The MCP server that keeps your AI grounded, coherent, and honest — across every turn.
npx context-first-mcpWorks instantly with Claude Desktop · Cursor · VS Code · any MCP client · Vercel remote — zero API keys needed.
37 research-backed tools across 7 layers — context health, state, sandboxing, persistent memory, advanced reasoning, truthfulness verification, orchestration, structured research, and autonomous file export. One
context_loopcall replaces 6–7 individual tools and returns a unified action directive.
Long AI conversations fail in predictable ways. Context-First fixes all four:
Failure Mode | What Goes Wrong | Context-First Solution |
Context Drift | AI forgets earlier decisions and intent as the conversation grows |
|
Silent Contradiction | New inputs silently overrule established facts — the AI doesn't notice |
|
Vague Execution | AI proceeds on underspecified requirements, producing misaligned output |
|
Hallucinated Success | Tool outputs look successful but didn't actually achieve the goal |
|
Related MCP server: nautilus-compass
37 production-ready tools grouped into 7 layers — plus 1 orchestrator that runs them all:
context_loop ─────────────────────────────────────────────────────────────────
├─ Layer 1 · Context Health (9 tools) recap, conflict, ambiguity, depth …
├─ Layer 2 · Sandbox (3 tools) discover_tools, quarantine, merge
├─ Layer 3 · Persistent Memory(6 tools) store, recall, compact, graph …
├─ Layer 4 · Advanced Reasoning(5 tools) InftyThink, Coconut, KAG, MindEvo …
├─ Layer 5 · Truthfulness (7 tools) NCB, IOE, verify_first, self_critique…
└─ State + Research Pipeline + Export (7 tools)One call. One directive. One score.
{
"directive": {
"action": "clarify",
"contextHealth": 0.62,
"instruction": "Resolve with the user: (1) Is this a firm requirement? (2) Which framework?",
"autoExtractedFacts": { "deploy_to": "Vercel" },
"suggestedNextTools": ["verify_execution", "quarantine_context"]
}
}npx context-first-mcp{
"mcpServers": {
"context-first": {
"command": "npx",
"args": ["-y", "context-first-mcp"]
}
}
}{
"mcp": {
"servers": {
"context-first": {
"command": "npx",
"args": ["-y", "context-first-mcp"]
}
}
}
}{
"mcpServers": {
"context-first": {
"url": "https://context-first-mcp.vercel.app/api/mcp"
}
}
}Tool | Purpose |
| One-call orchestrator. Runs 8 stages (ingest→recap→conflict→ambiguity→entropy→abstention→discovery→synthesis) and returns a single |
| Extracts hidden intent, key decisions, and produces consolidated state summaries |
| Compares new input against ground truth; surfaces contradictions |
| Identifies underspecified requirements and generates clarifying questions |
| Validates whether tool outputs actually achieved the stated goal |
| Proxy-entropy scoring via lexical diversity, contradiction density, hedge frequency, and n-gram repetition (ERGO) |
| 5-dimension confidence scoring — abstains with questions rather than hallucinating (RLAAR) |
| Detects conversation drift from the original intent |
| Evaluates response depth against question complexity |
Tool | Purpose |
| Retrieve confirmed facts and task status |
| Lock in ground truth — subsequent conflict checks run against these values |
| Reset specific keys or all state |
| Compressed conversation history with intent annotations |
Tool | Method | Purpose |
| MCP-Zero + ScaleMCP | Natural-language tool routing — returns only semantically relevant tools, reducing context bloat by up to 98% |
| Multi-Agent Quarantine | Create isolated memory silos for sub-tasks, preventing intent dilution |
| Multi-Agent Quarantine | Merge silo results with noise filtering — only promoted keys return to main context |
Tool | Purpose |
| Store findings, decisions, and intermediate results with metadata |
| Retrieve relevant memories by semantic query |
| Compress and consolidate memory entries |
| Build and query a knowledge graph from stored memories |
| Inspect memory store contents and statistics |
| Deduplicate and organize memory entries |
Tool | Method | Purpose |
| InftyThink | Infinite-depth reasoning with adaptive stopping |
| Coconut | Chain-of-Continuous-Thought in latent space |
| ExtraCoT | Compress chain-of-thought while preserving reasoning fidelity |
| MindEvolution | Evolutionary search over the solution space |
| KAG-Thinker | Knowledge-augmented generation with structured thinking |
Tool | Purpose |
| Probe model consistency across paraphrased prompts |
| Detect whether model reasoning is trending toward or away from truth |
| Neighborhood consistency check across semantically equivalent inputs |
| Verify logical coherence of reasoning chains |
| Pre-verification before committing to claims |
| Intrinsic-extrinsic self-correction |
| Structured self-critique with improvement suggestions |
Tool | Purpose |
| Structured research orchestration across |
| Writes every verified report chunk and/or every raw evidence batch to disk in a single call. |
Every core algorithm traces back to a published paper:
Algorithm | Paper | arXiv | Tool |
MCP-Zero | Active Tool Request |
| |
ScaleMCP | Semantic Tool Grouping |
| |
ERGO | Entropy-based Quality |
| |
RLAAR | Calibrated Abstention |
|
Implementation highlights:
Proxy Entropy (ERGO): 4 response-level proxy signals (lexical diversity, contradiction density, hedge-word frequency, n-gram repetition) replace inaccessible token-level logprobs. Composite score above threshold triggers adaptive context reset.
TF-IDF Discovery (MCP-Zero): Pure TypeScript, zero external dependencies. Indexes all tool descriptions at startup; cosine similarity routes queries to the top-k relevant tools only.
Inference-Time Abstention (RLAAR): 5-dimension confidence scoring replaces the RL training loop. Abstains with targeted questions when confidence < threshold — no hallucination fallback.
Tool | Description |
| Writes research artifacts directly to disk. It can automatically expand and write every verified report chunk without asking the LLM to loop |
context_loop (single MCP tool call)
├── Stage 1: INGEST — Store messages to session history
├── Stage 2: RECAP — Extract intents, decisions, summaries
├── Stage 3: CONFLICT — Detect contradictions against ground truth
├── Stage 4: AMBIGUITY — Check for underspecified requirements
├── Stage 5: ENTROPY — Monitor output quality degradation (ERGO)
├── Stage 6: ABSTENTION — Multi-dimensional confidence check (RLAAR)
├── Stage 7: DISCOVERY — Suggest relevant next tools (MCP-Zero)
└── Stage 8: SYNTHESIS — Combine signals → action recommendation + LLM directiveSynthesis Priority: abstain > reset > clarify > proceed
Each stage runs with independent error isolation — a failure in one stage doesn't block the others. The result includes per-stage timing, status, and detailed results for observability.
The context_loop response includes a top-level directive object designed for LLM consumption — a compact, actionable instruction that replaces the need to parse nested stage results:
{
"directive": {
"action": "clarify",
"instruction": "Before proceeding, resolve these issues with the user:\n1. Could you specify exactly what you mean?\n2. Is this a firm requirement or still open for discussion?",
"questions": ["Could you specify exactly what you mean?", "Is this a firm requirement?"],
"contextHealth": 0.62,
"autoExtractedFacts": { "framework": "React", "deploy_to": "Vercel" },
"suggestedNextTools": ["verify_execution", "quarantine_context"]
}
}context_loop Workscontext_loop (single MCP tool call)
├── Stage 1: INGEST — Store messages to session history
├── Stage 2: RECAP — Extract intents, decisions, summaries
├── Stage 3: CONFLICT — Detect contradictions against ground truth
├── Stage 4: AMBIGUITY — Check for underspecified requirements
├── Stage 5: ENTROPY — Monitor output quality degradation (ERGO)
├── Stage 6: ABSTENTION — Multi-dimensional confidence check (RLAAR)
├── Stage 7: DISCOVERY — Suggest relevant next tools (MCP-Zero)
└── Stage 8: SYNTHESIS — Combine signals → action + directiveSynthesis priority: abstain > reset > clarify > proceed
Each stage runs with independent error isolation. The directive response field carries everything an LLM needs:
Field | Description |
|
|
| Plain-language guidance for the LLM's next step |
| Aggregated clarifying questions (ambiguity + abstention + conflicts) |
| 0–1 composite score. 1 = healthy, 0 = degraded |
| Key-value facts auto-extracted from user messages and stored as ground truth |
| Relevant tools the LLM should consider next |
Smart defaults: currentInput is auto-inferred from the last user message. Facts like "use React" are extracted and stored automatically.
The #1 mistake: LLMs treat
context_loopas optional. It's not — it's the backbone.
The server ships with four compliance mechanisms that require zero configuration:
Server Instructions — Full usage protocol injected at MCP handshake via ServerOptions.instructions
Bootstrap Gate — First non-context_loop call appends a strong redirect reminder
Cross-Tool Reminders — After 3 consecutive calls without context_loop, reminders appear in tool responses
MCP Prompts — context-first-protocol and research-protocol prompt templates available on demand
When using Context-First MCP:
1. Call context_loop BEFORE any complex task
2. Call context_loop every 2–3 tool calls
3. Call context_loop AFTER generating long-form output
4. ALWAYS follow directive.action (proceed/clarify/reset/abstain/deepen/verify)
5. Use memory_store to save findings; memory_recall to retrieve themresearch_pipeline orchestrates memory, phase control, reasoning, and autonomous file writing. It is not a web crawler — bring your own sources from web search, GitHub, fetch tools, PDFs, or any other MCP.
Phase 1 · Init research_pipeline(init) → sets up state, enables autonomous file writing
Phase 2 · Gather ONE web search → research_pipeline(gather) → file written to disk → repeat
Phase 3 · Analyze research_pipeline(analyze) → reasoning engines produce clean analysis file
Phase 4 · Verify research_pipeline(verify) → context health gate (non-blocking)
Phase 5 · Finalize research_pipeline(finalize) → synthesis.md + all batch files on disk
Automation shortcut:
export_research_files(outputDir, exportVerifiedReport=true) → write all report chunks
export_research_files(outputDir, exportRawEvidence=true) → write all evidence batchesAutonomous file writing is always on. Files are written to ./context-first-research-output/ by default — no LLM cooperation required. Pass outputDir to override.
┌──────────────────────────────────────────────────────────────┐
│ @xjtlumedia/context-first-mcp-server │
│ (Core — shared logic) │
│ │
│ Layer 1: Context Health (9 tools) │
│ Layer 2: Sandbox (3 tools) │
│ Layer 3: Persistent Memory (6 tools) │
│ Layer 4: Advanced Reasoning(5 tools) │
│ Layer 5: Truthfulness (7 tools) │
│ State (4) · Orchestrator · Pipeline · Export │
└──────────────┬───────────────────────┬──────────────────────┘
│ │
┌──────▼──────┐ ┌──────▼────────┐
│ stdio-server │ │ remote-server │
│ (npx local) │ │ (Vercel) │
│ stdio │ │ Streamable │
│ 37 tools │ │ HTTP │
└──────────────┘ │ 37 tools │
└───────────────┘Core library (@xjtlumedia/context-first-mcp-server): All tool implementations. Zero external API keys — heuristic-based by default.
stdio-server (context-first-mcp): npx entry point, stdio transport, 37 tools.
remote-server: Vercel serverless, Streamable HTTP transport, 37 tools.
Try all 37 tools live in your browser at context-first-mcp.vercel.app.
git clone https://github.com/XJTLUmedia/Context-First-MCP.git
cd Context-First-MCP
pnpm install
# Build everything
pnpm build
# Run stdio server
cd packages/stdio-server && pnpm start
# Run frontend
cd packages/frontend && pnpm dev
# Tests
pnpm testSee CONTRIBUTING.md.
Context-First MCP · @xjtlumedia/context-first-mcp-server · context-first-mcp
Built for every developer tired of watching their AI lose the plot.
[CONTEXT & STATE] 13 sub-tools: recap, conflict, ambiguity, verify, entropy, abstention, grounding, drift, depth, get_state, set_state, clear_state, history. Auto-selects based on params or use 'check' to override. TIP: context_loop runs all health checks automatically — prefer context_loop for comprehensive analysis, use context_health for targeted checks.
| Name | Required | Description | Default |
|---|---|---|---|
| sessionId | No | Session identifier | default |
| check | No | Override: run a specific check. If omitted, auto-selects based on params. clear_state shares params with get_state — use this override to disambiguate. | |
| params | No | Parameters for the underlying tool(s), minus sessionId. Multiple checks run if params match more than one tool. recap: {messages[], lookbackTurns?}; conflict: {newMessage}; ambiguity: {requirement, context?}; verify: {goal, output, expectedIndicators?}; entropy: {outputs[], threshold?, autoReset?}; abstention: {claim, requiredKeys[], threshold?}; grounding: {assistantOutput, claim?}; drift: {windowSize, turn?, health?, breakdown?}; depth: {content, minDepthWords?, minDepthSentences?}; get_state: {keys?}; set_state: {key, value, source?}; clear_state: (use check override); history: {maxTokens} |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
No annotations are provided, so the description carries full burden. It mentions 'auto-selects based on params' and the override capability, but doesn't disclose critical behavioral traits like what happens when multiple checks run, error handling, performance characteristics, or what 'health' assessment entails.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Reasonably concise with two sentences plus a TIP. The first sentence is dense but informative. Could be slightly more front-loaded with purpose before implementation details.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given 3 parameters, 100% schema coverage, no output schema, and no annotations, the description provides good usage guidance but lacks behavioral context for a complex tool with 13 sub-tools. The absence of output schema means the description should ideally explain what results to expect.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, so baseline is 3. The description adds meaningful context about 'clear_state shares params with get_state — use this override to disambiguate' and mentions 'Multiple checks run if params match more than one tool,' which provides valuable semantic information beyond the schema.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description states the tool has '13 sub-tools' for health checks and can auto-select or override, but it's vague about what 'health' means in this context. It distinguishes from sibling 'context_loop' but doesn't clearly articulate the core purpose beyond being a collection of checks.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Explicitly provides when-to-use guidance: 'prefer context_loop for comprehensive analysis, use context_health for targeted checks.' This clearly distinguishes from the sibling tool and gives specific usage context.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
[ORCHESTRATOR — CALL THIS FIRST] CALL THIS TOOL every 2-3 turns and at the start of ANY task. It is the single most important tool — it replaces calling recap, conflict, ambiguity, entropy, grounding, drift, depth, and discovery tools individually.
What it does: Runs ALL context health checks in one call. Auto-extracts facts from conversation, detects contradictions, checks answer quality, and tells you exactly what to do next.
Returns a 'directive' object with:
action: proceed | clarify | reset | abstain | deepen | verify
instruction: Plain English telling you what to do
contextHealth: 0-1 score
autoExtractedFacts: Key facts pulled from conversation
suggestedNextTools: What tools to call next
constraints: Machine-readable rules you must follow
ESSENTIAL for: research tasks, multi-step workflows, long conversations, preserving context across turns, knowledge management, and any task requiring memory or fact-checking.
Minimal call: { "messages": [{"role":"user","content":"","turn":1}] } — most fields have smart defaults.
| Name | Required | Description | Default |
|---|---|---|---|
| sessionId | No | default | |
| messages | No | Recent conversation messages. Include at least the last 2-3 user/assistant exchanges. Example: [{role:'user', content:'explain X', turn:1}, {role:'assistant', content:'X is...', turn:2}]. If empty, the loop runs with reduced context. | |
| currentInput | No | The current user message or task description. Auto-inferred from last user message in messages array if omitted. | |
| claim | No | A specific assertion or answer to fact-check for confidence evaluation | |
| discoveryQuery | No | What capability do you need? e.g. 'store research findings' or 'compress reasoning chain' | |
| lookbackTurns | No | How many turns to analyze (use 15-20 for research or long conversations) | |
| entropyThreshold | No | Entropy spike detection threshold (0-1) | |
| abstentionThreshold | No | Abstention confidence threshold (0-1) |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries the full burden of behavioral disclosure. It effectively describes the tool's behavior: it 'runs ALL context health checks,' 'auto-extracts facts,' 'detects contradictions,' 'checks answer quality,' and 'tells you exactly what to do next.' It also details the return structure ('directive' object with specific fields) and provides a minimal call example. However, it lacks information on potential side effects, error handling, or performance characteristics like rate limits.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is appropriately front-loaded with critical usage instructions, but it is verbose with some redundancy (e.g., repeating the tool's importance). Sentences like 'It is the single most important tool' and 'ESSENTIAL for:' could be more concise. While most content is valuable, the structure could be tighter to improve readability without losing essential information.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the tool's complexity (8 parameters, no annotations, no output schema), the description is fairly complete. It explains the tool's purpose, usage, behavior, and return structure in detail. However, it lacks an output schema, so the description must fully describe return values, which it does with the 'directive' object fields. Gaps include no error handling details and limited parameter semantics, but overall, it provides sufficient context for effective use.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The schema description coverage is high (88%), so the baseline is 3. The description adds minimal parameter semantics beyond the schema: it mentions 'most fields have smart defaults' and provides a minimal call example for 'messages.' However, it does not explain the purpose or interaction of parameters like 'sessionId,' 'claim,' or 'discoveryQuery,' nor does it clarify how parameters like 'currentInput' are 'auto-inferred.' The description compensates somewhat but not significantly.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool's purpose with specific verbs ('runs ALL context health checks in one call') and resources ('auto-extracts facts from conversation, detects contradictions, checks answer quality'). It explicitly distinguishes this tool from its siblings by stating it 'replaces calling recap, conflict, ambiguity, entropy, grounding, drift, depth, and discovery tools individually,' making the differentiation unambiguous.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides explicit usage guidelines: 'CALL THIS TOOL every 2-3 turns and at the start of ANY task' and 'ESSENTIAL for: research tasks, multi-step workflows, long conversations, preserving context across turns, knowledge management, and any task requiring memory or fact-checking.' It also implicitly suggests when not to use it (for simpler tasks not requiring these features) and positions it as a replacement for multiple sibling tools, offering clear alternatives.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
[EXPORT] Automatically writes research artifacts to disk. It can expand and write every verified report chunk without asking the LLM to loop finalize manually, and it can also write every gathered raw-evidence batch even when verify has not passed yet.
| Name | Required | Description | Default |
|---|---|---|---|
| sessionId | No | default | |
| outputDir | Yes | Directory where research export files will be written. Prefer an absolute path so the caller knows exactly where the artifacts landed. | |
| baseFileName | No | Base filename prefix for all written research artifacts. The helper sanitizes it into a filesystem-safe ASCII stem. | research_export |
| exportVerifiedReport | No | When true, automatically expands and writes the full verified report by looping all finalize chunks internally. This path remains blocked until verify has passed. | |
| exportRawEvidence | No | When true, writes every gathered research batch as raw evidence files even if verify has not passed, separating evidence capture from narrative approval. | |
| maxChunkChars | No | Maximum size for each written markdown file. Large batches are automatically split across multiple files when needed. | |
| overwrite | No | Whether existing export files may be overwritten. Defaults to false so exports do not silently clobber prior artifacts. | |
| finalSummary | No | Optional final summary override for verified report export. If omitted, the helper uses the stored pipeline summary or existing analysis and verification outputs. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries the full burden of behavioral disclosure. It does an excellent job describing key behaviors: automatic file writing, internal chunk processing, blocking behavior until verification passes, separation of evidence capture from narrative approval, and file splitting for large batches. The only gap is lack of information about error handling or permission requirements.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is efficiently structured in two sentences that each earn their place. The first establishes the core export functionality, while the second elaborates on the two distinct export modes. It's appropriately sized for an 8-parameter tool with complex behavior, though it could be slightly more front-loaded with the most critical information.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a complex export tool with 8 parameters, no annotations, and no output schema, the description provides substantial context about what the tool does and how it behaves. It covers the two main export modes, automation aspects, and file handling. The main gap is lack of information about return values or error conditions, which would be helpful given the absence of an output schema.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
With 88% schema description coverage, the baseline is 3. The description adds meaningful context about parameter behavior: it explains that the tool 'can expand and write every verified report chunk' (relates to exportVerifiedReport), 'can also write every gathered raw-evidence batch even when verify has not passed yet' (relates to exportRawEvidence), and implies automation that affects multiple parameters. This provides valuable semantic understanding beyond the schema's technical descriptions.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool's purpose with specific verbs ('writes research artifacts to disk', 'expand and write every verified report chunk', 'write every gathered raw-evidence batch') and distinguishes it from sibling tools by focusing on export functionality. It explicitly mentions automation capabilities that differentiate it from manual processes.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides clear context about when to use the tool ('automatically writes research artifacts', 'without asking the LLM to loop finalize manually') and distinguishes between two export modes (verified reports vs raw evidence). However, it doesn't explicitly mention when NOT to use this tool or name specific alternatives among the sibling tools.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
[MEMORY] 6 sub-tools: store (hierarchical ingest), recall (adaptive gate retrieval), compact (compress with integrity), graph (knowledge graph with PageRank), inspect (tier status), curate (importance-based curation). Auto-selects based on params or use 'action' to override. TOOL NAME: memory (use underscores).
| Name | Required | Description | Default |
|---|---|---|---|
| sessionId | No | Session identifier | default |
| action | No | Override: run a specific memory action. If omitted, auto-selects based on params. store — pass {role, content}; recall — pass {query}; compact — pass {targetRatio?}; graph — pass {action:'query'|'stats'|'recompute'}; inspect — pass {tier?}; curate — pass {action:'top'|'filterByDomain'|'mostReused'|'prune'} | |
| params | No | Parameters for the underlying tool. store: {role, content, metadata?}; recall: {query, maxResults?, turnCount?, entropy?, conflicts?}; compact: {targetRatio?, preserveRecency?}; graph: {action:'query'|'stats'|'recompute', startEntity?, depth?}; inspect: {tier?, runIntegrityCheck?}; curate: {action:'top'|'filterByDomain'|'mostReused'|'prune', domainTag?, threshold?} |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries the full burden of behavioral disclosure. It mentions 'auto-selects based on params' which describes decision logic, and gives brief behavioral hints for each sub-tool (e.g., 'compress with integrity', 'knowledge graph with PageRank'). However, it doesn't disclose important behavioral traits like whether operations are read-only or destructive, performance characteristics, error handling, or authentication requirements.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is poorly structured and contains unnecessary elements. It starts with '[MEMORY]' which adds no value, includes implementation details like 'use underscores' that don't help the agent, and has a confusing mix of tool documentation and usage instructions. The information about parameter mappings could be presented more clearly. Multiple sentences don't earn their place.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the complexity (6 sub-tools with different behaviors), no annotations, and no output schema, the description is incomplete. While it covers the basic action-parameter mappings, it doesn't explain what the tool returns, error conditions, or the semantics of operations like 'compact' or 'curate'. For a complex multi-function tool with no structured metadata, more comprehensive documentation would be expected.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The schema description coverage is 100%, so the baseline is 3. The description adds significant value by explaining the semantic mapping between 'action' values and required 'params' structures. For example, it specifies that 'store' requires {role, content}, 'recall' requires {query}, etc. This goes beyond what the schema provides by clarifying how parameters interact with actions.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description lists the 6 sub-tools (store, recall, compact, graph, inspect, curate) which gives a vague sense of purpose, but it doesn't clearly state what the overall 'memory' tool does. It mentions 'hierarchical ingest', 'adaptive gate retrieval', etc., but these are technical terms that don't clearly explain the tool's function. The description focuses on implementation details rather than stating the core purpose.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides clear guidance on when to use specific actions: 'Auto-selects based on params or use 'action' to override.' It explains the default behavior (auto-selection) and how to override it. However, it doesn't provide guidance on when to use this tool versus its siblings (context_health, context_loop, etc.), which would be needed for a perfect score.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
[REASONING] 5 engines: inftythink (iterative bounded reasoning), coconut (multi-perspective latent analysis), extracot (reasoning chain compression), mindevolution (evolutionary search), kagthinker (structured logical decomposition with dependency DAG). Auto-selects based on params or use 'method' to override.
| Name | Required | Description | Default |
|---|---|---|---|
| sessionId | No | Session identifier | default |
| method | No | Override: run a specific reasoning method. If omitted, auto-selects based on params. inftythink — iterative bounded reasoning (default for raw problems); coconut — multi-perspective latent-space analysis; extracot — compress existing reasoning steps; mindevolution — evolutionary search over seed solutions; kagthinker — structured logical decomposition with dependency graph | |
| params | No | Parameters for the underlying reasoning engine. inftythink: {problem, priorContext?, maxSegments?, maxSegmentTokens?, summaryRatio?}; coconut: {problem, maxSteps?, breadth?, enableBreadthExploration?}; extracot: {reasoningSteps[], problem?, maxBudget?, targetCompression?, minFidelity?}; mindevolution: {problem, criteria?, populationSize?, maxGenerations?, seedResponses[]}; kagthinker: {problem, knownFacts?, maxDepth?, maxSteps?} |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries full burden. It discloses the existence of 5 reasoning engines and auto-selection behavior, but doesn't describe performance characteristics, rate limits, authentication needs, or what constitutes successful/unsuccessful execution. The description adds some behavioral context but leaves significant gaps.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is appropriately sized with two sentences that efficiently convey the core functionality. The first sentence lists all engines, and the second explains the selection mechanism. No redundant information is present, though the engine names could be better integrated with their descriptions.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a complex tool with 3 parameters (including a nested object), no annotations, and no output schema, the description is incomplete. It doesn't explain what the tool returns, error conditions, or provide examples of typical use cases. The parameter descriptions in the schema help, but the description alone leaves significant gaps for agent understanding.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, so the baseline is 3. The description adds value by explaining the auto-selection logic ('Auto-selects based on params') and providing high-level descriptions of each method option, which complements the schema's technical enum values. However, it doesn't elaborate on how params influence auto-selection beyond what's implied.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool provides reasoning with 5 different engines and auto-selection capability. It specifies the verb 'reasoning' and resource 'engines', but doesn't distinguish this from sibling tools like 'context_loop' or 'truthcheck' which might also involve reasoning processes.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implies usage through the mention of auto-selection based on params and method override, but doesn't explicitly state when to use this tool versus alternatives like 'context_loop' or 'truthcheck'. No specific exclusions or comparison to sibling tools is provided.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
[PIPELINE] RECOMMENDED for research tasks. Orchestrates all underlying Context-First layers through 6 phases (init→gather→review→analyze→verify→finalize). NEW: plan→draft→review→fix loop — like compile→test→fix in coding. Init generates a research outline (12+ sections). Each gather adds depth to one section with quality gate (25K char / 500 line min — multiple gathers per section expected). Review runs quality tests and identifies gaps. CRITICAL: Interleave web search and gather — after EACH search, IMMEDIATELY call gather with deeply written content. Do NOT batch searches. Each gather writes a file to disk. After sufficient gathers, call review to run quality tests. Fix failed sections by gathering again with metadata.targetSection=N. Coverage must reach 60% before analyze. Autonomous file writing is ALWAYS ON — files are written to disk during gather, analyze, and finalize phases. Provide outputDir to control destination, or let the pipeline auto-create a temp directory. Finalize works even if verify hasn't passed. It does not browse the web or invent source material for you; use it to structure, preserve, pressure-test, and export sourced findings collected from web, GitHub, fetch, or other MCP tools.
| Name | Required | Description | Default |
|---|---|---|---|
| sessionId | No | default | |
| phase | Yes | Pipeline phase. Call in order: init → gather (repeatable) → review → analyze → verify → finalize. Each phase auto-chains the appropriate layer tools internally. Analyze is blocked until gathered evidence clears the weak-evidence gate and coverage ≥ 60%. | |
| content | Yes | Phase-specific content: init=task description, gather=WRITE a deeply researched section based on your LATEST web search. CRITICAL WORKFLOW: Do ONE web search, then IMMEDIATELY call gather. Repeat. Do NOT do multiple searches before calling gather — content gets lost to compaction. You are a research AUTHOR: use the search result as input to write a comprehensive section with specific facts, data, analysis, relationships, and expert commentary. Each gather call writes one file to disk immediately — this is the pipeline's core output. analyze=problem/question to reason about (runs on accumulated gather files), verify=draft output to verify (runs on accumulated files), finalize=final summary to persist (synthesizes all files) | |
| outputDir | No | RECOMMENDED. When provided, the pipeline autonomously writes enriched research files to this directory during gather, analyze, and finalize phases. This eliminates the need for the LLM to write files manually — the pipeline writes them itself, like how export_research_files works but incrementally per-phase. Files survive context compaction because they are on disk, not just in memory. Prefer an absolute path. | |
| baseFileName | No | Base filename prefix for autonomously written files when outputDir is set. Produces files like research.batch-001.topic-slug.md, research.analysis-001.md, research.synthesis.md | research |
| messages | No | Recent conversation messages for context_loop | |
| claim | No | Specific claim to fact-check (verify phase) | |
| metadata | No | Optional metadata for memory storage and finalize/export controls. Supported conventions: sourceTools during gather, maxChunkChars during finalize, exportChunkIndex during finalize chunk retrieval, outline (Array<{title, description}>) during gather to set the research outline, targetSection (number) during gather to expand/append depth to a specific outline section (multi-gather accumulation). |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description fully discloses behavioral traits: autonomous file writing to disk during phases, quality gates (25K char/500 line min, 60% coverage requirement), phase dependencies (analyze blocked until coverage threshold), and critical workflow constraints (immediate gather after search). It explains operational mechanics like file persistence and phase chaining.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is front-loaded with purpose and workflow but becomes verbose with repetitive instructions (e.g., multiple warnings about immediate gather). Some sentences could be condensed (e.g., overlapping explanations of file writing). It's informative but not optimally concise, with minor redundancy.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the tool's complexity (8 parameters, no annotations, no output schema), the description is highly complete: it explains the multi-phase process, behavioral constraints, file management, and integration with other tools. It compensates for lack of structured fields by detailing usage, dependencies, and outputs sufficiently for an agent to operate it correctly.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is high (88%), so baseline is 3. The description adds value by clarifying parameter usage in context: e.g., content's role per phase (init=task description, gather=write based on latest search), outputDir's purpose for autonomous file writing, and metadata conventions like targetSection for multi-gather accumulation. However, it doesn't fully detail all 8 parameters beyond schema hints.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool orchestrates research tasks through 6 phases (init→gather→review→analyze→verify→finalize), specifying it structures, preserves, pressure-tests, and exports sourced findings. It distinguishes from siblings by focusing on research orchestration rather than isolated functions like export_research_files or memory.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Explicit guidance is provided: 'RECOMMENDED for research tasks,' with detailed workflow instructions (e.g., interleave web search and gather, do not batch searches, call phases in order). It contrasts with alternatives by noting it does not browse the web or invent source material, implying use of other MCP tools for sourcing.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
[SANDBOX] 3 sub-tools: discover (semantic tool search via TF-IDF), quarantine (isolated state sandbox), merge (merge/discard silo). Auto-selects based on params or use 'action' to override.
| Name | Required | Description | Default |
|---|---|---|---|
| sessionId | No | Session identifier (used by quarantine/merge) | default |
| action | No | Override: run a specific action. If omitted, auto-selects based on params. discover — pass {query}; quarantine — pass {name}; merge — pass {siloId, action:'merge'|'discard'} | |
| params | No | Parameters for the underlying tool. discover: {query, maxResults?, minScore?}; quarantine: {name, inheritKeys?, ttl?}; merge: {siloId, action:'merge'|'discard', promoteKeys?} |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries full burden and does well by disclosing key behavioral traits: the tool has three distinct sub-tools with different purposes, auto-selects based on parameters unless overridden, and describes what each sub-tool does (search, isolation, merge/discard). It doesn't mention performance characteristics like rate limits or error handling, but covers the core behavior adequately.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is perfectly concise and well-structured: one sentence identifies the three sub-tools with their purposes, and a second sentence explains the auto-selection and override mechanism. Every word earns its place with no redundancy, and key information is front-loaded.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the tool's complexity (three sub-tools with different behaviors), no annotations, and no output schema, the description does an excellent job explaining what the tool does and how to use it. The main gap is lack of information about return values or error conditions, which would be helpful given the absence of output schema. However, for a sandbox tool with clear parameter guidance, it's mostly complete.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, so the baseline is 3. The description adds significant value by explaining the semantic relationship between parameters: how 'action' overrides auto-selection, and how 'params' should be structured differently for each sub-tool (query for discover, name for quarantine, siloId+action for merge). This clarifies usage beyond the schema's technical definitions.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool's purpose: it's a sandbox with three specific sub-tools (discover, quarantine, merge) and explains their functions (semantic tool search, isolated state sandbox, merge/discard silo). It distinguishes from siblings by specifying its unique multi-action nature and auto-selection behavior, which none of the listed sibling tools appear to share.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides explicit guidance on when to use each sub-tool: 'discover' for semantic search with query parameters, 'quarantine' for isolation with name parameter, and 'merge' for merging/discarding with siloId and action. It also explains the auto-selection logic and how to override it with the 'action' parameter, giving clear alternatives within the tool itself.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
[TRUTHFULNESS] 7 tools: probe (linguistic truth signals), truth_direction (truth vector projection), ncb (perturbation robustness), logic (formal logical consistency), verify_first (5-dimension verification), ioe (confidence-based correction), self_critique (iterative refinement). Auto-selects or use 'check' to override. Set cascade=true for auto-correction on low scores.
| Name | Required | Description | Default |
|---|---|---|---|
| sessionId | No | Session identifier | default |
| check | No | Override: run a specific truthfulness check. If omitted, auto-selects based on params. probe — linguistic truth proxy signals; truth_direction — truth vector projection; ncb — perturbation robustness; logic — formal logical consistency; verify_first — 5-dimension verification; ioe — confidence-based self-correction; self_critique — iterative multi-criteria refinement | |
| cascade | No | If true, after primary checks, auto-run ioe_self_correct → self_critique when any extracted truthfulness score falls below 0.5. | |
| params | No | Parameters for the underlying tool(s), minus sessionId. probe: {assistantOutput, includeHistory?}; truth_direction: {assistantOutput, includePriorOutputs?}; ncb: {originalQuery, response}; logic: {claims[], includeGroundTruth?}; verify_first: {candidateAnswer, question, context?}; ioe: {response, question?, priorAttempts?}; self_critique: {solution, criteria?, maxIterations?, question?} |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries full burden and does well: it discloses the tool's multi-method approach, auto-selection behavior, cascade functionality for correction, and scoring threshold (below 0.5 triggers correction). It explains the tool's operational logic beyond basic input-output, though it could mention performance characteristics or error handling more explicitly.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is appropriately sized for a complex tool with 4 parameters and 7 methods. It front-loads the 7 tools list, then explains auto-selection and cascade behavior. While dense, every sentence adds value; it could be slightly more structured but remains efficient without wasted words.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the tool's complexity (4 parameters, nested objects, 7 methods) and no annotations/output schema, the description does well: it covers purpose, usage modes, parameter effects, and behavioral logic. It explains the multi-tool approach and cascade correction, though it doesn't detail return formats or error cases, which would be helpful given the absence of output schema.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, so baseline is 3. The description adds meaningful context: it explains that 'check' overrides auto-selection and lists what each enum value represents (e.g., 'probe — linguistic truth proxy signals'), providing semantic clarification beyond the schema's technical descriptions. It also explains the cascade parameter's effect, adding operational understanding.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description explicitly states the tool's purpose: it performs truthfulness checking using 7 specific methods (probe, truth_direction, ncb, logic, verify_first, ioe, self_critique). It distinguishes itself from siblings by focusing on truth verification rather than context management, reasoning, or file operations. The description provides a clear verb ('truthfulness checking') and resource ('7 tools') with specific differentiation.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides explicit guidance on when to use this tool: it mentions auto-selection of methods or manual override with 'check' parameter, and specifies cascade=true for auto-correction on low scores. It distinguishes usage scenarios between automatic and manual modes, though it doesn't explicitly mention when NOT to use it or alternatives among siblings, but the context is sufficiently clear for a complex tool.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Recent tool additions, removals, and schema changes observed during successful MCP inspections. Dates show when Glama detected each change.
v0.1.2context_healthcontext_loopexport_research_filesmemoryreasonresearch_pipelinesandboxtruthcheckv0.1.1abstention_checkcheck_ambiguitycheck_depthcheck_groundingcheck_logical_consistencyclear_statecoconut_reasoncontext_loopdetect_conflictsdetect_driftdetect_truth_directiondiscover_toolsentropy_monitorexport_research_filesextracot_compressget_history_summaryget_stateinftythink_reasonioe_self_correctkagthinker_solvememory_compactmemory_curatememory_graphmemory_inspectmemory_recallmemory_storemerge_quarantinemindevolution_solvencb_checkprobe_internal_statequarantine_contextrecap_conversationresearch_pipelineself_critiqueset_stateverify_executionverify_firstv0.1.0abstention_checkcheck_ambiguitycheck_depthcheck_groundingcheck_logical_consistencyclear_statecoconut_reasoncontext_loopdetect_conflictsdetect_driftdetect_truth_directiondiscover_toolsentropy_monitorexport_research_filesextracot_compressget_history_summaryget_stateinftythink_reasonioe_self_correctkagthinker_solvememory_compactmemory_curatememory_graphmemory_inspectmemory_recallmemory_storemerge_quarantinemindevolution_solvencb_checkprobe_internal_statequarantine_contextrecap_conversationresearch_pipelineself_critiqueset_stateverify_executionverify_firstThe tools have distinct high-level purposes (e.g., context management, memory, reasoning, research), but the sub-tools within each main tool (like context_health's 13 sub-tools or memory's 6 sub-tools) create significant internal overlap and ambiguity. For example, context_health and context_loop both handle context checks, with context_loop described as replacing many individual checks, which could confuse an agent about when to use each. The auto-selection features mitigate this somewhat, but the boundaries between tools like context_health, context_loop, and truthcheck are not clearly defined, leading to potential misselection.
The naming is inconsistent across tools, with a mix of styles: some use snake_case (context_health, context_loop, export_research_files), others use single words (memory, reason, sandbox, truthcheck), and research_pipeline uses a hybrid format. There is no predictable verb_noun pattern, and the sub-tools within each main tool further add to the inconsistency (e.g., inftythink vs. extracot in reason). While the names are readable, the lack of a uniform convention makes the set harder to navigate and predict.
With 8 main tools, the count is reasonable for a server focused on context management and research workflows, as it covers key areas like health checks, memory, reasoning, and pipeline orchestration. However, the extensive sub-tools (e.g., 13 in context_health) make the effective surface larger, which could feel heavy but is justified by the server's complex domain. The count is slightly high but still appropriate given the scope, avoiding extreme over- or under-provisioning.
The tool set provides comprehensive coverage for context-aware AI tasks, including context health monitoring (context_health, context_loop), memory storage and retrieval (memory), reasoning engines (reason), research pipeline management (research_pipeline), truth verification (truthcheck), sandboxing (sandbox), and export functionality (export_research_files). There are no obvious gaps; it supports full lifecycle operations from initialization to analysis and export, with tools like context_loop and research_pipeline ensuring no dead ends in workflows. The domain is well-covered with tools that interlock effectively.
Unclaimed servers have limited discoverability.
If you are the server author, to access and configure the admin panel.
Persistent memory and drift detection for AI agents across session restarts.
Context engineering for AI coding agents: product context, project missions, and 360 memory.
Memory that reasons: continual learning for stateful agents. Better context, fewer tokens.
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/XJTLUmedia/Context-First-MCP'
If you have feedback or need assistance with the MCP directory API, please join our Discord server
