Live survey304 models · 14 providers · 23,592 auditsAudited implement · SAFE →What changedHow we verifyllms.txt

MCP server · Audited from source

headroom-ai MCP server

by headroomlabs-ai/headroomApache-2.0 licenceAudited 13 Nov 2026 · source 25b1342e6242

npm headroom-ai 0.44.0stdiostreamable-http★ 79,509
✕BLOCKGrade F · 37/100 · 13 Nov 2026 · 25b1342

Is it safe? No — BLOCK, grade F (37/100). It exposes 3 tools: 2 read, 1 write and 0 destructive. The audit recorded 98 findings (4 critical, 0 high, 15 medium, 74 low). Maintainer's description: Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.

Ask an AI about this pageGChatGPTCClaudePPerplexityGeGeminiXGrok
More
Cite

Cite this page

ModelsAtlas. “headroom-ai MCP server (headroomlabs-ai/headroom)”. Retrieved 13 Nov 2026. https://modelsatlas.com/mcp/headroomlabs-ai/headroom

Why this is BLOCK. The audit found 4 critical and 0 high findings; the most severe is Hard-coded secrets at .gitguardian.yaml:34. Each finding is listed below with its file and line.

Why this page is still public. People search for this server by name. Hiding it would leave them with the README's word alone; here they see what the code does. We list it so you can decide, and we do not show connect commands for BLOCK servers.

Trust audit

Audit 0.4.1 · full source read 13 Nov 2026

Trust score37/100✕BLOCK F
4 critical0 high15 medium74 low5 info

98 findings, listed below with the file and line each was found on. Grade A needs a sandboxed behavioural run too; this audit read the source only.

What the source does

  • FilesystemDeclared9 observations
  • NetworkDeclared9 observations
  • ShellDeclared6 observations
  • DependenciesNot pinnedSome ranges float
  • SecretsFoundSee the findings

Findings (98)

Sorted by severity · the 25 most severe of 98 are listed

CRITICALHard-coded secretssecret.anthropic · .gitguardian.yaml:34
F-03 Hard-coded secretsCWE-798 · CWE-321Layer L1

Excerpt withheld: it is a secret, and we never print one.

Why
No further detail recorded for this rule.
Fix
—
CRITICALHard-coded secretssecret.anthropic · crates/headroom-core/benches/auth_mode.rs:41
F-03 Hard-coded secretsCWE-798 · CWE-321Layer L1

Excerpt withheld: it is a secret, and we never print one.

Why
No further detail recorded for this rule.
Fix
—
CRITICALHard-coded secretssecret.anthropic · crates/headroom-core/benches/auth_mode.rs:62
F-03 Hard-coded secretsCWE-798 · CWE-321Layer L1

Excerpt withheld: it is a secret, and we never print one.

Why
No further detail recorded for this rule.
Fix
—
CRITICALHard-coded secretssecret.anthropic · crates/headroom-proxy/src/cache_stabilization/drift_detector.rs:761
F-03 Hard-coded secretsCWE-798 · CWE-321Layer L1

Excerpt withheld: it is a secret, and we never print one.

Why
No further detail recorded for this rule.
Fix
—
MEDIUMCode injectioncode.dynamic_import · headroom/learn/registry.py:32
F-02 Code injectionCWE-78 · CWE-94 · CWE-95Layer L1
mod = importlib.import_module(f"headroom.learn.plugins.{mod_name}")
Why
No further detail recorded for this rule.
Fix
—
MEDIUMNetwork egressnet.beacon_words · deploy/beacon/test-rollup.mjs:65
F-04 Network egressCWE-200 · CWE-319Layer L1
const beacon = (install, id, seq) =>
Why
No further detail recorded for this rule.
Fix
—
MEDIUMNetwork egressnet.beacon_words · deploy/beacon/test-rollup.mjs:96
F-04 Network egressCWE-200 · CWE-319Layer L1
[`sessions/${PART}/a.json`]: [beacon('i1', 's1', 3), beacon('i1', 's2', 1)].join('\n'),
Why
No further detail recorded for this rule.
Fix
—
MEDIUMNetwork egressnet.beacon_words · deploy/beacon/test-rollup.mjs:97
F-04 Network egressCWE-200 · CWE-319Layer L1
[`sessions/${PART}/b.json`]: beacon('i1', 's1', 9),
Why
No further detail recorded for this rule.
Fix
—
MEDIUMNetwork egressnet.beacon_words · deploy/beacon/test-rollup.mjs:98
F-04 Network egressCWE-200 · CWE-319Layer L1
[`sessions/${PART}/c.json`]: beacon('i1', 's1', 7),
Why
No further detail recorded for this rule.
Fix
—
MEDIUMNetwork egressnet.beacon_words · deploy/beacon/test-rollup.mjs:100
F-04 Network egressCWE-200 · CWE-319Layer L1
[`sessions/${PART}/d.json`]: beacon('i2', 's1', 2),
Why
No further detail recorded for this rule.
Fix
—
MEDIUMNetwork egressnet.raw_ip · Dockerfile:212
F-04 Network egressCWE-200 · CWE-319Layer L1
CMD curl --fail --silent "http://127.0.0.1:${HEADROOM_PORT:-8787}/readyz"
Why
No further detail recorded for this rule.
Fix
—
MEDIUMNetwork egressnet.raw_ip · Dockerfile:255
F-04 Network egressCWE-200 · CWE-319Layer L1
CMD ["python3", "-c", "import os, urllib.request; urllib.request.urlopen('http://127.0.0.1:' + (os.environ.get('HEADROOM_PORT') or '8787') + '/readyz', timeout=5)"]
Why
No further detail recorded for this rule.
Fix
—
MEDIUMObfuscation / stealthobf.homoglyph · scripts/export_kompress_v2_onnx.py:243
F-07 Obfuscation / stealthCWE-506 · CWE-94Layer L1
"Verify: max|Δscore|=%.4f  keep-decision agreement=%.1f%% (fp32 ~100%%, int8 ~98-100%%)",
Why
No further detail recorded for this rule.
Fix
—
MEDIUMHard-coded secretssecret.anthropic · crates/headroom-core/tests/auth_mode.rs:172
F-03 Hard-coded secretsCWE-798 · CWE-321Layer L1

Excerpt withheld: it is a secret, and we never print one.

Why
No further detail recorded for this rule.
Fix
—
MEDIUMHard-coded secretssecret.generic · docs/scripts/verify-sdk-wire.mjs:40
F-03 Hard-coded secretsCWE-798 · CWE-321Layer L1

Excerpt withheld: it is a secret, and we never print one.

Why
No further detail recorded for this rule.
Fix
—
MEDIUMHard-coded secretssecret.generic · tests/test_compression/test_masks.py:247
F-03 Hard-coded secretsCWE-798 · CWE-321Layer L1

Excerpt withheld: it is a secret, and we never print one.

Why
No further detail recorded for this rule.
Fix
—
MEDIUMHard-coded secretssecret.generic · tests/test_compression/test_masks.py:254
F-03 Hard-coded secretsCWE-798 · CWE-321Layer L1

Excerpt withheld: it is a secret, and we never print one.

Why
No further detail recorded for this rule.
Fix
—
MEDIUMHard-coded secretssecret.generic · tests/test_compression/test_universal.py:354
F-03 Hard-coded secretsCWE-798 · CWE-321Layer L1

Excerpt withheld: it is a secret, and we never print one.

Why
No further detail recorded for this rule.
Fix
—
MEDIUMHard-coded secretssecret.generic · tests/test_compression_safety_rails.py:375
F-03 Hard-coded secretsCWE-798 · CWE-321Layer L1

Excerpt withheld: it is a secret, and we never print one.

Why
No further detail recorded for this rule.
Fix
—
LOWInventory / provenanceinv.hidden_file · .actrc
F-13 Inventory / provenanceCWE-1104Layer L0
.actrc
Why
hidden member outside the usual dotfiles
Fix
review its purpose
LOWInventory / provenanceinv.hidden_file · .actrc.local.example
F-13 Inventory / provenanceCWE-1104Layer L0
.actrc.local.example
Why
hidden member outside the usual dotfiles
Fix
review its purpose
LOWInventory / provenanceinv.hidden_file · .changelog.md
F-13 Inventory / provenanceCWE-1104Layer L0
.changelog.md
Why
hidden member outside the usual dotfiles
Fix
review its purpose
LOWInventory / provenanceinv.hidden_file · .commitlintrc.json
F-13 Inventory / provenanceCWE-1104Layer L0
.commitlintrc.json
Why
hidden member outside the usual dotfiles
Fix
review its purpose
LOWInventory / provenanceinv.hidden_file · .env.act.example
F-13 Inventory / provenanceCWE-1104Layer L0
.env.act.example
Why
hidden member outside the usual dotfiles
Fix
review its purpose
LOWCode injectioncode.deserialize · tests/test_pipeline_outcome_stage.py:129
F-02 Code injectionCWE-78 · CWE-94 · CWE-95Layer L1
restored = pickle.loads(pickle.dumps(snap))
Why
deserialises untrusted bytes into live objects
Fix
use json or yaml.safe_load

Tools it exposes

3 declared in source · 0 destructive

2 read1 write0 destructive
ToolWhat it doesRisk
headroomShow what Headroom snipped from each request (addwrite
headroom_retrieveRetrieve original uncompressed content from Headroomread
lookupLocal fixture.read

Safer alternatives

Audit history

  1. ✕BLOCK F 37 / 100

    First audit · 25b1342e6242

README badge

Badges show the current verdict, BLOCK included

ModelsAtlas: BLOCK FMODELSATLASBLOCK · F

A badge cannot be hidden by the maintainer while the verdict is BLOCK. Fix the findings, push, and the re-audit updates it.

Questions about headroom-ai MCP server

What is the headroom-ai MCP server?
Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.
What tools does headroom-ai expose?
3 in total: 2 read-only, 1 that write, and 0 that can delete or overwrite. Every one is listed on this page with its risk.
Is headroom-ai safe to connect to an agent?
No — not without reading the findings first. The audit graded it F (37/100) and found 4 critical or high issues in the source. Each one is listed on this page with the file and line it is on.
What credentials does headroom-ai need?
It reads ANTHROPIC_API_KEY, API_KEY, AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_SESSION_TOKEN, CAPTURE_PROXY_TOKEN, CAPTURE_PROXY_TOKEN_HOSTS, CLAUDE_CODE_OAUTH_TOKEN, FOO_API_KEY, GCP_ACCESS_TOKEN, GEMINI_API_KEY and GITHUB_COPILOT_API_TOKEN from the environment. Give it a token scoped to the least it needs — an agent that can be talked into calling a tool can be talked into calling it with your credentials.
How does headroom-ai run?
It speaks stdio and streamable-http, so it runs as a local process your client starts. It is published on npm as headroom-ai at 0.44.0.
How current is this page?
The grade is for one exact copy of the source (25b1342e6242), read on 2026-11-13. The repository is watched and re-audited when it changes.
I already installed headroom-ai. What should I do?
Remove it from your client's MCP configuration, then rotate any API keys and tokens that were in your environment while it ran.
Can the verdict change?
Yes. The repository is watched and re-audited when its source changes. If the findings are removed the verdict updates, and the history on this page shows it.

Sources

  • headroomlabs-ai/headroom on GitHub — the source we read, its licence, stars and activityObserved · 13 Nov 2026
  • ModelsAtlas trust audit — BLOCK verdict for commit 25b1342e6242Observed · 13 Nov 2026

Spotted something wrong? Report a correction · Methodology

Keep going in your AI

Opens with this page as the source. We add nothing else to the prompt.