GTM MCP Bench
This benchmark measures how effectively agents are able to prospect, create campaigns, and manipulate data using the MCPs available for common GTM tools.
Why we built this. We wanted to understand where agents get stuck when using MCPs to complete real outbound workflows. These results are a first step, and we plan to build on this benchmark by adding more harnesses & models, and more tasks covering a broader range of use cases. We use the results to improve Cluster’s MCP.
How it works. We compare the performance of the MCPs across 26 tasks spanning common outbound workflows like campaign creation, prospecting, and list building.
- Campaign creation: 8 tasks
- Prompt to campaign: 9 tasks
- Qualification: 7 tasks
- CSV import: 2 tasks
The tasks use anonymized queries derived from real user examples.
For each (task, provider) combination we start a fresh Codex CLI agent (Using GPT 5.5 xhigh) with a task brief, any supplied inputs, and the provider's MCP. The agent discovers tools and works through the task. We record its time and token usage, then verify the agent produced the expected outputs to pass the task (e.g. creating a campaign).
We use the publicly hosted MCP server for each provider. Tests were last run on: .
Task coverage
| MCP provider | Campaign creation8 tasks | Prompt → campaign9 tasks | Qualification7 tasks | CSV import2 tasks | Total coverage26 tasks |
|---|---|---|---|---|---|
| Cluster | 8 / 8 | 9 / 9 | 7 / 7 | 2 / 2 | 26 / 26 |
 Lemlist | 7 / 8 | 9 / 9 | 7 / 7 | 2 / 2 | 25 / 26 |
 Clay | 0 / 8 | 0 / 9 | 7 / 7 | 0 / 2 | 7 / 26 |
 HeyReach | 4 / 8 | 0 / 9 | 0 / 7 | 1 / 2 | 5 / 26 |
Coverage means supported by the MCP configuration tested, not a successful run or a claim about the entire product. Clay’s connected MCP lacked campaign creation; HeyReach’s lacked email and native research.
Campaign creation
The agent receives recipients and message copy, then saves a draft with the required sequence, variables, and delays. Tasks range from 3 to 1,000 contacts. Seven tasks are shared by Cluster and Lemlist; the average compares those tasks.
| Task | Cluster | Lemlist | Clay | HeyReach |
|---|---|---|---|---|
| Basic email — 3 leads | $0.48 | $1.04 | Not supported | Not supported |
| Basic LinkedIn — 3 leads | $0.54 | $1.03 | Not supported | $1.03 |
| Personalized LinkedIn — 40 leads | $0.69 | $1.42 | Not supported | $1.09 |
| LinkedIn + email — 40 leads | $0.64 | $1.53 | Not supported | Not supported |
| Connection requests, CSV — 87 leads | $0.54 | $0.88 | Not supported | $3.46 |
| Connection requests, Markdown — 87 leads | $0.64 | $1.22 | Not supported | $4.28 |
| Email — 1,000 leads | $0.54 | $0.97 | Not supported | Not supported |
| LinkedIn + image — 3 leads | $0.45 | Not supported | Not supported | Not supported |
| Average | $0.587 tasks | $1.167 tasks | Not supported | $2.464 tasks |
The average uses seven tasks for Cluster and Lemlist, and four supported tasks for HeyReach. All measured runs are included. Costs cover the agent’s model tokens, excluding product credits and subscriptions.
Prompt to campaign
Given a targeting brief, the agent finds people, researches them, and saves a personalized draft campaign. Nine tasks are shared by Cluster and Lemlist.
| Task | Cluster | Lemlist | Clay | HeyReach |
|---|---|---|---|---|
| Restaurant-group alumni | $1.38 | $7.21 | Not supported | Not supported |
| Experienced benefits brokers | $2.52 | $4.33 | Not supported | Not supported |
| Boston ecommerce operators | $2.27 | $10.21 | Not supported | Not supported |
| AI-native GTM teams | $2.51 | ≥ $12.87 | Not supported | Not supported |
| Indian developer decision-makers | $1.59 | $6.14 | Not supported | Not supported |
| Retail/media directors and VPs | $1.44 | $6.32 | Not supported | Not supported |
| Enterprise CIOs | $1.57 | $9.21 | Not supported | Not supported |
| JLL & CBRE leasing brokers | $1.95 | $7.82 | Not supported | Not supported |
| CPO / COO targeting — 40 leads | $1.43 | $2.33 | Not supported | Not supported |
| Average | $1.859 tasks | ≥ $7.389 tasks | Not supported | Not supported |
One selected trial per provider per task; all measured runs are included. ≥ marks a lower bound because one Lemlist run timed out with partial token accounting. Completion checks the saved campaign structure.
Example task: Northeastern alumni in restaurant groups
One selected trial per provider. These figures describe this task only.
Task brief · condensed
Find 10 Northeastern University alumni currently working at US-headquartered restaurant companies.
Create a personalized draft LinkedIn campaign. Save a distinct, factual opening line and a supporting public URL for each person. Do not send.
Agent run
21 turns
$1.38 in tokens
4m 36s
Lemlist49 turns
$7.21 in tokens
9m 42s
Each block is one agent turn.
Output summary · both providers
LinkedIn campaign
Connection request
Wait 1 day
Personalized message
Wait 3 days
Follow-up message
Personalized opening lines and supporting URLs saved per contact.
Both drafts passed structural checks. Research accuracy was not independently audited. Token costs exclude product credits.
Qualification
Given a list and qualification criteria, the agent uses each platform’s native research tools to return a decision and explanation for each entry. Results cover six shared tasks across Cluster, Lemlist, and Clay.
| Task | Cluster | Lemlist | Clay | HeyReach |
|---|---|---|---|---|
| Air-export fit | $0.64 | $7.25 | $4.56 | Not supported |
| Conditional qualification tiers | $0.77 | $1.55 | $1.10 | Not supported |
| Current technical role | $0.66 | $2.30 | $2.42 | Not supported |
| Production agents & HQ | $0.60 | $4.87 | $1.81 | Not supported |
| R&D with exceptions | $0.72 | $2.28 | $2.85 | Not supported |
| Security operators | $0.73 | $5.61 | $2.16 | Not supported |
| Average | $0.696 tasks | $3.986 tasks | $2.486 tasks | Not supported |
Tasks are weighted equally, with all measured runs included. The 100-person task was excluded after a callback failure.
Example task: Production agents and headquarters
9MCP calls
Configure a research column, run the batch, wait for the results.
Lemlist22MCP calls
Set up a campaign and AI variable, add leads, then check and troubleshoot.
Clay41MCP calls
Separate company-research requests and result checks for each company.
This describes the company-research tools used in this run; Clay also has batch contact enrichment. A Lemlist fallback hit the harness’s approval policy.
CSV import
The agent imports a small list or appends two CSVs into one destination. The two tasks are each run five times per supported provider and checked for preservation of the supplied rows and fields.
| Task | Cluster | Lemlist | Clay | HeyReach |
|---|---|---|---|---|
| Small CSV — 6 rows | $0.52 | $0.86 | Not supported | $0.91 |
| Two CSVs — 135 rows | $0.44 | $0.74 | Not supported | Not supported |
| Average | $0.482 tasks | $0.802 tasks | Not supported | $0.911 task |
Each task averages five trials, including failures. The average row weights supported tasks equally. Completion is the share of trials that fully passed the row and field preservation checks.
Methodology
Run by Cluster to evaluate our MCP alongside other providers. Results reflect selected runs.
Each trial starts a fresh Codex CLI agent (Using GPT 5.5 xhigh) with a task brief, any supplied inputs, and one provider’s publicly hosted MCP server. The agent discovers the available tools and decides how to work through the task.
We record agent time, model-token usage, and tool calls, including recovery from errors. After the agent finishes, we check the saved product state and recorded evidence against the task’s requirements to determine whether it passed.