FluidMPC native
v0.1.0 Workspace

Local MCP server / 21,505 reasoning cells / one control plane

Twenty-one thousand skills. Your agent reads sixteen.

FluidMPC holds the catalogue, scores it against the task in front of your agent, and installs only what the evidence supports. The optional workspace adds one MCP connection for the Project Contract, reviewed integrations and a human approval before anything reaches production.

Open the workspace Download the desktop app How a prompt becomes a plan
Catalogue
local, one JSON
Control plane
opt-in workspace
Approval
before production
Licence
MIT
Control plane / optional workspace

One connection. Four layers. One approval.

Your agent should not choose between twenty MCP servers and a twelve-page tool dump. One endpoint gives it the capability it needs, the decisions you already made, an executable procedure and a place to stop until you approve the exact production action.

Sign in with Freebuff Explore the sandbox
  1. 01

    Connect

    Add one stdio server to Codex, Claude Code, Cursor or Claude Desktop. No second entry for skills, context or pipelines.

    fluidmpc-native --mcp
  2. 02

    Contract

    Frontend, database, accounts, billing, deployment and the rules that never change, held once for every agent.

    get_project_context
  3. 03

    Execute

    Validation, executable skills and browser checks run in order. Each step records its result instead of disappearing.

    run_pipeline
  4. 04

    Approve

    Production pauses. You see the target, the action and the risk, then approve or reject the exact change.

    request_approval
01

Shared context

Decisions, rules and boundaries live in one Project Contract. A new agent starts from the same facts instead of asking again.

02

Scoped access

Every integration lists the permissions it would request. Nothing connects until you review and confirm each one.

03

Executable skills

Procedures carry the questions to ask and the checks that prove the work, so agents follow the same method every time.

04

Approval gates

A production deliver is refused unless an approval exists for that run. Rejecting stops the pipeline cleanly.

05

Run history

Inputs, steps, durations, browser results and the person who approved it — the record survives after the run ends.

06

Scheduled pipelines

Run validation and QA on a cadence. Scheduled runs still stop at the same gate before production.

Project Contract revision 4
frontend
Next.js 16 · Vercel
database
Postgres · Supabase
accounts
Freebuff OIDC · teams
billing
Stripe · seat plans
deployment
https://northwind.example
permanent rules

Nothing reaches production without an approved run.

A message is readable only by members of its workspace.

run_8f2c · Release waiting approval
  1. Contract loaded5 decisions · 3 rules · 0.4s
  2. Validation passedfrontend · database · accounts · deployment
  3. security-review5 procedure steps · 3 checks
  4. Browser check200 · 214ms · contains “Example Domain”
  5. Approve productiontarget production · high risk
  6. Deliverblocked until the approval exists
needs youDeploy Northwind to production

High risk · exact action, not a category.

Approve Reject

The catalogue and the second brain stay on your disk. The workspace is opt-in: it stores the contract, integrations and run history so the same project works from the desktop app and the browser.

01

The route, not the pile.

A catalogue is only useful if something narrows it. FluidMPC narrows it per task, in order, and shows its work at each step.

01

Select

Installed is active. Thirty-four presets stand as doctrine — mission, verification, minimal change, comprehension — and any of them can be muted per workspace when the task is not about code.

mutual exclusion: none · presets cannot be uninstalled, only muted

02

Route

route_task hands the task to the local router. Laya, an optional self-hosted model on loopback, adds a field of view: task type, complexity, tool need, security sensitivity. The catalogue adds the evidence — lexical and latent-concept scoring over every cell.

laya offline? the plan still returns, on rules alone

03

Plan

One RoutePlan: primary picks with the exact words that matched, an adjacent pool of weaker but real evidence, the funnel counts, and the reason it stopped at sixteen. The agent states what it took and why, then leaves the rest unused.

relevance beats completeness · no model call on the server

04

Load

run_skill returns a cell's working instructions — the checklist, the failure modes, the order of operations. Your model does the reasoning. FluidMPC never pretends to.

unselected cells are refused, not quietly served

02

One window, if you want it.

The MCP server is the product. The deck is how you curate it: every figure printed in it is measured from your own rows, and a figure that does not exist is an em dash — never a zero, never a sample.

FluidMPC overview: readouts, agent pulse, family coverage and recent work
Overview Four live readouts, the agent pulse, fourteen-family coverage, and the packs worth installing next.
FluidMPC skill hub: the catalogue as a sortable table with family rail
Skill Hub The catalogue as a real table — family rail, status filters, live run and latency columns.

No title bar

The frame is a single floating island that opens downward, never wider. Your window is the screen.

Keyboard first

ctrl k for the palette, / to search, j/k to move, i to install, ? for the rest.

Honest history

Sample rows are labelled, and excluded from your measurements. A dry run says it is a dry run.

03

Twenty-seven packs. Fourteen families.

The catalogue is reference data compiled into the binary: public, versioned, and installed with its notices. Hover a family to see what it is for.

  • Engineering5,105

    Build, ship and operate software: languages, frameworks, tests, infrastructure and the review before merge.

  • Modes5,111

    How the agent operates: mission discipline, reading the room, and the rules that hold work to its instruction.

  • Research1,944

    Find, triangulate and cite — and say what the evidence does not support.

  • Data & AI1,484

    Numbers, pipelines and models: schemas, metrics, analytics, training, evaluation and inference.

  • Security1,458

    Audit, verify, redact and probe before anything leaves the building.

  • Automation1,187

    Triggers, retries, integrations and agents: the plumbing that runs without a person watching.

  • Design946

    Interface, visual and interaction craft: hierarchy, tokens, states, accessibility and brand.

  • Business981

    Commercial facts: pricing, revenue, contracts, marketing, sales, hiring and operations.

  • Reasoning829

    How an agent argues with itself: assumptions, alternatives, critique, calibration and the decision record.

  • Vision584

    Read images, charts, screenshots, documents and video — and mark what the pixels do not settle.

  • Writing554

    Prose, documentation and the edit that makes a reader act: structure first, then sentence, then voice.

  • Product524

    Discovery to spec: the problem, the scope, the proof of done and the plan that ships it.

  • Communication456

    Email, meetings, decks and the update that reaches the person who has to decide.

  • Memory342

    Context budgets, recall, provenance and the facts worth carrying into the next session.

Total 21,505 27 curated packs
04

Always on, and still yours.

Thirty-four presets are active in every workspace. They are the difference between an agent that finishes a job and one that reports progress. Mute one — it stays in the catalogue, it just stops instructing this workspace.

Installed cells are how you add to that baseline. Presets cannot be uninstalled; only muted.

Mission & discipline

  • mission-complete
  • goal-contract
  • definition-of-done
  • instruction-re-reader
  • no-guess-protocol
  • blocked-path-finder
  • verification-before-stop
  • completion-gate
  • skill-router

Change safety

  • minimal-change-engineer
  • change-impact-analyst
  • regression-sentinel
  • release-readiness-gate
  • diff-reviewer
  • test-weaver
  • stack-trace-surgeon
  • refactor-cartographer

Contracts & trust

  • contract-designer
  • schema-architect
  • secure-code-reviewer
  • dependency-auditor
  • secret-sweep
  • injection-scout
  • threat-model

Words & memory

  • spec-writer
  • technical-writer
  • context-window-budgeter
  • fact-ledger
  • gap-finder

Understanding code

  • understand
  • understand-diff
  • understand-chat
  • understand-explain
  • understand-onboard
05

Two processes, one file.

The window and the server write the same workspace. That is the whole architecture problem, and it is solved rather than hidden: a lock, a three-way merge, a revision counter, and a damaged file moved aside instead of overwritten.

workspace.json 
{
  "schemaVersion": 2,
  "revision": 26,
  "name": "FluidMPC Lab",
  "installed": [{ "skillId": "mission-complete" }],
  "notes": [{
    "id": "note_1a16ee2a1e6a",
    "title": "Dynamic island",
    "body": "One capsule, opened downward.\nBuilt on [[FluidMPC]].",
    "tags": ["interface"],
    "stub": false,
    "updatedAt": 1792854507699
  }],
  "runs": [{ "status": "success", "completion": 100 }],
  "mcpCalls": 22,
  "appearance": {
    "accent": "phosphor",
    "theme": "phosphor",
    "density": "comfortable",
    "customThemes": []
  }
}

Notes are nodes, [[links]] are edges, a missing title becomes a stub. One file, held by one lock, merged by revision.

  • Serialised writes

    A lock file holds each read-modify-write. A lock older than five seconds is reclaimed.

  • Three-way merge

    Installs and runs from both sides survive; a removal applies to the other side; the newer edit wins a note conflict.

  • Nothing lost quietly

    The previous generation stays as .bak, writes go through a temp file and a rename, and an unparseable file is moved to workspace.corrupt-…json.

06

One server to register.

Eighteen tools, four resources, JSON-RPC 2.0 over stdio. Laya is never a second entry — route_task reaches it for you and returns its signals inside the same plan.

mcp.json 
{
  "mcpServers": {
    "fluidmpc": {
      "command": "fluidmpc-native",
      "args": ["--mcp"]
    }
  }
}

Started with --mcp it draws nothing. Started normally it opens the deck. Same binary, same workspace, one lock between them.

Tools

  • search_skills
  • recommend_skills
  • route_task
  • get_skill
  • install_skill
  • uninstall_skill
  • list_collections
  • list_categories
  • list_skills_by_category
  • run_skill
  • workspace_manifest
  • presets
  • list_runs
  • brain_search
  • brain_read
  • brain_write
  • brain_graph
  • brain_tags

The Connect screen launches the binary, runs initialize and tools/list, and reports protocol, version and tool count — so a broken client is diagnosed before an agent sees it.

FluidMPC connect screen: four client configurations, the tool and resource counts, and the handshake test
Connect Four client configurations, the whole tool surface, the resources, and the optional local Laya router — one server, one command.
07

Take it offline.

No runtime, no account, no telemetry. Install it or unpack it, run it, register the server. The catalogue and the fonts are compiled in; Laya stays an optional sidecar you host yourself.

Windows · x64 FluidMPC-0.1.0-setup.exe The 9.6 MB installer: per-user or per-machine, a Start menu shortcut and an uninstaller. Windows 10 or newer. 9.6 MB · GUI + MCP server · symbols stripped Download the installer Windows · portable FluidMPC-0.1.0-windows-x64.zip Unpack the 11 MB zip and run fluidmpc-native.exe. No installer, no registry entries, no admin rights. 11 MB · portable folder · GUI + MCP server Download the portable zip

Verify either file against SHA256SUMS.txt. Windows may show SmartScreen because this build is not code-signed. Linux and macOS builds are compiled on their target platform from the same crate; the scripts are in fluidmpc/installer.

Build it yourself

Rust 1.85+, edition 2024. Fat LTO, one codegen unit, symbols stripped — hardened, not encrypted. Windows, Linux and macOS installer scripts are in fluidmpc/installer.

terminal 
cd fluidmpc
cargo build --release --bin fluidmpc-native

# Windows helper
build.bat

# Linux / macOS installers
installer/build-installer-linux.sh
installer/build-installer-macos.sh
08

Plans for every stage.

The catalogue is local and free. The workspace adds accounts, shared context, reviewed integrations, scheduled pipelines and the approval gate. Every plan is free while the control plane is in early access.

Free

$0

For your first connected project.

  • 1 project
  • 3 reviewed integrations
  • Public skills and pipelines
  • Manual runs, 5 a day
  • Project Contract and approval gate
Start free

Builder

$0

For the products you are building.

  • 5 projects
  • 10 reviewed integrations
  • Premium skills and packs
  • Scheduled pipelines
  • Browser checks and run history
Explore Builder

Pro

$0

For a complete release loop.

  • 50 projects
  • Every integration
  • Scheduled pipelines
  • Browser checks
  • Advanced history and tokens
Go Pro

Billing is not active. Plans and limits exist so the shape is honest; nothing is charged during early access.

What is FluidMPC?

A native desktop instrument and an MCP server for the skill catalogue, plus an optional control plane for the decisions, integrations and pipelines around a product.

Do I need an account?

No. The catalogue, router, second brain and MCP server work entirely offline. An account only adds the shared workspace and desktop sync.

How does the approval gate work?

A pipeline that would deliver to production creates an approval request with the target, the action and the risk. The run pauses. Nothing continues until a person approves it, and a rejection stops the run cleanly.

Which clients can connect?

Codex, Claude Code, Claude Desktop, Cursor or any MCP client that speaks JSON-RPC over stdio. One server entry reaches the catalogue, the workspace tools and the pipelines.

Where does the data live?

The catalogue and second brain stay in the workspace file on your disk. The hosted workspace stores projects, contracts, integrations, runs and approvals in D1, scoped to your Freebuff account.

09

What changed.

The release notes for the control plane and the desktop instrument. Newest first; every entry links to the screen that shows it.

  1. Ship composer and a wider MCP control plane

    The workspace overview now takes a request directly: type / for a pipeline, @ for an integration, or a sentence to get a route plan with the capabilities that match and the context still missing. The MCP server grew to 29 tools, including get_project_context, contract writes, approvals and pipeline runs that sync to the same account.

  2. Accounts, workspaces and desktop sync

    Sign in with Freebuff, create a project, and link the desktop app with a workspace token. The contract and pipelines are shared; runs and approvals land in the same history.

  3. Approval gate before production

    A run that would deliver to production now creates an approval request and stops. The approver sees the exact target, action and risk; a rejection stops the run cleanly.

  4. Pipeline builder and browser checks

    Build a pipeline from context, validation, skills, review, approval and delivery steps. Browser checks fetch the deployed page and assert status and content.

  5. Project Contract editor

    Frontend, database, accounts, billing, deployment, permanent rules and open questions in one editable contract, loaded by every agent.

  6. Integrations with scope review

    Sixteen integrations list the permissions they would request. Nothing connects until every scope is confirmed, and disconnecting is always available.

Start with the sandbox

Connect one endpoint. Load the contract. Run the first pipeline.

Explore the workspace Sign in with Freebuff