Atlas / MCP servers / kastalien-research / Thoughtbox

ThoughtboxBLOCK

Name: Thoughtbox
Author: kastalien-research

mcp/kastalien-research/thoughtbox

MCP server by kastalien-research · ai agents · claude · claude code · Audited 2026-10-07

⚑ Report / re-scan Source ↗

Thoughtbox is an intention ledger for agents. Evaluate AI's decisions against its decision-making.

Verdict

BLOCK

Grade

Trust score

30 /100

Exposed tools

164 118r · 42w · 4d

Transport

streamable-http · stdio

License

MIT

Stars

Overview

From the repository's own README, as read at the audited commit. Badges and raw HTML are left out.

Multi-agent collaborative reasoning that's auditable. Thoughtbox is an MCP server for structured, multi-agent reasoning, with a companion web app for workspace and inspection flows. Every step is recorded as a structured thought in a persistent reasoning ledger that can be visualized, exported, and analyzed.

Runtime modes: Local development can use filesystem or in-memory storage. Deployed mode uses Supabase-backed storage, and the current production MCP server runs on Cloud Run.

Observatory UI showing a reasoning session with 14 thoughts and a branch exploration (purple nodes 13-14) forking from thought 5.

Code Mode

Thoughtbox exposes exactly two MCP tools using the Code Mode pattern:

`thoughtbox_search` — Write JavaScript to query the operation/prompt/resource catalog. The LLM has full programmatic filtering power over the catalog.
`thoughtbox_execute` — Write JavaScript using the tb SDK to chain operations. Access thoughts, sessions, knowledge, notebooks, hub, observability, and protocol tools through a unified namespace.

Workflow: search to discover available operations, then execute code against them. Use console.log() for debugging — output is captured in response logs.

This replaces per-operation tool registration with a two-tool surface that scales without context window bloat.

Multi-Agent Collaboration

The Hub is the coordination layer. Agents register with role-specific profiles, join shared workspaces, and work through a structured problem-solving workflow — all via thoughtbox_execute.

The workflow: register → create workspace → create problem → claim → work → propose solution → peer review → merge → consensus

Workspace primitives:

Problem — A unit of work with dependencies, sub-problems, and status tracking (open → in-progress → resolved → closed)
Proposal — A proposed solution with a source branch reference

Read from source at commit 437eeaaba9a8OBSERVED · 2026-10-07

Connect

Built from this server's own package name, version and transport as found in its source — not copied from anyone's documentation, so it cannot drift against a page we do not control. Replace the environment placeholders with a token scoped to the least it needs.

claude-code (npm)

claude mcp add thoughtbox -- npx -y @kastalien-research/[email protected]

Exposed tools (164)

118 read · 42 write · 4 destructive. Blast radius: 4 tools can delete or overwrite — an agent that can be talked into calling a tool can be talked into calling this one.

Tool	Risk	Description
`ARCHITECT`	read	Structural design specialist
`DEBUGGER`	read	Root cause analysis specialist
`M10-test`	read	Testing per-session isolation
`MANAGER`	read	Delegation and team coordination specialist
`RESEARCHER`	read	Parallel hypothesis investigation specialist
`REVIEWER`	read	Code and proposal review specialist
`SECURITY`	read	Risk and vulnerability detection specialist
`add_dependency`	write	Add a dependency between problems. The problem cannot be claimed until its dependency is resolved.
`affected`	read	Transitive dependents of a claim: claims reachable by reverse depends_on edges. Cycle-safe and depth-capped (default 10, max 20); each result carries its edge distance.
`arch-review`	read	Architecture review workspace
`assert`	write	Create a new typed claim in a hub workspace with status asserted. Returns the claim with its generated id.
`bad`	read	...
`bind_final_validator`	read	Pin a notebook code cell as the predicate that gates terminalState=
`blocked_problems`	read	List problems that are blocked by unresolved dependencies.
`branch`	read	Toolhost for managing reasoning branches.
`branch-management`	write	Spawn and merge reasoning branches
`branch-retrieval`	read	List and inspect branches
`branch_get`	read	Retrieve a specific branch with its metadata and all thoughts.
`branch_list`	read	List all branches for a session with thought counts and status.
`branch_merge`	write	Merge branch results back into the main track.
`branch_read`	read	Read thoughts from this branch, ordered by thought number.
`branch_spawn`	write	Create a new reasoning branch from a main-track thought.
`branch_status`	read	Get the status of the current branch (thought count, context).
`branch_thought`	read	Record a thought on this branch. Thoughts are numbered sequentially within the branch.
`caching-decision`	read	Redis selected
`cell-operations`	write	Add, update, list, and retrieve cells
`changed_since`	read	Digest query: claims whose status changed strictly after a timestamp, oldest transition first. For session-start recall and natural-boundary syncs (spec §11.1). Optionally narrowed to one hub workspace. Read-only.
`checkpoint`	write	Submit a diff for Cassandra adversarial audit. The audit result (approved/rejected) and optional feedback are recorded.
`claim_diff`	read	Claim-level diff between two branches in a workspace: added, removed, shared, superseded claims and crossing contradicts edges (SPEC-MERGE-EVIDENCE diffBranchClaims; the same computation feeding merge-evidence notebooks). A claim belongs to a branch when an evidenceRef is
`claim_problem`	read	Claim a problem to work on. Auto-generates a branch name if not provided. Sets status to in-progress.
`clear_folder`	destructive	Whether to clear the .interleaved-thinking folder after completion, keeping only final-answer.md (default: false)
`collab-ws`	read	Collaboration workspace
`collection-ds`	read	Collection
`complete`	read	End the Theseus session with a terminal state. Bridges a knowledge entry if a summary is provided.
`create_problem`	write	Define a new problem to be solved within a workspace.
`create_proposal`	write	Propose a solution to a problem. References a thought branch containing the work.
`create_sub_problem`	write	Create a sub-problem under an existing parent problem. Sub-problems inherit workspace scope.
`create_workspace`	write	Create a new collaboration workspace. The creating agent becomes the coordinator.
`decision`	read	...
`decision-1`	read	...
`decision-2`	read	...
`delete`	destructive	Remove a stored variable. Returns { deleted: boolean } (false if the name was not set).
`deployment-ds`	read	Deployment
`endorse_consensus`	read	Endorse an existing consensus marker to show agreement.
`entity-management`	write	Create, read, and observe entities in the knowledge graph
`evidence-engine`	write	Persist, run, inspect, cancel, and retrieve artifacts for Notebook Evidence Engine modes
`execution`	write	Run code cells and install dependencies
`extract_claims`	read	Extract atomic claims from a text artifact
`get`	write	Read back a value stored earlier in this MCP session via tb.vars.set. Throws a clear error naming the variable if it was never set (or the session restarted) — use tb.vars.list() to see what exists.
`get_profile_prompt`	read	Get the behavioral prompt for a specific profile role. Includes domain-specific mental models and guidelines.
`get_session`	read	Read a reasoning session through the broker proxy
`graph-structure`	write	Create relations and traverse the graph
`health`	read	Check health of Thoughtbox infrastructure services (Thoughtbox server, Supabase OTEL store). Optionally filter to specific services.
`init`	write	Start a Theseus refactoring session with a declared file scope. All subsequent operations are scoped to these files until a visa expands the boundary.
`input`	read	Your task/question. May include
`integration-test`	read	Integration test workspace
`interleaved-thinking`	read	Use this Thoughtbox server as a reasoning workspace to alternate between internal reasoning steps and external tool/action invocation. Enables structured multi-phase execution with tooling inventory, sufficiency assessment, strategy development, and execution.
`invalidate`	read	Mark a claim invalidated. Append-history-style: the claim row is preserved (no hard delete). Idempotent on already-invalidated claims; rejected for superseded claims.
`join_workspace`	read	Join an existing workspace. Returns current workspace state including problems and proposals.
`knowledge_add_observation`	write	Add an observation to an existing entity. Observations are timestamped notes that track how an entity evolves over time.
`knowledge_create_entity`	write	Create a new entity in the knowledge graph. Entities represent insights, concepts, workflows, decisions, or agents.
`knowledge_create_relation`	write	Create a directed relation between two entities. Relations form the edges of the knowledge graph.
`knowledge_get_entity`	read	Retrieve full details of a specific entity by ID, including all properties and metadata.
`knowledge_list_entities`	read	List entities with optional filtering by type, visibility, name pattern, or date range.
`knowledge_query_graph`	read	Traverse the knowledge graph starting from an entity. Follows relations up to a specified depth, with optional filtering by relation type.
`knowledge_stats`	read	Get aggregate statistics for the knowledge graph: entity counts by type, relation counts, and general health metrics.
`link`	write	Create a typed edge between two claims in the same workspace. depends_on edges feed the affected traversal. Idempotent per (from, to, kind).
`list`	read	List the names (and serialised sizes in bytes) of all variables stored in this MCP session.
`list_consensus`	read	List all consensus markers in a workspace with endorsement counts.
`list_mcp_assets`	read	Overview of all MCP capabilities, tools, resources, and quickstart guide
`list_problems`	read	List all problems in a workspace with their status and assignments.
`list_proposals`	read	List all proposals in a workspace with their review status.
`list_workspaces`	read	List all available workspaces. Does not require registration.
`mark_consensus`	read	Record a consensus decision. Links to a thought reference for traceability.
`merge_proposal`	write	Merge an approved proposal. Requires at least one approval review.
`never`	read	never
`notebook-management`	write	Create, list, load, and export notebooks
`notebook_add_cell`	write	Add a cell to a notebook (title, markdown, or executable code)
`notebook_cancel_run`	write	Cancel a running notebook evidence run.
`notebook_create`	write	Create a new headless notebook for literate programming or Notebook Evidence Engine workflows with JavaScript/TypeScript support. See thoughtbox://notebook/capabilities for evidence modes and templates.
`notebook_export`	read	Export a notebook to .src.md format. Always returns the notebook content as a string. Optionally writes to a filesystem path if provided. - STDIO mode: Provide
`notebook_fitness`	read	Read the fitness ledger for a runbook template (SPEC-AGX-SUBSTRATE §7).
`notebook_get_artifact`	read	Retrieve a Notebook Evidence Engine artifact by artifactId.
`notebook_get_cell`	read	Get complete details of a specific cell including content and execution results
`notebook_get_run`	write	Retrieve a run record, including its typed status, outputs, and artifact references.
`notebook_install_deps`	write	Install pnpm dependencies defined in the notebook
`notebook_instantiate`	read	Reconstruct a live notebook from a persisted, versioned runbook template and
`notebook_list`	read	List all active notebooks with their metadata
`notebook_list_cells`	read	List all cells in a notebook with their metadata
`notebook_list_runs`	read	List notebook evidence runs, optionally filtered by notebookId.
`notebook_load`	read	Load a notebook from .src.md format. Accepts exactly one source: -
`notebook_persist`	read	Persist the current notebook document. Always stores an in-process artifact for replay/export; with a durable backend configured (FileSystem locally, Supabase deployed) the document also persists durably — upsert by notebookId, latest wins — and the response
`notebook_run_cell`	write	Execute a code cell and capture output (stdout, stderr, exit code).
`notebook_start_run`	write	Execute a notebook
`notebook_update_cell`	write	Update the content of an existing cell
`notebook_validate`	write	Run a code cell as a deterministic predicate over JSON-serialisable observed data. The cell receives the observed value as a JSON file pointed to by process.env.TB_OBSERVED_PATH and must write its verdict to process.env.TB_VERDICT_PATH as {
`outcome`	read	Record whether tests passed or failed after a modification. Tracks the B (brittleness) counter.
`parallel-verification`	write	Execute parallel hypothesis exploration using Thoughtbox. Pass your task/question - optional max_branches param can be included or will default to 3.
`persistent`	read	...
`plan`	write	Record a primary action step with a pre-committed recovery step. Each move must be paired with a notebook code cell that will deterministically decide that move
`post_message`	write	Post a message to a problem
`post_system_message`	write	Post a system message to a problem
`probe`	read	Issue a single outbound broker proxy call
`query`	read	List claims in a workspace filtered by type, status, creating agent, and/or statement substring.
`quick_join`	read	Register and join a workspace in a single call. Combines register + join_workspace for efficient onboarding.
`read_channel`	read	Read messages from a problem
`ready_problems`	read	List problems that are ready to claim (no unresolved dependencies, status is open).
`reflect`	read	Form a falsifiable hypothesis when the state step tracker reaches S=2 (both primary and backup moves failed). Must include explicit disproof criteria. Required before the protocol allows further action steps. Failed moves are forbidden; generate new primary + backup.
`register`	read	Register as an agent in the hub. Required before any other hub operation. Returns a unique agentId.
`remove_dependency`	destructive	Remove a dependency between problems.
`research`	read	Research workspace
`review`	read	...
`review-ws`	read	Review workspace
`review_proposal`	read	Review a proposal with approve/request-changes/reject verdict.
`runbook_add_await_cell`	write	Author an await cell (SPEC-AGX-SUBSTRATE B6): a claim subscription plus a
`runbook_advance`	read	Pull-based advancement of a durable runbook instance (tb.runbook.advance —
`runbook_status`	read	Read-only snapshot of a durable runbook instance: derived status
`session`	read	Toolhost for managing Thoughtbox reasoning sessions. List, search, retrieve, resume, export, and analyze sessions.
`session-analysis`	read	Analyze session structure and quality metrics
`session-analysis-guide`	read	Process guide for qualitative analysis of reasoning sessions. Teaches how to identify key moments and record learnings to the knowledge graph.
`session-management`	read	Resume and export sessions
`session-retrieval`	read	List, search, and retrieve session details
`session_analyze`	read	Analyze the structure and quality metrics of a reasoning session. Returns objective metrics (linearity, revision rate, branch depth, convergence) - qualitative analysis is done client-side using the session-analysis-guide resource.
`session_export`	read	Export a session to markdown format, optionally using cipher notation for compression. Useful for injecting session context into hooks or sharing reasoning chains.
`session_get`	read	Retrieve full details of a specific reasoning session, including all thoughts, branches, and metadata.
`session_info`	read	Retrieve detailed information about a specific reasoning session including thought count, duration, and metadata.
`session_list`	read	List previous reasoning sessions for the current user. Returns session metadata including title, tags, thought count, and timestamps.
`session_query_thoughts`	read	Structured queries over a session
`session_resume`	read	Load a previous session into the active ThoughtHandler, allowing continuation of reasoning from where it left off. After resuming, subsequent thoughtbox calls will append to this session.
`session_resume_latest`	read	Resume the most recently updated session in the current workspace without knowing its ID. Convenience wrapper around session_list + session_resume; optionally filter by tags. Returns the same payload as session_resume, or { success: false } when the workspace has no sessions.
`session_search`	read	Search for reasoning sessions by title or tags using a keyword query.
`session_timeline`	read	Retrieve the chronological timeline of OTEL events (tool calls, API requests, errors) for a Claude Code session. Events are ordered by timestamp.
`sessions`	read	List reasoning sessions with optional filtering by status (active, idle, or all) and a result limit.
`sleep_forever`	read	Sleeps past every budget to exercise timeout enforcement
`status`	write	Fetch a merge commit by id: status, evidence notebook id + hash, verdict, attribution, and decision timestamps. Read-only.
`subscribe`	read	Register a subscriber (agent or runbook cell ref) on a claim. Defaults to the acting agent. Idempotent per (claim, subscriber).
`supersede`	read	Replace a claim with a new one: asserts the replacement and marks the old claim superseded with a superseded_by pointer. The old claim is preserved.
`support`	read	Attach evidence to a claim and mark it supported. Rejected for invalidated or superseded claims.
`task`	read	The task to complete using interleaved thinking approach
`test`	read	...
`test-ws`	read	...
`thoughtbox-interleaved-guides`	read	IRCoT-style interleaved reasoning guides centered on Thoughtbox as the canonical reasoning workspace. Mode parameter: research, analysis, or development.
`thoughtbox_core_outcomes`	read	Causal-lift core outcomes: end-user tasks where only the final result matters
`thoughtbox_execute`	write	Run JavaScript using the \
`thoughtbox_negative_controls`	read	Causal-lift negative controls: simple, direct tasks where Thoughtbox should stay out
`thoughtbox_notebook`	write	Notebook toolhost for literate programming with JavaScript/TypeScript. Create, manage, and execute interactive notebooks with markdown documentation and executable code cells.
`thoughtbox_search`	read	Discover Thoughtbox operations, prompts, resources, and public tool surfaces by querying this catalog with JavaScript.
`thoughtbox_session`	read	Toolhost for managing Thoughtbox reasoning sessions. List, search, retrieve, resume, export, and analyze sessions.
`thoughtbox_theseus`	read	Theseus Protocol: friction-gated refactoring for autonomous agents. Prevents scope drift via boundary locking, test-write locks, epistemic visas, and adversarial auditing (Cassandra). Operations: - init: Start a refactoring session with declared file scope (args: { scope: [
`thoughtbox_thought`	write	Advanced reasoning tracking tool. Submit thoughts, track state changes, audit decisions, and build branches or revisions.
`thoughts_limit`	read	Maximum number of thoughts to use (default: 100)
`unsubscribe`	destructive	Remove a subscriber from a claim. No-op when the subscription does not exist.
`update_problem`	write	Update problem status or resolution. Status transitions: open → in-progress → resolved → closed.
`v1-decision`	read	Agreed on Redis
`visa`	read	Request an epistemic visa to touch an out-of-scope file. Requires justification and acknowledgment of the scope-creep anti-pattern.
`whoami`	read	Get current agent identity, role, and workspace memberships.
`workspace_digest`	read	Get a comprehensive digest of workspace state: agents, problems, proposals, and consensus.
`workspace_status`	read	Get current status of a workspace including agent activity and problem summary.
`ws`	read	test
`ws-1`	read	...
`ws-2`	read	...
`ws-a`	read	A
`ws-b`	read	B
`x`	read	x

Trust audit

BLOCKgrade F · trust 30/100 Do not install this without reading the findings. The audit found something that could harm you or your machine.

Layer	What it checks	Result
L0	Provenance & inventory	WARN
L1	Static analysis of the code	FAIL
L2	Instruction surface (what it tells the agent)	FAIL
L3	Class-specific surface	WARN
L4	Behavioural (sandbox)	SKIPPED

What the source does

Filesystem: declared (8 observation(s))
Network: declared (10 observation(s))
Shell: declared (5 observation(s))
Dependencies: not all pinned
Secrets in source: found

Findings (25)

HIGHCode injection · code.eval_exec · CWE-78, CWE-94, CWE-95

scripts/utils/validate-handoff.mjs:18

await exec(`git merge-base --is-ancestor ${sha} HEAD`, { cwd: ROOT });

Why it matters. evaluates text as code

Fix. remove; use a parser or a dispatch table

HIGHHard-coded secrets · secret.db_uri · CWE-798, CWE-321

.agents/skills/supabase/SKILL.md:14

- **Local DB URL**: `postgresql://postgres:[email protected]:54322/postgres`

HIGHHard-coded secrets · secret.db_uri · CWE-798, CWE-321

.claude/skills/supabase/SKILL.md:14

- **Local DB URL**: `postgresql://postgres:[email protected]:54322/postgres`

HIGHPrompt injection · prompt.zero_width · CWE-94, CWE-1427

apps/web/.roo/skills/writing-clearly-and-concisely/signs-of-ai-writing.md:165

> Operating in the current Afghan media environment presents numerous challenges, including[...] Despite these challenges, Amu TV has managed to continue to provide a vital service to the Afghan popul

Why it matters. invisible characters in instruction text

Fix. strip non-printing characters

MEDIUMInventory / provenance · inv.binary · CWE-1104

apps/web/src/lib/repo/timeline.ts

timeline.ts

Why it matters. a compiled or binary member cannot be reviewed from source

Fix. ship source, or explain the binary in the README

MEDIUMInventory / provenance · inv.binary · CWE-1104

src/http/drift-http.ts

drift-http.ts

Why it matters. a compiled or binary member cannot be reviewed from source

Fix. ship source, or explain the binary in the README

MEDIUMInventory / provenance · inv.binary · CWE-1104

src/merge-evidence/claim-diff.ts

claim-diff.ts

Why it matters. a compiled or binary member cannot be reviewed from source

Fix. ship source, or explain the binary in the README

MEDIUMInventory / provenance · inv.symlink · CWE-1104

apps/web/.augment/skills/nextjs-supabase-auth

apps/web/.augment/skills/nextjs-supabase-auth