blankstate-mcp

Protocol-based sensing and validation for autonomous AI agents.

PyPI: [blankstate-mcp](https://pypi.org/project/blankstate-mcp/) · import package: blankstate_mcp

The Blankstate MCP Server exposes Blankstate measurement, Programmatic Dive, and Stage / World authoring as MCP tools. It lets any MCP-compatible agent (Cursor, Claude Desktop, and others) sense interactions against Blankstate Protocols, validate actions before executing them, run a Stage dive (open → act → status → close), author a Stage, run and read its Worlds, and check quota status.

Protocols are universal sensors — the same protocol measures any content modality: conversations, commands, documents, agent actions. The measurement engine determines signal, not the input type.

Features

• Twenty-three MCP tools
  • Sense: bks_sense, bks_validate, bks_status
  • Dive: bks_dive_open, bks_dive_act, bks_dive_status, bks_dive_close
  • Stages: bks_stage_list, bks_stage_create, bks_stage_get, bks_stage_edit, bks_stage_check, bks_stage_estimate, bks_stage_run, bks_stage_versions, bks_stage_restore, bks_stage_variant
  • Worlds: bks_world_list, bks_world_get, bks_world_set, bks_world_review, bks_world_rename, bks_world_delete
• Token-only boot: python -m blankstate_mcp with only BLANKSTATE_API_TOKEN — Dive tools work without protocols
• Optional Stage default: Atlas copies BLANKSTATE_STAGE_ID into MCP env. bks_dive_open uses it when you omit stage_id
• Protocol and Metric support: Sense against a single Protocol or a Metric that aggregates multiple Protocols with weights
• SGM 1.0 and 1.5: Full output for both engine versions (resonance, evidence, and extended analysis in v1.5)
• Three modes: block (reject action), feedback (allow + suggest), audit (log only)
• Drop-in integration: Works with any MCP-compatible agent

Quick Start

1. Get an API Token

Sign up at atlas.blankstate.ai and generate an API token. Find your Protocol ID at API Dashboard → Explorer → /protocols.

2. Install the server

Use a virtual environment or pipx so dependencies stay isolated:

pipx install blankstate-mcp
# or: python3 -m pip install blankstate-mcp

3. Add to Your Agent Configuration

Point the MCP host at the same Python you installed into. Two common patterns:

Option A — run as a module (python3 must have blankstate-mcp installed):

{
  "mcpServers": {
    "blankstate": {
      "command": "python3",
      "args": ["-m", "blankstate_mcp"],
      "env": {
        "BLANKSTATE_API_TOKEN": "your-token-here",
        "BLANKSTATE_STAGE_ID": "stg_xxxxxxxx", // optional — default for bks_dive_open
        "BLANKSTATE_PROTOCOLS": "proto-xxxxxxxx-...:1.0" // optional — only for bks_sense / bks_validate
      }
    }
  }
}

Option B — console script (after pipx install blankstate-mcp, the blankstate-mcp executable is on PATH):

{
  "mcpServers": {
    "blankstate": {
      "command": "blankstate-mcp",
      "args": [],
      "env": {
        "BLANKSTATE_API_TOKEN": "your-token-here",
        "BLANKSTATE_STAGE_ID": "stg_xxxxxxxx", // optional — default for bks_dive_open
        "BLANKSTATE_PROTOCOLS": "proto-xxxxxxxx-...:1.0" // optional — only for bks_sense / bks_validate
      }
    }
  }
}

Windows: Use "command": "python" or the full path to python.exe, or the launcher, e.g. "command": "py" with "args": ["-3", "-m", "blankstate_mcp"], depending on how Python is installed. If the host cannot find the interpreter, set command to the absolute path returned by where python (cmd) or Get-Command python (PowerShell).

Important: Use the actual Protocol UUID from Atlas (e.g. proto-6d28a7b8-...:0.1), not a friendly name. Find yours at Atlas → API Dashboard → Explorer → /protocols. Token is required. Protocol and Stage id are optional. Stage can work without protocol UUIDs.

4. Done

Your agent now has the twenty-three tools listed below. Dive works with the token alone. If Atlas filled in BLANKSTATE_STAGE_ID, bks_dive_open can run without passing stage_id. Stage writes need stages:read and stages:write; estimating, starting, pausing, and carrying on a World needs runs:dry.

MCP Tools

Once connected, your agent has these tools:

bks_sense

Measure any content against a Protocol or Metric. Returns a resonance score, per-metamarker breakdown, glass-box evidence, and fidelity assessment.

Input: content (required), protocol_id?, metric_id?, language?

Protocols are universal — they measure conversations, commands, documents, and agent actions the same way. v1.5 protocols provide extended signal analysis output.

When metric_id is provided, all component protocols are evaluated and scores are aggregated by weight into a single composite score.

bks_validate

Validate a proposed agent action (before executing it) against all configured Protocols and Metrics. The action arguments are content-extracted and evaluated.

Input: tool_name (required), tool_args (required)

Returns ALLOWED, BLOCKED (with evidence and matched nuances), or FEEDBACK depending on the configured mode. In block mode a blocked action is returned as an MCP error, stopping the agent from proceeding.

bks_status

Check API connectivity, authentication, remaining ICS credits, and the full list of configured Protocols and Metrics.

Input: (none)

bks_dive_open

Open a Programmatic Dive against a published Stage (POST /api/v1/dive/open). Returns dive_id, rules, the opening observation, and envelope.

Input: stage_id?, session_id?, password?, agent?, seed?

stage_id is optional when BLANKSTATE_STAGE_ID is set in MCP env. A tool argument still wins over the env default.

agent is { id?, version? } — your audit tag, not Blankstate's model.

bks_dive_act

Submit one action. turn is an idempotency key. Returns the next observation and running totals, or a StageCard if the dive auto-closes.

Input: dive_id (required), action (required string or DiveAction), turn?

bks_dive_status

Inspect an open or closed Dive without advancing the world.

Input: dive_id (required)

bks_dive_close

Close the Dive and emit the StageCard (POST /api/v1/dive/{dive_id}/close). Idempotent: closing an already-closed dive returns the same card.

Input: dive_id (required), summary?

Stage id is not baked into the package. Pass it on bks_dive_open, or set BLANKSTATE_STAGE_ID (Atlas MCP snippet) as the default. Use a Stage your token is entitled to dive (stg_… from Atlas).

Stage tools

Stage a situation from your agent. A Stage is the situation you author and version; a World is one run of it. Stages are staged as Proxy Worlds (world: "proxy", dived by agents). Creating and editing are free. Writes are applied whole or refused whole: a refusal names each item and why (issues), and an edit on an old rev is refused with the current rev. The token needs stages:read, stages:write to create or edit, and runs:dry to estimate, start, pause and carry on Worlds.

A World started here is the World Atlas runs and lists. It follows the Stage's run plan and this run's options, and counts in its unit (days, or turns of a conversation, with the interactions so far): it goes on until the Stage's own stop, this run's budget, day or time, your organisation's ICS runs out, or you pause or stop it. A World at the end of its history can be made longer with run.history (that keeps the version) and then carried on. Opening it in Atlas carries it on in the page. What a token may spend is the ICS cap set on its connection in Atlas, and a daily allowance; each token may also start a set number of Worlds a day (refused with daily_cap past it, until midnight UTC).

Tool	Input	Does
bks_stage_list	none	Your Stages: id, title, world, rev, Worlds run. Free
bks_stage_create	world ("proxy"), premise, scale?, ops?	A Stage from a premise, with typed operations in the same call. Applied whole or refused whole. Returns id, rev 1, and the outline. Free
bks_stage_get	stage_id	The outline: actants, links, principles, Moments, objectives and targets, protocols, locks, drawn agents, surfaces and their mix, rev
bks_stage_edit	stage_id, if_rev, ops	Typed operations against the rev you read. An objective can be read on the seeded agents (cases), as a share (rate_*), or on the World's events (reach_*, count_*). run.history makes the history longer and keeps the version. An edit on an old rev is refused with the current rev
bks_stage_check	stage_id	What is thin, each with a question, and the gaps. Free
bks_stage_estimate	stage_id, span, seed?	What span days (or turns) of a World ask on a seed: interactions written, how many are already kept, agents, time (minutes, minutes_max), the ICS range, contrast pairs (pair_passages), and what each objective read on events would count (expect). Free
bks_stage_run	stage_id, wait_seconds?, seed?, budget_ics?, until_day?, minutes?, pauses?, target?	Start a World. seed repeats a draw on the same version. Run options pause it by budget, day, time, or a reached objective. Waits up to wait_seconds, then returns where it is. Spends ICS as it runs
bks_world_set	stage_id, world_id, state (running | paused | stopped), plus the same run options when state is running	Pause, stop, or carry on a World. Carrying on is refused at the end of its history until run.history makes that history longer
bks_world_list	stage_id	The Stage's Worlds, newest first: version, status, the day each is on, who writes it while it runs, and how many objectives it met. Free
bks_world_get	stage_id, world_id, trajectories?	One World's status and outcome: objectives, spread over other seeds, levers, proposals, ICS, and next_steps once paused or stopped. trajectories: true adds each protocol's trajectory evidence. Free
bks_world_review	stage_id, world_id, keep?, discard?	Keep or discard what a World proposes, by id. A discard holds for that World only. Free
bks_world_rename	stage_id, world_id, name	Rename a World (at most 80 characters). Free
bks_world_delete	stage_id, world_id	Delete a World. One still running stops first. Free
bks_stage_versions	stage_id, from?, to?	Kept versions, and what changed between two. Free
bks_stage_restore	stage_id, if_rev, version	Make a kept version the Stage again, against the rev you read
bks_stage_variant	stage_id, world_id, label, varies, question?	A Variant Stage that varies system, population, environment, and/or clock, and holds the rest

Every tool declares a title, whether it is read-only or destructive, and an outputSchema; its answer comes as structuredContent next to the text. The server also sends instructions and the ready-made prompts build_stage, run_world, read_world, and improve_stage for the tools it offers.

Installation

Install from PyPI (recommended: pipx for a global CLI isolated from system Python):

pipx install blankstate-mcp

Other options:

python3 -m pip install blankstate-mcp
# editable from the blankstate-py monorepo:
# python3 -m pip install -e "./packages/sdk" -e "./packages/mcp"

The MCP config must use the same python3 (or the blankstate-mcp script from that environment) that has the package installed.

Configuration

Environment Variables

Variable	Description	Default
BLANKSTATE_API_TOKEN	Required. Bearer token from Atlas
BLANKSTATE_STAGE_ID	Optional. Default Stage for bks_dive_open (Atlas MCP tab)
BLANKSTATE_API_URL	Optional. Do not set this unless you need a custom host (local/dev).	https://api.blankstate.ai
BLANKSTATE_PROTOCOLS	Optional. Comma separated Protocol IDs (sense / validate)
BLANKSTATE_METRICS	Optional. Comma separated Metric IDs
BLANKSTATE_THRESHOLD	Score threshold (0.0 to 1.0)	0.7
BLANKSTATE_MODE	block, feedback, or audit	block

Config File

For advanced configuration, create ~/.blankstate/config.json:

{
  "apiToken": "your-token-here",
  "stageId": "stg_xxxxxxxx",
  "defaultThreshold": 0.7,
  "defaultMode": "block",
  "protocols": [
    {
      "id": "circuit-breaker:1.0",
      "mode": "block",
      "threshold": 0.7,
      "tools": ["exec", "write"]
    },
    {
      "id": "code-quality:1.0",
      "mode": "feedback",
      "threshold": 0.5,
      "tools": ["write"]
    }
  ]
}

Protocol Configuration Options

Option	Type	Description
id	string	Protocol ID with version (e.g., circuit-breaker:1.0)
mode	block	feedback
threshold	number	Score threshold (0.0 to 1.0) for triggering the action
tools	string[]	Tool types this Protocol applies to

Metric Configuration Options

Option	Type	Description
id	string	Metric ID from Atlas
mode	block	feedback
threshold	number	Score threshold (overrides Atlas default if set)
tools	string[]	Tool types this Metric applies to

Tool Types

Type	Description
exec	Shell commands, terminal execution
write	File creation, modification, patches
browser	Web navigation, form submission
messaging	Email, Slack, Discord, etc.
all	All tool types (default)

Modes

Block Mode (Default)

When a Protocol triggers above the threshold, the tool call is rejected:

[BLANKSTATE] Action Blocked

Protocol: circuit-breaker:1.0
Nuances Matched: scope_violation, destructive_pattern
Evidence: targeting home directory with recursive deletion

Message: Action blocked by Protocol circuit-breaker:1.0.
The requested action has been suspended for safety review.

Feedback Mode

Tool call executes, but the agent receives improvement suggestions:

{
  "result": { ... },
  "blankstate_feedback": [
    {
      "source": "code-quality:1.0",
      "score": 0.65,
      "suggestions": ["missing_error_handling", "no_input_validation"]
    }
  ]
}

Use feedback mode to help agents write better code or follow best practices.

Audit Mode

Tool call executes and evaluations are logged for compliance tracking:

{
  "result": { ... },
  "blankstate_audit": [
    {
      "timestamp": "2026-01-31T12:00:00Z",
      "tool": "execute_command",
      "source": "compliance-check:1.0",
      "score": 0.3,
      "action": "allowed"
    }
  ]
}

Available Protocols

Browse available Protocols and their capabilities at doc.blankstate.ai/protocol-versions.

Circuit Breaker (circuit-breaker:1.0)

Designed for autonomous agent safety. Detects scope violations, destructive patterns, credential exposure, and risky communication patterns.

Custom Protocols

Create your own Protocols in Atlas and reference by ID:

{
  "protocols": [
    { "id": "my-org/custom-policy:1.0", "mode": "block" }
  ]
}

Metrics

Metrics aggregate multiple Protocols into a single weighted score. Metrics are defined in Atlas and referenced by ID:

{
  "metrics": [
    {
      "id": "agent-safety-score",
      "mode": "block",
      "threshold": 0.7
    }
  ]
}

The protocol weights and composition are managed in Atlas. The threshold here is optional and overrides the Atlas default if you need different behavior for MCP.

Use Metrics when you want to combine multiple Protocols with weights, or when you need a pre configured safety suite.

Programmatic Usage

You can use the MCP server or the tool wrapper from Python (blankstate_mcp):

import asyncio

from blankstate_mcp import BlankstateMCPServer, load_config, validate_config


async def main() -> None:
    config = load_config()
    validate_config(config)
    server = BlankstateMCPServer(config)
    await server.start()
    await server.wait_until_stopped()


asyncio.run(main())

Or use the tool wrapper directly (validate before executing a tool in your own agent loop):

import asyncio

from blankstate_mcp import create_tool_wrapper, load_config, validate_config


async def main() -> None:
    config = load_config()
    validate_config(config)
    wrapper = create_tool_wrapper(config)

    result = await wrapper.wrap_tool_call(
        "execute_command",
        {"command": "rm -rf ~/"},
    )
    if result.blocked and result.result is not None:
        print("Blocked:", result.result.message)
    else:
        print("Proceed with execution")


asyncio.run(main())

Integration Examples

MCP Compatible Agents

Any agent that supports the Model Context Protocol can use Blankstate. Add to your agent's MCP configuration file:

{
  "mcpServers": {
    "blankstate": {
      "command": "python3",
      "args": ["-m", "blankstate_mcp"],
      "env": {
        "BLANKSTATE_API_TOKEN": "your-token",
        "BLANKSTATE_STAGE_ID": "stg_xxxxxxxx",
        "BLANKSTATE_PROTOCOLS": "circuit-breaker:1.0"
      }
    }
  }
}

The exact config file location depends on your agent. Common locations:

• ~/.config/<agent>/config.json
• ~/<agent>/settings.json

Direct API Integration

For agents without MCP support, you can call the Blankstate API directly. See the API documentation for details.

ICS Billing

Blankstate uses ICS (Interaction Computed Signal) billing:

• ICS consumed only on detection: no charge for tool calls that pass validation
• Usage tracked per API token in the Atlas dashboard
• Credits managed in Atlas: individual accounts receive monthly free credits

This means you can monitor thousands of tool calls and only pay when something is caught. Pricing and credit allocation are configured in your Atlas account.

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                      YOUR MACHINE                                │
│                                                                  │
│  ┌────────────────┐         ┌──────────────────────────────┐    │
│  │   Agent        │  MCP    │   blankstate-mcp             │    │
│  │   (Any MCP     │ ──────> │                              │    │
│  │   compatible   │         │  Intercepts tool calls       │    │
│  │   agent)       │         │  Calls configured Protocols  │    │
│  └────────────────┘         │  Applies mode decisions      │    │
│                             └──────────────┬───────────────┘    │
│                                            │                     │
└────────────────────────────────────────────┼─────────────────────┘
                                             │
                                             │ HTTPS + Bearer Token
                                             ▼
┌──────────────────────────────────────────────────────────────────┐
│                    BLANKSTATE CLOUD                              │
│                                                                  │
│   Protocol evaluation (pure sensors)                             │
│   Returns score, nuances, evidence                               │
│                                                                  │
└──────────────────────────────────────────────────────────────────┘

Contributing

git clone https://github.com/blankstate-ai/blankstate-py.git
cd blankstate-py
python3 -m pip install -e "./packages/sdk" -e "./packages/mcp[dev]"
pytest packages/mcp/tests -q

Troubleshooting

Issue in Cursor / Claude Desktop (Windows)

Some hosts spawn subprocesses without your shell PATH. Use the full path to python.exe (or to the blankstate-mcp script in your pipx/venv Scripts folder) in "command", or configure the host’s environment so python / blankstate-mcp resolves correctly.

"Missing API token"

Set BLANKSTATE_API_TOKEN in the env block of your MCP config, or in ~/.blankstate/config.json.

"stage_id" is required unless BLANKSTATE_STAGE_ID is set

bks_dive_open needs a Stage. Pass stage_id on the tool call, or set BLANKSTATE_STAGE_ID in MCP env (Atlas copies this on the Stage page).

"No protocol or metric configured" (on bks_sense / bks_validate)

The server boots with a token alone. bks_sense and bks_validate still need a protocol or metric at call time. Set BLANKSTATE_PROTOCOLS or BLANKSTATE_METRICS in the env block:

"BLANKSTATE_PROTOCOLS": "proto-xxxxxxxx-...:1.0"

Both can be set at the same time. Protocols are evaluated directly; Metrics aggregate multiple protocols. Dive tools do not need this.

Protocol ID format

Use the UUID-based ID from Atlas (e.g. proto-6d....:0.2), not a friendly name. Find your Protocol ID at Atlas → API Dashboard → Explorer → /protocols.

"API request failed: 401"

Token is invalid or expired. Generate a new one at atlas.blankstate.ai.

"API request failed: 404"

Protocol not found. Verify the ID and version exist in your Atlas account.

Links

• Blankstate
• Atlas Dashboard
• API Reference
• GitHub Issues

License

MIT License. See LICENSE for details.