Workfile MCP server

@illodev/workfile is a repository-native protocol for coordinating Work, Docs, History and durable project Memory between humans and software agents.

Markdown files in the repository are canonical. The CLI, HTTP API and local UI use the same core services, collection registry, index and validation rules. No exclusive state is kept in the browser or in a database.

Work, Docs, History and Memory share the common ProjectRecord index. The core, CLI, HTTP server and MCP runtime are authored in TypeScript and distributed as compiled ESM with public declarations. The local UI is precompiled and included in the package, and semantic search runs on-device through the optional @illodev/workfile-search-local workspace package.

Try the live demo — it replays this repository's own workspace: the real cards, releases, incidents and learnings of Workfile's development. Mutations work per browser session and reset on reload.

https://github.com/user-attachments/assets/c9cd3035-6729-4cda-9172-984829ab5dbc

Used by

Boundaries

Workfile records work. It does not configure agents.

The two get confused because both live next to the same repository. Ecosystem configurators — gentle-ai is a good example — install a persona, curated skills, model routing, MCP servers and review gates into the agents you already use, across many agents at once. Their question is how your agent works. Workfile's question is what was done, who holds it and on what evidence, and its answer is Markdown files that outlive the agent, the session and this package.

They compose. A well-configured agent still needs somewhere durable to write down what it did.

What is here, and is not a configurator's job:

• The repository is canonical. A card is a file in the pull request: reviewed in the diff, reported by workfile doctor when malformed. No exclusive state in a browser, a database or ~/.config. Remove the package and the records stay readable.
• Claims are enforced, not agreed. Ownership is checked at the mutation, so a card another actor holds refuses your transition with CARD_CLAIM_OWNER_MISMATCH instead of quietly accepting it — a guarantee no sentence in a prompt can make.
• review is not done. done requires evidence from somewhere the code actually ran. A merge is not evidence.
• Humans read the same records. The UI, the rendered changelog and the releases are derived from exactly what the agent writes; there is no machine view and human view to keep in sync.

What is deliberately absent: Workfile does not install or update agents, ship a persona, route models or curate a skill catalogue. It syncs its own protocol into the instruction files an agent already reads (workfile agents sync) and exposes every operation over MCP — vendor neutral, but a server, not an ecosystem.

Requirements

• Node.js 22 or newer.
• npm, pnpm, yarn or Bun may invoke the package.

Install

Every workfile … command in this README requires the package to be installed — pnpm dlx / npx one-offs run a command and discard the binary afterwards:

pnpm add -D @illodev/workfile     # per repository (recommended)
pnpm workfile doctor              # dependency bins run through pnpm / npx

pnpm add -g @illodev/workfile     # or globally: `workfile` lands on your PATH
workfile doctor
wf doctor                         # `wf` is the same binary, for typing by hand

wf is an alias, not a rename: both names reach the same entry point, and the help and error hints answer in whichever one you typed. Keep the long form in anything generated or shared. wf only resolves once the package is installed, and an unrelated wf exists on the registry — so npx wf would fetch someone else's tool where npx workfile fails outright.

pnpm dlx @illodev/workfile init is fine for one-shot initialization, but keep the package as a devDependency afterwards: that is what makes the project* scripts that init adds to package.json resolve. That prefix is an npm script namespace — pnpm project opens the UI, pnpm project:doctor runs the checks — and has nothing to do with the old binary name.

TypeScript API

The published surface exposes JavaScript and declarations through conditional package exports. TypeScript consumers receive typed configuration, workspace, record, search and integration contracts from the root package and every documented subpath:

import {
    defineProject,
    type CardStatus,
    type ProjectConfig,
    type ProjectRecord
} from "@illodev/workfile";
import { createSemanticSearchProvider } from "@illodev/workfile/search";

const config: ProjectConfig = defineProject({
    schemaVersion: 2,
    name: "Billing",
    cards: {
        areas: ["api", "web"]
    }
});

const status: CardStatus = "doing";

The CLI and UI do not require TypeScript in consuming projects. React, Primer, Vite and the UI type packages are build-only dependencies; the installed package serves bundled browser assets from dist/ui.

Workspace

A project is discovered through project.config.mjs and normally stores protocol-owned files under .project/:

project.config.mjs
.project/
├── VERSION
├── cards/
│   └── archive/
├── assets/
├── docs/
├── changelog/
│   ├── unreleased/
│   └── releases/
├── memory/
│   ├── learnings/
│   ├── decisions/
│   ├── incidents/
│   ├── conventions/
│   └── context/
├── agents/
└── .cache/

Minimal configuration — a plain object, not defineProject(...). The loader applies defineProject itself, and an import here is a bare specifier the file can only resolve with node_modules present, which breaks the two consumers that run without one: a pnpm dlx-initialized workspace before the package is installed, and the generated CI job's npx run on a clean clone. The JSDoc annotation keeps editor typing without a runtime import:

/** @type {import("@illodev/workfile").ProjectConfigInput} */
export default {
    schemaVersion: 2,
    name: "My project",
    cards: {
        areas: ["api", "web", "infra", "docs"]
    },
    docs: {
        sources: [
            "README.md",
            "docs/**/*.md",
            "apps/*/README.md",
            ".project/specs/**/*.md"
        ]
    },
    changelog: {
        releaseStrategy: "semver",
        defaultVisibility: "public"
    },
    memory: {
        collections: [
            "learnings",
            "decisions",
            "incidents",
            "conventions",
            "context"
        ]
    },
    agents: {
        targets: ["agents-md", "cursor"]
    },
    ci: {
        targets: ["github"]
    },
    mcp: {
        allowMutations: true
    },
    search: {
        semanticWeight: 0.35,
        maxProviderRecords: 500
    }
};

Project-specific areas, paths and vocabularies are resolved at runtime and exposed through the effective schema. The eight Work statuses and the schema-v2 memory collection semantics remain protocol contracts.

Work

Cards are managed Markdown records under .project/cards/. The Work module provides hierarchy, dependencies, claims, scope, status transitions, archives, assets and conflict-aware writes.

workfile card list --json
workfile card show T-0042 --json
workfile card create --title "Implement runtime schema" --area infra
workfile card create --json-input card.json   # body, parent, source and tags in one call
workfile card claim T-0042 --scope apps/api,packages/sdk   # actor resolves itself
workfile card transition T-0042 review
workfile card patch T-0042 --json-input changes.json --expected-revision sha256:...
workfile card archive T-0042
workfile card reopen T-0042 --status backlog

Docs

Docs combines two sources without copying existing documentation:

• Indexed documents discovered from configured globs. They receive deterministic PATH-* IDs and remain read-only through the protocol.
• Managed documents stored in .project/docs/ with stable DOC-NNNN IDs, typed frontmatter and revision-aware mutations.

Managed documents are read recursively, so they can be grouped in folders — including folders you create by hand. IDs stay global and sequential: a folder is organization, not identity. New documents follow docs.layout (kind, the default, groups them by document kind; flat writes them to the managed root) and --folder overrides it.

workfile doc list --query billing
workfile doc show DOC-0012 --json
workfile doc create --title "Deployment runbook" --kind runbook --status current
workfile doc create --title "Rate limiting" --folder architecture/billing
workfile doc move DOC-0012 --folder architecture
workfile doc patch DOC-0012 --json-input changes.json --expected-revision sha256:...

The doctor detects broken local links, unresolved related or superseded records, missing scope paths and stale review/source relationships.

History

History uses atomic change fragments rather than asking multiple branches or agents to edit one shared CHANGELOG.md.

Unreleased fragment:

---
id: CHG-0042
title: Add portable history workspace
type: added
area: infra
visibility: public
cards: [T-0042]
created: 2026-07-28
updated: 2026-07-28
---

A release consumes selected fragments, moves them beneath the release directory and creates a canonical REL-NNNN record. Public or internal changelogs are derived output.

workfile changelog list --unreleased
workfile changelog add --title "Add portable history" --type added --area infra
workfile changelog preview
workfile changelog release 0.4.0 --title "History and Memory"
workfile changelog render --visibility public
workfile changelog render --visibility public --write
workfile changelog verify

Release versions can use semver, calendar or freeform validation according to configuration. Fragments and releases participate in the same workfile search and backlink graph as cards, docs and memory.

Memory

Memory is a set of typed, atomic and lifecycle-aware records rather than a single growing conversation transcript:

Collection	Prefix	Purpose
Learnings	LRN	Reusable observations with confidence and occurrence signals
Decisions	ADR	Proposed, accepted, rejected or superseded decisions
Incidents	INC	Operational events, severity, timing and corrective actions
Conventions	CONV	Durable rules followed by humans and agents
Context	CTX	Useful but potentially expiring project state

workfile memory list --collection learnings --status active
workfile memory add learning --title "Atomic fragments avoid merge conflicts" \
  --confidence high
workfile memory add decision --title "Keep Markdown canonical" --status accepted
workfile memory add incident --title "Release pipeline stalled" --severity high
workfile memory graduate LRN-0004 --to CONV-0002,DOC-0012
workfile memory supersede ADR-0003 --by ADR-0009
workfile memory patch CTX-0002 --json-input changes.json --expected-revision sha256:...
workfile memory verify

The doctor checks invalid lifecycle states, missing graduation/supersession targets, expired context and incomplete incident resolution metadata.

Unified index

Every module normalizes its files as ProjectRecord entries through a common collection registry. The derived process-local index provides:

• weighted full-project text search;
• lookup by stable record ID;
• outgoing references and incoming backlinks across all four domains;
• card source: links and local Markdown links;
• module-specific health, lifecycle and freshness signals;
• module and collection counts.

Canonical state always remains on disk. The server cache is short-lived, invalidatable and fully rebuildable.

workfile search "billing architecture"
workfile search release --kind change,release,memory --limit 25 --json

Initialization

The initializer can run interactively or deterministically in automation. It detects the package manager, monorepo folders, likely card areas, documentation sources, existing agent environments and CI providers. A dry run exposes the exact filesystem plan.

pnpm dlx @illodev/workfile init
pnpm dlx @illodev/workfile init --yes \
  --agents agents-md,claude,cursor,copilot --ci github
workfile init --dry-run --json

The generated project.config.mjs exports a plain object, so a workspace initialized via pnpm dlx remains loadable before the package is installed locally. Existing files are not overwritten unless --force is explicit. .project/.cache/ is added to .gitignore; all canonical protocol files remain tracked.

Hosted demo

The UI ships with a demo mode for static hosting (Vercel, GitHub Pages, any file server). npm run build:demo builds the UI with an in-memory API that replays a snapshot of a seeded workspace: every view works and mutations behave normally for the session, then reset on reload. The repository includes a vercel.json, so importing it into Vercel deploys the demo with zero configuration.

pnpm run demo:data   # reseed and resnapshot packages/workfile/ui/src/demo-data.json
pnpm run build:demo  # static demo build into packages/workfile/dist/demo

Regular builds tree-shake the demo layer and snapshot out of the bundle.

Releasing

Releases publish from CI via npm trusted publishing (OIDC) — no npm token is stored in the repository. Before cutting, if any UI change shipped since the last tag, run pnpm run screenshots: it recaptures every view and publishes the README and landing-page stills itself, and they belong in the commit that moved the shell, not in the version bump. The circuit:

1. Cut the changelog: workfile changelog release <version> and workfile changelog render --write.
2. Bump and tag: npm version <version> then git push && git push --tags. The version hook carries every packages/* package inside the same bump — workspace packages always ship the core's version.
3. The Release workflow verifies the tag matches package.json (and that no workspace version drifted), runs check:release (build, typechecks, tests, audit and a packaged-tarball smoke) with pnpm, and publishes the core and every workspace package with the npm CLI under latest.

There is no prerelease channel: every published version is one npm install away, and a v*-rc.* tag fails the release rather than publishing. That is a consequence of trusted publishing rather than a preference — OIDC authorizes npm publish and no other registry write, so CI cannot move a dist-tag off a release candidate once it has been set.

Agent Protocol

Canonical instructions and workflows live under .project/agents/. Compact managed blocks are synchronized into supported environments without replacing unrelated user content:

AGENTS.md
CLAUDE.md
.cursor/rules/workfile.mdc
.github/copilot-instructions.md

workfile agents sync
workfile agents sync --targets agents-md,claude,cursor,copilot
workfile agents check
workfile agents context --card T-0042

Managed blocks carry the package version and a SHA-256 digest. agents check and workfile doctor report missing, unmanaged or stale generated instructions. Agent context is bounded and prioritizes the selected card, direct relationships, active conventions, unresolved incidents and non-expired context instead of loading all workfile memory.

Model Context Protocol

Workfile includes a local, dependency-free MCP server using UTF-8, newline-delimited JSON-RPC over stdio. It delegates every operation to the same core services used by the CLI and HTTP API, speaks both the modern (2026-07-28) and legacy (2025-11-25) protocol revisions, and exposes 32 tools, four resources and three prompts. Mutation tools disappear entirely in --read-only mode.

Point a client at it without installing anything. This is the invocation the official registry publishes for io.github.illodev/workfile, and what most clients will build for you from that listing:

{
  "mcpServers": {
    "workfile": {
      "command": "npx",
      "args": ["-y", "@illodev/workfile", "mcp"]
    }
  }
}

mcp there is a subcommand, not a binary: npx resolves the bin whose name matches the package and hands it everything that follows. Append --root PATH when the client starts somewhere other than the workspace — it searches upwards for .project/ otherwise — and --read-only to serve the read tools alone, with every mutation refused.

With the package installed, the same server is a subcommand away:

workfile mcp
workfile mcp inspect --json
workfile mcp config --read-only --json

For Claude Code the same surface ships as a plugin — the MCP server plus /claim, /card-context, /next and /done commands, a skill, and hooks that turn card claims into an executable guard rail — with no generated files committed to the repository:

/plugin marketplace add illodev/workfile
/plugin install workfile@illodev

The full contract — tool inventory, resources, prompts, process hygiene and the plugin's surface — is documented in docs/mcp.md.

The server is listed on Glama, which builds it in a container and inspects the capabilities it reports:

Search integrations

Lexical search remains deterministic and local. Hosts may inject an optional semantic provider programmatically; Workfile never selects a vendor or sends repository content over the network by itself.

First-party: local embeddings

@illodev/workfile-search-local runs embeddings on-device (onnxruntime-web, ONNX on CPU, Xenova/multilingual-e5-small quantized) — repository content never leaves the machine. Declare it in project.config.mjs with a guarded import, because the config must also load where the package cannot resolve (the generated CI job runs npx on a clean clone):

export const integrations = await (async () => {
    try {
        const { localSearchIntegration } = await import(
            "@illodev/workfile-search-local"
        );
        return [localSearchIntegration()];
    } catch {
        return []; // package absent: search stays lexical
    }
})();

export default {
    // …
    search: { provider: "local-embeddings" }
};

Know the cost model before wiring it: the first hybrid search embeds every uncached candidate record — minutes of sustained CPU on a few-thousand-record workspace, triggered by whichever surface searches first (CLI, board UI, or the MCP server an agent loads). The provider caps ONNX at half the cores by default, persists per batch so an interrupted pass resumes instead of restarting, and reports progress on stderr; sizing search.maxProviderRecords to your corpus makes every record eligible. Details and options in packages/search-local/README.md.

Bring your own

import {
    createSemanticSearchProvider,
    searchProjectRecordsHybrid
} from "@illodev/workfile/search";

const provider = createSemanticSearchProvider({
    id: "company-embeddings",
    async search({ query, records }) {
        // Return [{ id, score }] with scores between 0 and 1.
        return rankWithYourApprovedProvider(query, records);
    }
});

const result = await searchProjectRecordsHybrid(index.records, query, {
    provider,
    semanticWeight: 0.35
});

The adapter boundary makes external data disclosure an explicit host decision and keeps the canonical Markdown/index implementation provider-independent.

Experimental integration registry

Programmatic hosts can group approved semantic search and health adapters in a small, vendor-neutral registry:

import {
    createIntegrationRegistry,
    defineProjectIntegration
} from "@illodev/workfile/integrations";

const integrations = createIntegrationRegistry([
    defineProjectIntegration({
        id: "company.platform",
        semanticSearchProvider: provider,
        async healthCheck({ workspace, index }) {
            return [];
        }
    })
]);

The registry is accepted by the MCP server and doctor APIs. It is intentionally limited in the current RC: vendor-specific issue trackers, deployment systems and credentials are not part of the canonical package. The boundary can mature from real integrations without committing the schema to GitHub, GitLab, Jira or a deployment provider.

CI templates

CI files use the same managed-file contract and can be generated for GitHub Actions, GitLab CI or a generic shell runner:

workfile ci sync --targets github,gitlab,generic
workfile ci check

Templates run both the workfile doctor and agent synchronization check against the pinned Workfile version.

Legacy migration

The v1 .planning system can be planned and applied with deterministic collision checks:

workfile migrate plan --source .planning
workfile migrate apply --source .planning --mode copy
workfile migrate apply --source .planning --mode move

Valid legacy cards and assets become canonical v2 Work records. Old proposals, changelogs, learnings and malformed records are preserved under .project/sources/legacy-planning/ rather than being silently reinterpreted with an incompatible schema. Every applied migration writes .project/migrations/legacy-planning.json with source, destination, digest and result metadata.

General CLI

workfile init
workfile schema --json
workfile doctor --json
workfile ui
workfile ui --read-only --host 0.0.0.0 --allowed-host board.example.com

Commands return stable machine-readable errors with --json. A stale revision exits with code 3; configuration errors exit with code 2; validation and not-found errors exit with code 1. The complete command surface is documented in docs/cli.md.