, Azure Boards and Jira.
• Drive runs from chat — two-way Telegram, Slack, Discord and Microsoft Teams channels (Teams needs a public HTTPS endpoint through a tunnel; the others dial out). Pick the events that notify — questions and approvals, finishes, failures, pauses (cost limits included), Away hours — in Settings › Runs › Chat notifications (Advanced), and hit Test. Answer back with commands: /answer questions and forms, /approve or /retry at a gate, /direct a live run, /pause, /stop, /resume, /status, /cost, /use <name>, /mute 2h; /help lists them all. Commands are deny-by-default: only chats on the channel's allowlist may send them. Setup: Telegram, Slack, Discord, Teams.

Chat

• A chat that is the home page — Chat (the first sidebar row, and the worca logo), or the last row of the top bar's search (⌘K / Ctrl K), opens a chat that answers from Worca's own records: runs, their diffs, progress and clarify answers, workflows and agents, schedules, team metrics and policy, models, Actions, and tasks from installed trackers. It reads code in read-only, detached worktrees with an allowlisted git (log, diff, blame, grep…), takes attached text files, images and PDFs, and has no file-editing tools of its own (in agent mode, below, it changes files only through commands you ask for). Settings › Chat (Advanced) sets the turn limit, a per-turn cost cap, whether it may change and test-run your scripts, and opt-in web access with an allowed-domain list; chat spend counts toward the total budget, and a team policy can preset the limits or narrow web access, never widen it.
• One scope per chat — the composer's scope pill sets No project, a project or a workspace, at every interface level. Changes act only there. When a No-project message is clearly about one of your projects, Ask asks with a card — Switch to or Continue without project — and the same message continues after you choose; it never switches on its own. A project you create or add from a chat card, the chat home's Add project tile or the scope menu's Add project… becomes that chat's scope.
• Changes are cards you confirm — work becomes a run card (project, workflow, guardrails, branches and brief, all editable) that starts nothing until you click Start run, or Schedule… to run it later (docs/scheduled-runs.md). Ask for "auto", or for work no saved workflow fits, and it drafts a workflow card you tune and Save as workflow — or, when the work needs a script, a gate or a new agent, a graph card with draft cards for the new agents and scripts. Model and provider, workspace, Actions, team-metrics, team-policy and Away mode changes, creating, adding or cloning a project and reading a new website are cards too: nothing changes until you apply one. A started run becomes a live card — status, elapsed, cost, its agents and its workflow graph, with Open run — where you Pause, Resume or Stop the run and answer its questions; once it ends, the card shows the result, the top changed files, its artifacts and Create PR. While the chat follows a live run, the composer's → chip (named after the run) sends it a direction instead of a message. The chat still posts when the run waits for an answer, pauses, fails or finishes. A pull request opened from a chat card keeps reporting there: the card shows its Watch PR status, and each fix run that Watch PR or Resolve starts on it appears in the chat as a live run card, with what started it and how it ended.
• Agent mode with a shared terminal — Chat can run commands (tests, builds, linters, git) in a run's or a project's folder, in up to 3 terminals per chat that appear as tabs in the terminal pane, where you can type too. Each command shows in the chat as a card with Stop and Output; it changes files only through commands, and only when you ask. Shells get a cleaned environment and a command check, which is not a sandbox; a hosted Worca allows none unless WORCA_TERMINAL_REMOTE=1. See docs/terminal.md.
• Pinpoint — select text in a chat answer, a markdown artifact or the Diff tab to Explain it, ask for Alternatives, leave a note or comment, or send it to the chat; diff comments can become a fix run.
• Per-agent models for one run — on a run card for a saved workflow, Chat can pick a model and effort for each agent step from the catalog (judgement steps on the top price tier, mechanical ones on the cheapest capable model, Anthropic first, then OpenAI; a premium-tier model such as Fable 5.1 only when you name it). Each pick sits beside the project's own setting with its reason on hover; change or reset any of them before Start run. The picks apply to that run only — turn on Save as project defaults to keep them (docs/models.md).
• Files in the chat — when Ask makes or finds a file you should see (a screenshot it took with run_command, a chart, a report, a run artifact), it puts it into the reply as a files card: images inline, other files as rows, and a click opens a viewer with zoom, next/prev and Download. Show in folder opens the folder the file came from. Files are copied into the chat; credential files and anything outside the chat's folders are refused (docs/show-files.md).
• Memory — say "remember …" or state a lasting preference, and Ask saves it to Worca's memory: global for how you work, per project for facts about one repository. Pipeline agents load the same memory; view, edit, delete or restore it on Settings › Memory and on a project's Memory tab.
• Voice — the composer's mic dictates once, lets you talk and read the replies, or holds a hands-free conversation that reads replies aloud (hold the mic to start it). Whisper and Kokoro run in the browser, so audio never leaves the computer, or point either side at your own speech server; it needs https:// or localhost. It listens in ~99 languages and the built-in voice speaks English. See docs/speech.md and docs/speech-languages.md.

Browser extension

• Chat next to the page you are reading — a sidebar for Chrome, Edge and Firefox chats with your local Worca. Each message carries the page's URL, title and your selection, the page text if you switch it on, and any passage you quote. Pinpoint works on web pages too: select text, then Explain it, Discuss it or send it to the chat. You approve the pairing in Worca with a matching code, and Settings › General › Browser extension lists and revokes paired browsers. See docs/browser-extension.md.
• Ships with Worca — the built extension is in the npm package: open Settings › General › Browser extension, or run worca extension, for the folder to load; each GitHub release also has Chrome, Edge and Firefox zips (Firefox unsigned for now).

Away mode

• Runs keep going while you're away — during your away hours, or after you turn on Away in the account menu (the corner at the foot of the sidebar), Worca answers what a run would wait on: clarifying questions, mid-step questions, input forms, review-loop gates, proposed workflows and failed-step retries. It trusts the agent's recommendation when the agent is sure enough. Otherwise a read-only review (it can only read files), on the run's engine, scores each option against your Worca memory, how easily it can be undone, scope, the codebase's conventions and cost. An answer it is unsure of is still given, but flagged please check.
• You choose when and where — Settings › Runs › Away mode (Advanced level) sets the away hours and time zone, Only runs I marked or All runs, the method, the model and effort that weigh the options, the limits, and Always wait for me on… for kinds Worca must never answer. A project's Away mode tab overrides any of these for that project. You mark a run with Mark this run: worca may answer for me on New pipeline, or with --night on the CLI. A marked run is also answered by day once a question has waited 30 minutes. On the run page, Away mode on this run (As set up, Answer for me now, Never on this run) overrides one run.
• Bounded on purpose — while critical or major issues remain, a review loop gets at most one extra fix round (configurable), then continues with a flag. A failed step is retried up to 3 times, then the run pauses. A run pauses after 20 answers by default. An optional spend cap, counted across all projects from the start of the away stretch, pauses runs that reach it. Your own cost caps always hold; Worca passes the team's soft cap only if you allow it.
• Every answer on record — the run page lists Answered for you (question, answer, reason and the deciding model), one group per ask, with the answers to check marked and listed first. A note above it reads like "5 answers while you were away — 2 to check". A short notice says when away hours start and end, and a violet dot sits on your avatar while you are away; turning Away off in the account menu ends it.
• A small easter egg — while you are away with Keep awake on, leave the Worca window in front and untouched for 5 minutes and a screensaver takes over: the wordmark drifts across the window with one line on how your runs and chats are doing. Moving the mouse or pressing a key brings the page back. AWAY_SCREENSAVER in ui/public/away-screensaver.mjs turns it off.

Scheduled runs

• Run it later — start a pipeline once or on a repeat ("every weekday at 02:00", "every 15 minutes from 09:00 to 18:00") from New pipeline (Start run › Schedule…), an Ask card, the CLI (--at, --every, --cron) or the API (POST /api/run with scheduledFor or repeat). A series has overlap and failure-streak policies — by default it pauses itself after 3 failures in a row; Settings › Runs › Scheduled runs holds the defaults. See docs/scheduled-runs.md.
• Schedules view — sidebar Schedules (Advanced; shown at every level while anything is scheduled) with three tabs: Activity, a feed of every miss, failure, late start, skip and self-pause, whose problems raise an amber badge; Once (Run now, Change time, Cancel); Repeating (Run now, Skip next, Edit, a pause switch, Delete). Upcoming and missed runs also appear in Runs; worca schedule (list, show, run-now, move, cancel, skip, pause, resume, log) does the same from a terminal.
• When runs start — there is no daemon: a run starts while worca ui (or a worca … --at … --wait terminal) is up and the machine is awake. A slot missed by less than the grace window (6 h by default) starts late; past it, the run is marked missed and waits for Run now.
• Run trains — a run can start after another run instead of at a time (--after <id>, or Run after on a run page), by default only when that run ends done (--after-any also accepts a failed or stopped one), and on its feature branch with --source-from-previous — refactor → tests → docs, each on the branch the previous one left.
• Chat schedules too — "run the dependency upgrade every weekday at 2am" becomes a run card whose main button is Schedule. It lists, pauses, resumes and skips schedules on request; moving, editing, cancelling or deleting one is a card you confirm.

Costs & budgets

• Cost on every run — the run page shows each run's estimated cost as it goes; Expert mode adds per-step cost on the workflow graph. Figures are Claude Code's client-side estimates, not billing.
• Statistics (Advanced) — pick Today, This week, This month, Last month, All time or a Custom range for Spent, Saved (estimated human hours × your developer rate, minus spend), Pipeline spend, Chat, Time worked, Pipelines finished and PRs merged, plus Spend per and Runs per charts by hour, day or month (runs split into finished, stopped and failed). With a total limit, a ring around your avatar in the sidebar fills toward it (amber from 80%, red once new runs are blocked); the account menu's spend card has the figures. Filter or split spend by person (or by tag, see Shaper, cost tags and keep awake) and export the split as CSV; see docs/statistics.md.
• Hard limits — set in Settings › Runs › Budget & cost limits, or with worca config. A per-pipeline cost limit pauses a runaway run before its next step; Continue without cap (this pipeline) or worca resume <id> --ignore-cost-cap lifts it for that run. A total cost limit counts pipeline and Chat spend over a weekly (from Monday) or monthly (from the 1st) window; once reached it pauses running pipelines at their next step and blocks new runs, resumes and Ask turns until the window resets or you raise the limit. Team caps live in Team policy.

Team metrics

• A shared, git-backed record — every finished run (done, failed or stopped) is pushed as one file to an orphan worca-metrics branch on the project's own origin; there is no separate metrics server. Records carry the git user name (or none, a team choice made once at enable time), never an email address.
• Opt-in per project, with delegation — Set up team metrics… on a project's Team tab records on the repository itself (joining the branch if a teammate already made it), or delegates to another project that already records. A workspace records through one metrics home; Route all to metrics home delegates every member without a branch of its own. Include my runs stops only your machine from recording.
• A Team metrics page — sidebar Team metrics (Expert), for one project or workspace and a range from This month to All time or Custom: Spend, Runs, Cost per run, Duration, Autonomy and Review cycles tiles (plus Saved when runs carry human-hour estimates), spend and runs per week, breakdowns by workflow, ticket, actor and model mix, the runs table and Export CSV.
• A Timeline for planners — work items as bars on a calendar (click the header to drill from month to week to day), grouped by work item or person, with Shipped (the PR merged), In review, Needs attention, Median lead time and Agent spend tiles. Merge dates come from an optional GitHub Action or the GitHub CLI; without either the page still works and counts completed runs instead. With the Action, pull requests made outside Worca show too, so the calendar covers the whole team's delivery.
• CLI — worca metrics push [--project <path>] flushes pending run records, e.g. on a headless machine; worca metrics pr-workflow [--project <path>] [--force] [--print] writes that Action to .github/workflows/worca-metrics-pr-events.yml. See docs/team-metrics.md.

Team policy

• Team caps and expectations, shared through git — a lead publishes a policy to an orphan worca-policy branch on the project's origin (protect it so only maintainers push). Every teammate's Worca reads it on start and hourly; worca policy pull refreshes a CLI-only machine. A project carries its own policy or follows another project's; a workspace uses a policy home, a member each developer picks on their machine.
• Soft by design — each field is a default (applies until you set your own) or a soft expectation: for a cap the tighter of team and local applies, and an off-list model, a missing plugin or guardrails below the minimum tier warn and are recorded. Going past a team cap takes a confirmation (with a reason when the policy requires one) and is recorded to team metrics; unattended (--yes) runs warn instead of pausing. Only Chat's web fields bind: a policy can switch web access off or narrow its hosts, never widen them.
• What a policy can set — per-pipeline and per-period total cost caps (pause or warn on breach), an advisory pooled budget, Chat turn and cost limits, a default guardrail set and a minimum tier, allowed models and step defaults, marketplaces, required plugins and MCP servers, blocked plugins, a default workflow, human-in-the-loop, a minimum Worca version, Away mode defaults, and separate values for workspace runs. It can also ship its own guardrail sets and models. Required plugins and MCP servers come through a setup checklist and never install without your click (an opt-in per-home trust switch may install plugins, never MCP servers).
• The Team policy page — Manage › Team policy (Expert) shows the effective policy per project or workspace (team value, yours, what applies), the required plugins with Install… / Update…, and an editor that publishes in one commit. Turn a policy on from a project's Team tab (Set up team policy…); a workspace's Team tab picks its policy home.
• worca policy show|pull|init|setup — init --here creates the branch, init --follow <slug> follows another project, setup --install installs or updates the required plugins; continue past a team cap with --past-team-cap [--reason "…"] on a run or on worca resume. See docs/team-policy.md.

Models

• Four engines — agents run on Claude Code by default, or on OpenAI's Codex CLI, GitHub's Copilot CLI or Cursor's CLI agent, all three in beta. Pick one per run with Engine on New pipeline (Advanced) or --engine codex|copilot|cursor, or set a default on the Models page's Engine & models card or in a project's settings. Each needs its own CLI installed and signed in. The beta engines hold guardrail rules only in part, so Normal and Strict need Allow unguarded (--allow-unguarded-engine), or run them on Permissive (--guardrails permissive); Copilot and Cursor report no cost; Chat runs on Claude Code or Codex. A model's connection decides which engines can run it. See docs/models.md.
• Bring your own models — Add model on the Models page registers any model id (a proxy, a fine-tune, an alternative provider), declares which effort levels it supports, and attaches per-model routing env (e.g. ANTHROPIC_BASE_URL) that is merged into that model's agent spawns. Share as plugin… exports your models as a plugin scaffold: each env value is committed, omitted, or kept out of the repo as a secret each teammate enters once, after installing, under the plugin's Model secrets.
• No first-party account needed — the Title generation, Auto workflow model and PR description model cards on the Models page put Worca's own helper calls on a model you pick (left empty, a chat titles itself with its own model, a worca --model run with that model, and other runs with the built-in Haiku); endpoint-routed models carry Claude Code's internal haiku/sonnet/opus/fable tier keys so nothing falls back to the Anthropic API; and Hide built-in models drops the built-ins from every picker.
• GitHub Copilot and OpenAI-compatible endpoints, built in — Worca's own in-process bridge lets the Claude Code CLI run against a Copilot subscription (Claude models through Copilot's native Anthropic endpoint, thinking intact; GPT, Gemini and the rest translated to chat completions or the Responses API), any OpenAI-compatible server (llama.cpp, Ollama, LM Studio, vLLM, OpenRouter) or an Anthropic-compatible gateway — no LiteLLM, no second daemon. Sign in or set keys on the Providers page, then Import models… lists what a source serves, windows and capabilities included; worca models list and worca models import copilot|openai|openrouter do the same from a terminal. Both pages are Expert level. See docs/models.md.

Storage

• Nothing in your working tree — structured state (projects, workflows, run steps, Q&A, review verdicts) lives in one SQLite database, ~/.worca-cc/worca-cc.db; plan/review markdown and per-run artifacts live in a machine-wide store, ~/.worca-cc/store/<projectKey>/. Agent memory sits beside them in ~/.worca-cc/memory/.
• Keyed by repo identity — the project key is the repo's folder name plus a hash of its root (the parent of the shared .git), so every git worktree of a repo shares one history. A plain folder (a Folder project) is keyed by its own path.
• Move it — WORCA_HOME=<dir> (or Settings › General › Worca root folder, Advanced) puts everything under <dir>/.worca-cc/; only the global settings.json stays in your home folder.
• Projects folder — New (just a name) and Clone from URL put projects in ~/.worca-cc/projects/ (it follows WORCA_HOME and Settings › Worca root folder); New names the folder in lowercase with _ for spaces (Worca Presentation → worca_presentation), the project keeps the name as typed; WORCA_PROJECTS_ROOT or Settings › Projects root folder moves it. Existing projects stay where they are.
• One list for every project — Runs in the sidebar shows live, scheduled and finished runs from every project on the machine (see Pipeline). See docs/storage.md.

Shaper, cost tags and keep awake

• Shaper — Worca shapes a thin or large task first: at most 3 questions, a plan for the perspectives you pick, rule checks on the plan, and your approval before anything is built; a large task can be cut into slices. See docs/shape-it-first.md.
• Cost tags — key:value tags such as client:acme on New pipeline, run pages and Ask run cards (--tag on the CLI), optional mandatory tags under Settings › Tags, and spend By tag on Statistics and Team metrics.
• Keep awake — stops the computer sleeping while a run has agents working; a switch in the account menu. See docs/keep-awake.md.

Labs features

• Off until you turn them on — unfinished features ship behind machine-wide flags: Settings › Labs, worca flags on|off <key>, or WORCA_FLAGS=<key> at start. See docs/feature-flags.md.

Getting started & UI levels

• A guided first week — the first visit to New pipeline shows a one-time Welcome dialog with three doors (add a project, watch a run offline, run a real pipeline) and Skip. The Getting started pill under New pipeline in the sidebar, with its 3/9 count, opens a checklist of nine steps, from Connect Claude Code and Watch a run end to end to Group projects into a workspace and Set a team policy. Every tick is derived from what you have really done, never stored. Each tile, done ones included, starts a spotlight guide that rings the real control, says why, and moves on when you do the real thing; Esc or Skip leaves it. Hide from sidebar drops the pill and Settings › General › Getting started brings it back. See docs/getting-started.md.
• Three interface modes — Simple shows the core loop: New pipeline, Runs, Projects, Workflows (agents and scripts included), budget limits and Chat. Advanced adds Schedules, Statistics and Workspaces, plus branches, guardrails, per-agent models, the diff, pull requests, plugins, memory and the live log. Expert shows everything: Team metrics, Team policy, Models, diagnostics and every per-node tunable. Modes are cumulative, so raising one never removes anything.
• A view preference, not a permission — Interface mode in the account menu (the corner at the foot of the sidebar) names the current mode and lists the three. A choice applies at once, for every browser that opens this Worca; Settings › General › Interface mode › Change… opens the same choice as a dialog. A fresh install starts at Simple, one that already has projects or runs at Expert. A link to a page above your mode still opens it, under a banner offering Switch to …; values set in a higher mode keep applying, and a question a run waits on shows in every mode. A guide step that needs a higher mode asks first (Switch to Advanced?). See docs/ui-levels.md.

Install

npm install -g @worca/app     # puts `worca` on your PATH

Install with the npm of the Node that is first on your PATH in a new terminal. A global package lands under the Node that ran npm install -g: if nvm was active then, but a new terminal starts with the system or Homebrew Node, worca is not found there. Check in a fresh terminal:

command -v worca    # prints the path of the worca binary; nothing means it is not on PATH

npm install scripts. Newer npm versions warn about, or skip, the install scripts of dependencies you have not approved. Approve the two Worca needs:

npm install -g @worca/app --allow-scripts=node-pty,protobufjs

node-pty is optional and is used only by the web UI's terminal pane (and the commands Ask Worca runs in it); pipelines, mock or real, never load it. On Linux its install script compiles it, so without it the terminal falls back to plain pipes. On macOS the prebuilt spawn-helper can lack the execute bit; Worca sets it the first time a terminal opens, and worca doctor prints the exact chmod +x <path> when it cannot (prefix sudo for a root-owned global install). protobufjs's script only prints a version warning. See Without node-pty.

Requirements:

• Node.js >= 22.13.0 — the store is the built-in node:sqlite; on an older Node the CLI says so and exits.
• git — pull requests, Check out, Sync, Clone from URL and Use git need it (on Windows, Git for Windows' git.exe on PATH). Without git — on macOS, without the Xcode Command Line Tools — worca still opens and never runs git: it hides the branch, pull-request, Sync and Clone controls, a project run works on a copy of its folder (applied back with Apply to folder), a workspace run works on a copy of each member (each applied back on its own), and Workspaces, Marketplace, skill import, Team metrics and Team policy stay: GitHub-hosted sources work over HTTPS and other hosts say Needs git; installing git brings everything back without a restart. WORCA_GIT=0 hides them on purpose, WORCA_GIT=1 skips the check. The GitHub CLI (gh) is optional: it opens pull requests and issues.
• The Claude Code CLI (claude) on your PATH and signed in (claude, then /login; an ANTHROPIC_API_KEY works too) — for real runs only; mock mode starts no claude at all. WORCA_CLAUDE_BIN points Worca at another binary.
• Native Windows — both the native installer (claude.exe) and npm install -g @anthropic-ai/claude-code work: npm puts a claude.cmd shim on PATH that Node cannot spawn, so Worca runs the package's native claude.exe next to it instead. If that binary is still npm's placeholder (the package's postinstall didn't run) or there is none, Worca says what to do rather than a bare ENOENT; WORCA_CLAUDE_BIN can always point at a claude.exe directly.

In a container

Prefer the agents to run in a disposable Linux box instead of on your machine? The same Worca ships as an image (ghcr.io/sinishadjukic/worca, amd64 and arm64, plus a -full variant) with a Compose file. The agents cannot read your host credentials or touch files outside the project folders you mount; the box keeps its own Worca home and Claude login. Needs Docker or Podman with podman compose.

worca container up        # writes ~/.worca-cc/container/, starts the box; UI on http://localhost:4317
worca container login     # log Claude Code in, once

worca container help lists the rest (down, logs, pull, `sh