mcp-bastion

---

mcp-bastion sits between your MCP client (Claude Code, Cursor, Cline, Windsurf, Zed, Claude Desktop, or any MCP-compliant agent) and your MCP servers. It is client-agnostic — it works with any compliant client through configuration alone, with zero client-specific code — and non-invasive: your servers run unchanged, and removing Bastion is a one-line config revert.

📦 Package: mcp-bastion on npm · 🗂️ Official MCP Registry: io.github.Gowthaman90/mcp-bastion

🔒 Security, measured: on the open, vendor-neutral mcp-defense-bench v0.9.0, Bastion covers 66% of the MCP attack surface (24.6/37 vectors; 23 enforced) on development fixtures and 53% on the pre-registered held-out set, at zero false positives on matched controls — the broadest of the proxies measured.

📖 Launch story: Medium · dev.to

👤 Created & maintained by Gowthaman Arumugam — Independent Researcher. Companion benchmark: mcp-defense-bench.

Contents

• Why
• How it works
• Security model
• Features
• Quick start
• Demo
• Control tools
• Configuration
• Transports
• Per-user identity (gateway mode)
• Delegation checks
• Runtime authorization
• Runtime security
• Audit & compliance
• Client setup
• Architecture
• Development
• Roadmap
• Contributing
• Security
• License

Why

When an MCP server disconnects mid-session, the agent only sees a generic "No such tool available" error — indistinguishable from a tool that never existed — and it cannot reconnect; only a human can. Long agent sessions silently lose capabilities and fail in confusing ways.

Bastion closes that gap. It health-checks every server, auto-reconnects with backoff, and — crucially — exposes control tools so the agent itself can inspect connection health and recover a dropped server without human intervention.

Bastion now spans three layers: reliability (v0.1), runtime security (v0.2 — tool pinning / rug-pull & poisoning detection), and audit & compliance (v0.3 — pluggable sinks mapped to NIST AI RMF / OWASP LLM Top 10). See the roadmap.

How it works

Today your client connects directly to each server. With Bastion, your client connects to Bastion, which connects to those same servers on your behalf — so it sits in the tool-call path and can add reliability (and, later, security) transparently.

Before:   Client ─▶ server A / server B / server C

After:    Client ─▶ mcp-bastion ─▶ server A
                                  ─▶ server B
                                  ─▶ server C

Bastion is a standard MCP server to your client and a standard MCP client to each upstream. Because it speaks the protocol faithfully, it works with every compliant client automatically — the only per-client difference is where you put a few lines of config.

Security model

Bastion's security rests on three independent mechanisms. Heuristics feed all three and replace none of them: a heuristic finding is a risk signal, never a proof that a tool is safe.

Mechanism	What it gives you	Checks
Definition integrity	The tool, server and configuration you approved are the ones being called	tool pinning (rug pull), server identity, config drift, argument schema
Runtime authorization	Each caller reaches only the tools and argument values its policy grants, decided before any upstream contact	authorization rules, delegation checks (who acts for whom), MRTR continuation custody
Trustworthy audit evidence	Every decision is recorded without leaking secrets, and a changed or shortened log can be detected	audit events, hash chain, signed checkpoints, sink health, redaction

Every finding carries a kind: deterministic (an exact comparison against a pin, a schema, a policy or a token's claims) or heuristic (pattern matching that can misfire on benign text and can be evaded by a novel payload). What each enforcement profile does with them:

Check	Kind	observe	balanced (default)	strict
Tool pinning (rug pull), server identity	deterministic	warn	block	block
Runtime authorization	deterministic	warn	block	block
Delegation token checks	deterministic	warn	block	block
Delegation chain depth / cycles	deterministic	warn	warn	block
Cross-server data flow	deterministic	warn	block	block
Argument inspection: schema, plus command-injection and injected-instruction patterns	mixed	warn	block	block
MRTR (input_required) gate	mixed	warn	block	block
Description poisoning	heuristic	warn	warn	block
Response heuristics (injection, steering)	heuristic	warn	warn	block

Argument inspection blocks under balanced even though part of it is heuristic: a false positive costs the agent a retry, while a miss can run a command. Inline secret redaction in results (redactResponseSecrets) is on in every profile. Under observe nothing is blocked and each audit event records wouldBlock: what balanced would have done.

Known limits. Poisoning and response heuristics can be evaded by payloads they have not seen. Redaction covers known secret shapes, not every secret. Bastion does not sandbox the upstream server itself. Authorization trusts the identities it is given: bearer clients it authenticates itself, and delegation claims it verifies only in signature mode (see trustStructuralClaims).

Features

• 🔌 Client-agnostic — one binary, config-only integration; no per-client plugins.
• ♻️ Self-healing — health checks + capped exponential-backoff auto-reconnect for stdio servers.
• 🧭 Agent-recoverable — bastion__status and bastion__reconnect let the agent detect and fix drops itself, instead of hitting an opaque "no such tool" wall.
• 🧩 Transparent aggregation — merges many servers into one, with per-server tool namespacing to prevent collisions and tool-shadowing.
• 💬 Legible failures — a dropped server yields an actionable message, not a crash.
• 🛡️ Runtime security (new in v0.2) — pins each tool's definition and blocks "rug pulls" (a server changing a tool after approval); heuristically inspects descriptions for poisoning; detects cross-server shadowing. See Runtime security.
• 🔐 Runtime authorization (new in v1.4, opt-in) — per-caller, per-tool policy decided before any upstream contact: which client, user and agent may call which tool with which argument values, deny-overrides, default-deny, and an observe mode to preview a policy. See Runtime authorization.
• 📝 Audit & compliance (opt-in) — structured, hash-chained audit events to pluggable sinks (console / file / webhook / OTLP), mapped to NIST AI RMF & OWASP LLM Top 10; since v1.4 with HTTPS-only remote sinks, retrying delivery, per-sink redaction and Ed25519-signed checkpoints that make a shortened log detectable. See Audit & compliance.
• 🪶 Non-invasive & reversible — your servers run unchanged; uninstall is a config revert.
• 🧱 Enterprise-grade codebase — strict TypeScript, layered architecture, ESLint + Prettier, and unit + end-to-end tests.

Quick start

Bastion is published on npm as mcp-bastion — the npx command below fetches it automatically, so there's nothing to install first.

1. Add Bastion to your client, pointing it at a config file:

// your client's mcpServers config
{
  "mcpServers": {
    "bastion": {
      "command": "npx",
      "args": ["-y", "mcp-bastion", "--config", "bastion.config.json"],
    },
  },
}

2. List your real servers in bastion.config.json (moved verbatim from the client):

{
  "servers": {
    "github": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"] },
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/dir"],
    },
  },
  "reconnect": { "auto": true },
  "healthCheck": { "enabled": true },
}

3. Restart your client. Your tools now appear namespaced (e.g. github__create_issue) alongside Bastion's control tools. See bastion.config.example.json for the full set of options.

🔒 Security is on by default. Out of the box, Bastion runs the balanced enforcement profile: it blocks high-confidence attacks (rug-pulls, argument/command injection, cross-server exfiltration) and warns on heuristic ones (description/response poisoning), while redacting leaked secrets from tool results. Set security.enforcementProfile to observe (warn-only) or strict (block-all), or tune any individual control — see Runtime security.

Demo

See the whole thing in action — a server crashing mid-session and healing itself:

npm run demo

It boots Bastion in front of a server that crashes on command, shows the agent getting an actionable "reconnect" message instead of a cryptic error, and then the connection auto-recovering with no human involved. To record it as a GIF: asciinema rec demo.cast -c "npm run demo" && agg demo.cast assets/demo.gif.

Control tools

Bastion injects control tools so the agent can manage connections and review security itself, using only standard MCP calls:

Tool	Purpose
bastion__status	Health of every proxied server: connected / disconnected / reconnecting / failed, tool counts, last error.
bastion__reconnect	Reconnect a named server (argument: { "server": "<name>" }) without human intervention.
bastion__security	Per-tool security report: pin status (approved vs changed), poisoning findings, and shadowing.
bastion__compliance	Audit summary of recent activity mapped to NIST AI RMF / OWASP LLM Top 10 (requires audit.enabled).

Re-approval is operator-only. Clearing a rug-pull block is a security authority, so it is not an agent-callable tool — a prompt-injected agent must not be able to re-approve the very tool it was blocked from. A changed tool stays blocked until an operator clears it out-of-band; bastion__approve is not advertised and a client call to it is refused.

Configuration

Key	Type	Default	Description
servers	map	—	Upstream servers to proxy (required, at least one).
servers.<name>.command	string	—	Executable to launch (e.g. npx, node).
servers.<name>.args	string[]	[]	Arguments to command.
servers.<name>.env	map	—	Env overrides merged over the process env.
servers.<name>.cwd	string	—	Working directory for the spawned process.
reconnect.auto	boolean	true	Auto-reconnect after an unexpected disconnect.
reconnect.maxRetries	number	10	Max attempts before giving up (-1 = unlimited).
reconnect.initialBackoffMs	number	500	Initial backoff, doubled each attempt.
reconnect.maxBackoffMs	number	30000	Backoff ceiling.
healthCheck.enabled	boolean	true	Enable periodic liveness probing.
healthCheck.intervalMs	number	30000	Interval between probes.
healthCheck.timeoutMs	number	5000	Per-probe timeout.
namespace.strategy	prefix | passthrough	prefix	How upstream tool names are exposed.
namespace.separator	string	__	Separator used by the prefix strategy.
security.pinTools	boolean	true	Pin tool definitions and detect later changes.
security.onRugPull	block | warn	block	Action when a pinned tool's definition changed.
security.inspectDescriptions	boolean	true	Run poisoning heuristics on tool descriptions.
security.onPoisoning	block | warn	warn	Action on a high-severity poisoning finding.
audit.enabled	boolean	false	Record an audit event for every tool call.
audit.includeArgs	none|redacted|full	none	How tool arguments are recorded.
audit.tamperEvident	boolean	false	Hash-chain events so tampering is detectable.
audit.sinks	array	console	Destinations: console, file, webhook, otlp.
servers.<name>.transport	stdio | http	stdio	Local subprocess or remote endpoint.
servers.<name>.url	string	—	Remote MCP URL (required for http).
servers.<name>.headers	map	—	Headers for http upstreams (e.g. Authorization).
servers.<name>.forwardHeaders	map	—	http only. Upstream header ← incoming header, copied per request (gateway mode). Needs listen.mode: "http".
servers.<name>.onMissingForwardedHeader	error | omit	error	When a request lacks a forwarded header: answer a tool error without calling the upstream, or call it without.
servers.<name>.connect	startup | onDemand	startup	onDemand: connect on the first request that needs this upstream, with that request's forwarded headers.
servers.<name>.authHeaders	string[]	—	Extra header names that count as authentication in bastion__status (forwardHeaders keys always count).
servers.<name>.delegation	object	—	http only, needs forwardHeaders. Keyless checks on the delegate JWT in a forwarded header (below).
servers.<name>.delegation.verify	structural | signature	structural	signature also verifies the JWT against public keys (jwksUrl or jwks), with issuer and an algorithms allow-list.
delegation.gatewayAudiences	string[]	[]	Audiences that identify Bastion itself; a token for one of them (or with no aud) is passthrough.
delegation.oneUserPerRequest	boolean	true	All delegate tokens in one request must name the same user (iss + sub).
delegation.recordSubject	hash | plain	hash	Audit subject: HMAC pseudonym of iss|sub, or sub as is.
delegation.subjectKey	string (≥16)	see below	Pseudonym key; or MCP_BASTION_SUBJECT_KEY, then the audit integrity key, then random per process.
security.onDelegation	block | warn	block	Action on a delegation finding other than the chain rules.
security.onDelegationChain	block | warn	warn	Action on delegation-chain-depth / -chain-cycle (block under strict).
listen.mode	stdio | http	stdio	Serve Bastion over stdio or Streamable HTTP.
listen.host / listen.port	string / number	127.0.0.1 / 3000	Bind address for http mode.
listen.validateRoutingHeaders	boolean	true	Reject Mcp-* routing headers that disagree with the body (-32020, MCP 2026-07-28).
security.maxCacheTtlMs	number (ms)	3600000	Ceiling on upstream ttlMs cache hints forwarded downstream (MCP 2026-07-28).
servers.<name>.protocol	auto|legacy|2026-07-28	auto	Era negotiation per upstream (probe 2026-07-28, fall back; or pin).
listen.legacy	stateless | reject	stateless	Serve pre-2026-07-28 clients statelessly, or refuse them (-32022).
security.inspectInputRequests	boolean	true	Gate input_required rounds (credential phishing, model steering).
security.onInputRequired	block | warn	block	Action on a high-severity MRTR finding (warn strips the request).
security.requestStateKey	string (≥16)	random per process	HMAC key sealing requestState envelopes; or MCP_BASTION_REQUEST_STATE_KEY.
security.requestStateTtlSeconds	number	300	Lifetime of a sealed requestState envelope.
security.enforcementProfile	observe | balanced | strict	balanced	Sets every on* action as a group; see Security model. An explicit on* overrides it.
security.onAuthorization	block | warn	block	Action when authorization refuses a call (warn under observe).
security.redactResponseSecrets	boolean	true	Redact secret values from tool results before the agent sees them.
authorization.enabled	boolean	false	Turn on runtime authorization; see Runtime authorization.
authorization.default	deny | allow	deny	Decision for an identity no rule allows.
authorization.rules	array	[]	{ id, effect, principal, tools, args? } rules; deny overrides allow.
authorization.trustStructuralClaims	boolean	false	Accept user/agent from structural-mode delegate tokens.
listen.authToken	string	—	Bearer token for the client default. Required for a non-loopback bind unless listen.clients is set.
listen.clients	map	—	Named clients: { "<name>": { "tokenEnv": "VAR" } }, one bearer token each.
audit.integrityKey	string	—	HMAC key for the chain; or MCP_BASTION_AUDIT_KEY.
audit.onSinkFailure	continue | block	continue	block refuses calls while no sink can deliver.
audit.sinks[].includeArgs	none|redacted|full	audit.includeArgs	Per-sink argument recording.
audit.sinks[].headersEnv	map	—	webhook/otlp: header name → environment variable.
audit.sinks[].allowInsecureTransport	boolean	false	webhook/otlp: allow plain HTTP to a non-loopback host.
audit.sinks[].maxQueue	number	10000	webhook/otlp: events held while the receiver is down.
audit.checkpoint	object	—	Signed checkpoints: privateKeyFile, everyEvents (100), everySeconds (60), stateFile, sinks.

Transports

Bastion runs on the MCP TypeScript SDK 2.0 and speaks both protocol eras — the stateless 2026-07-28 revision and the pre-2026 initialize handshake — on both faces, per upstream and per client. When the protocol is stateless, the gateway is the only component that can still hold security state, which is why v1.0 adds requestState custody and an input_required consent gate (see CHANGELOG.md).

Bastion speaks two transports on both faces:

• stdio (default) — the client spawns Bastion, and Bastion spawns local servers.
• Streamable HTTP — connect to remote MCP servers (servers.<name> with transport: "http", a url, and optional auth headers), and/or serve Bastion over HTTP to multiple/remote clients (listen.mode: "http", or --http <port>).

HTTP upstreams configured without an authentication header are flagged (authenticated: false) in bastion__status and warned at connect time. A custom credential header counts once it is named in authHeaders; a forwarded header always counts.

Per-user identity (gateway mode)

When Bastion sits between an agent and sub-agents that must act as the signed-in user, each upstream call has to carry that user's identity, not one fixed token. forwardHeaders does this:

{
  "servers": {
    "pfqa": {
      "transport": "http",
      "url": "http://localhost:8011/mcp",
      "protocol": "legacy",
      "connect": "onDemand",
      "forwardHeaders": { "x-cso-identity": "x-cso-delegate-pfqa" },
    },
  },
  "listen": { "mode": "http" },
}

The map reads upstream header ← incoming header. For every upstream request made while serving a downstream HTTP request (initialize, tools/list, tools/call, ping), Bastion copies the value of the incoming x-cso-delegate-pfqa header onto the upstream request as x-cso-identity.

Bastion holds no credential in this mode. It only passes on what each request carries.

• Bound to the request. The incoming headers live in an AsyncLocalStorage for exactly the lifetime of one downstream request. A custom fetch on the upstream transport reads them at send time. Concurrent requests from different users share one upstream connection but never see each other's headers.
• Only listed headers travel. No other incoming header is forwarded. Hop-by-hop and transport-owned headers (connection, mcp-session-id, content-type, …) are rejected as forward targets at config load. While listen.authToken is set, the incoming authorization header (Bastion's own bearer token) cannot be forwarded.
• A forwarded header overrides a static headers entry of the same name.
• Missing header. By default (onMissingForwardedHeader: "error") the upstream is not called. The caller gets a tool error such as pfqa: the request carried no x-cso-delegate-pfqa header, so the call was not forwarded. Use "omit" to call the upstream anyway and let it decide.
• On-demand connection. With connect: "onDemand", nothing is sent at start-up. The first request that needs the upstream (tools/list, or a direct tools/call of pfqa__…) connects it with that request's headers. The tool list is then cached as usual. Health checks skip on-demand upstreams, which show as idle until first used. A failed connect returns to idle with no reconnect timer. connect: "startup" with forwardHeaders connects at boot with the static headers only, so use onDemand when the upstream authenticates initialize too.
• 401 is an authentication failure. It is reported as such in the tool result and in bastion__status (lastError), never as a crash or an unknown tool. One user's 401 does not tear down the connection other users share.
• Values are never recorded. Logs, audit sinks and bastion__status never contain a forwarded value. Audit events record only which headers were forwarded ("forwardedHeaders": ["x-cso-identity"]).
• Pinning is shared. Tool definitions are pinned on first sight and a change is blocked, whichever user's request saw it. Only the identity is per request.

forwardHeaders, onMissingForwardedHeader, connect and authHeaders are valid only on transport: "http" upstreams. forwardHeaders also needs listen.mode: "http". Other combinations are rejected at config load. See examples/identity-forwarding.config.json for three sub-agents behind one Bastion.

Delegation checks

Since v1.2.0. In gateway mode a caller (say, a master agent working for a signed-in user) sends one short-lived delegate JWT per sub-agent, each in its own header. Each sub-agent sees only its own token. Bastion sees the whole request, so it can check what none of them can, without holding any secret or key:

"servers": {
  "pfqa": {
    "transport": "http", "url": "http://localhost:8011/mcp", "protocol": "legacy", "connect": "onDemand",
    "forwardHeaders": { "x-cso-identity": "x-cso-delegate-pfqa" },
    "delegation": {
      "audience": "pfqa",                 // required: what this upstream's tokens must carry in `aud`
      "header": "x-cso-identity",         // forwardHeaders key carrying the JWT (default: the only one)
      "verify": "structural",             // or "signature" (below)
      "maxLifetimeSeconds": 120,          // default 3600
      "clockSkewSeconds": 30,             // default 30
      "allowedActors": ["master-quote-agent"],
      "requireActor": false,
      "maxChainDepth": 2
    }
  }
},
"delegation": { "gatewayAudiences": ["mcp-bastion"], "oneUserPerRequest": true, "recordSubject": "hash" }

Rule	Fires when	Default (balanced)
delegation-malformed	not a strict JWS compact token (see below), or a wrongly typed claim	block
delegation-audience-mismatch	no aud value equals this upstream's audience (a sibling's token)	block
delegation-passthrough	aud is absent, or any aud value is in gatewayAudiences	block
delegation-subject-mixing	the request's delegate tokens name more than one (iss, sub)	block
delegation-expired / -not-yet-valid	now > exp + skew / nbf > now + skew	block
delegation-lifetime-excess	exp − iat > maxLifetimeSeconds, or no exp	block
delegation-actor-not-allowed / -actor-required	current actor (act.sub, else act.client_id) not in allowedActors / no act	block
delegation-chain-depth / -chain-cycle	the act chain is longer than maxChainDepth / names an actor twice	warn
delegation-alg-disallowed	signature mode: alg is not in algorithms (case-sensitive; none and HMAC never are)	block
delegation-key-unknown	signature mode: no trusted key matches the kid, or the key set could not be fetched	block
delegation-signature-invalid	signature mode: the signature does not verify under the trusted key	block
delegation-issuer-mismatch	signature mode: iss is not the configured issuer	block

Chain depth and cycles concern prior actors, which RFC 8693 §4.1 calls informational, so they are medium severity and only warned on unless security.onDelegationChain (or the strict profile) says block. Everything else follows security.onDelegation (warn under observe).

• Strict decoding. Exactly three base64url segments, no padding, canonical encoding, JSON objects with unique member names (after escape decoding) and a string alg; an optional Bearer prefix is allowed. Anything else fails closed as delegation-malformed.
• One verdict per request. On first need Bastion decodes every delegation-configured token the request carries, runs the per-upstream checks and the cross-token subject check, and caches the verdict for the rest of the request. The checkers are exported as checkDelegation(policy, carriedHeaders, now) (structural, synchronous) and checkDelegationAsync(policy, carriedHeaders, now) (adds signature verification).
• Before any upstream contact. A blocked tools/call is refused before Bastion connects (on-demand connects included), with a tool error naming the rule, never a token or claim value, and an audit event with decision blocked_delegation. tools/list leaves out an upstream the request is blocked from. As a backstop, the upstream transport refuses to send a token whose verdict blocks (reconnects and other sends made inside the request).
• Who acted for whom. Calls to a delegation-configured upstream carry an audit delegation record: subject, issuer, audience, tokenId (jti), expiresAt, actors (current first), chainDepth, verified (structural, signature when the signature verified, or failed) and the findings. Never the token. In hash mode the subject is "hmac:" + the first 16 hex characters of HMAC-SHA256(key, iss|sub): an operator with the key can recompute it to find one user's records. The key is delegation.subjectKey, else MCP_BASTION_SUBJECT_KEY, else the audit integrity key, else a random key per process, whose pseudonyms do not link across restarts.
• Status. bastion__status shows, per upstream, delegation: { mode, audience, lastViolation }, where the last violation is a rule id and a time only.
• Config checks. delegation needs forwardHeaders; header must be one of its keys; two upstreams cannot share an audience, and no upstream audience may be a gateway audience; unknown keys are rejected. Signature mode needs issuer, a non-empty algorithms list and exactly one key source.

Signature mode

When the issuer signs delegate tokens with an asymmetric key (ES256, EdDSA, RS256, PS256, …), Bastion can verify them against the issuer's public keys:

"delegation": {
  "audience": "pfqa",
  "verify": "signature",
  "issuer": "https://spire.example/identity",               // exact match on `iss`
  "algorithms": ["ES256", "EdDSA"],                          // allow-list; HS256/384/512 and none are rejected
  "jwksUrl": "https://spire.example/.well-known/jwks.json"   // or "jwks": { "keys": [ … public JWKs … ] }
}

• Keys come only from the configured source. jku, x5u, x5c and embedded jwk headers are never followed. An inline jwks with private or secret members (d, p, q, dp, dq, qi, oth, k) is rejected at config load. jwksUrl must be HTTPS, except to a loopback host.
• The algorithm allow-list runs first, case-sensitively, before any key lookup: alg: none (in any casing) and HS256 signed with the public key are refused without a fetch. A critical header extension (crit, e.g. an unencoded payload) is refused.
• Remote key sets are fetched on first use and cached. A token with an unknown kid triggers at most one refetch per 30 s cooldown. A fetch failure or a 5 s timeout fails closed: the call is blocked as delegation-key-unknown, the upstream's lastError says the key set is unavailable, and further fetches wait for the cooldown.
• The structural rules still run on a verified token, under the same names, and the audit record says verified: "signature".

Honest limits.

• Structural mode authenticates nothing: a party that can forge tokens can forge consistent claims. It stops misrouting, passthrough, user mixing, stale relay and chain abuse by honest-but-wrong or prompt-injected callers. The upstream remains the verifier. Use signature mode (asymmetric tokens) for authentication at the gateway.
• Signature mode holds public keys only, so Bastion stays keyless: it can verify tokens but never mint one. It needs the issuer to sign with an asymmetric algorithm; HS256 tokens can only be checked structurally, because verifying them needs the shared secret. It trusts whatever the configured key set says, so the security of jwksUrl (HTTPS, the issuer's own host) is the security of the check.
• The one-user-per-request check covers only tokens configured with delegation.
• Replay inside the token's lifetime is not stopped (no jti cache yet; possible later, opt-in, bounded).

See examples/delegation.config.json for three sub-agents.

Runtime authorization

New in v1.4, opt-in. Decide who may call which tool, with which arguments, on every call and before any upstream server is contacted. Inactive unless authorization.enabled.

"listen": {
  "mode": "http",
  "clients": {                                   // one bearer token per client, read from the environment
    "docs-assistant":   { "tokenEnv": "BASTION_TOKEN_DOCS" },
    "release-pipeline": { "tokenEnv": "BASTION_TOKEN_RELEASE" },
    "ops-console":      { "tokenEnv": "BASTION_TOKEN_OPS" }
  }
},
"security": { "enforcementProfile": "observe" },   // preview first: log, never block
"authorization": {
  "enabled": true,
  "default": "deny",
  "rules": [
    { "id": "docs-read", "effect": "allow", "principal": { "client": "docs-assistant" },
      "tools": ["files__read_file", "files__list_dir"],
      "args": { "path": { "pathWithin": "/srv/docs" } } },
    { "id": "release-staging", "effect": "allow", "principal": { "client": "release-pipeline" },
      "tools": ["deploy__*"], "args": { "env": { "in": ["staging"] } } },
    { "id": "no-prod-rollback", "effect": "deny", "principal": "*",
      "tools": ["deploy__rollback"], "args": { "env": { "equals": "prod" } } },
    { "id": "ops-control", "effect": "allow", "principal": { "client": "ops-console" },
      "tools": ["bastion__*"] }
  ]
}

Who is calling. A call carries every identity Bastion has authenticated: the client (a named listen.clients bearer token, default for listen.authToken, or stdio), plus the user and the acting agent from the call's delegate token when there is one. Identities named in arguments, headers, _meta or input responses are ignored: a request cannot promote itself by claiming to be someone else.

How a call is decided.

• Every identity must be permitted. An agent acting for a user needs the user's grant and its own, so an agent's broad service grant is never used for a user who lacks the right (the confused deputy), and a user's right is never used by an agent that was not granted it. Earlier agents in a delegation chain grant nothing; only the current one counts.
• Deny overrides allow. Per identity, any matching deny refuses; otherwise an allow whose argument constraints all hold is needed. With default: "deny", no matching allow means no access.
• Rules. principal is "*" or one of client, user, agent (a name or a list). tools are namespaced names, exact or ending in one *. args constrain top-level arguments with equals, in, max, min and pathWithin. Matching is exact: no case folding, Unicode normalisation or type coercion; max/min accept JSON numbers only; pathWithin resolves . and .. and checks the boundary at a / (so /srv/docs-archive is not inside /srv/docs).
• Fail closed. A missing constrained argument fails an allow. A deny with args applies unless its constraints can be evaluated and do not hold, so a missing or wrongly typed argument cannot slip past a deny.
• Control tools (bastion__*) are authorized like any tool, under server name bastion. With default: "deny", grant them to an operator client, as ops-console above.

Where it runs. In tools/call, after the delegation checks and before any upstream contact, so a refused call never opens a connection carrying the caller's forwarded headers. Continuations (input_required retries) are re-authorized. tools/list hides tools a caller could never be allowed.

Rolling it out. Start under enforcementProfile: "observe": nothing is refused, and each audit event carries an authorization record (effect, deciding rule, violations such as agent:argument) plus wouldBlock: "blocked_authorization" where enforcement would have refused. When the log shows only refusals you intend, switch to balanced. Under balanced a refused call is answered Blocked by mcp-bastion: this caller is not authorized to call "<tool>" (rule <id>).; no other policy content is revealed.

Trust in user and agent identities. They come only from a trusted delegate token: one verified in signature mode, or, in structural mode, only with trustStructuralClaims: true. Set that only when the upstream verifies the same token itself, so a forged claim that passes Bastion is still rejected upstream. A rule naming a user or agent for a structural-mode upstream is a config error otherwise, and a token that is present but untrusted refuses the call (delegation:unverified).

Measured. This family of checks was pre-registered as test cases in mcp-defense-bench (tag authorization-prereg-2026-10-26) before this code was written. Per-caller MCP authorization also exists in other gateways (for example agentgateway's CEL rules and ToolHive's Cedar policies), and the benchmark's pre-registration commits to scoring them side by side.

Runtime security

New in v0.2. Bastion adds a security layer in the tool-call path (an interceptor pipeline), enabled by default:

• Rug-pull detection (tool pinning). Each tool's definition is pinned on first use. If a server later changes that definition, the tool is blocked (onRugPull: "block") until an operator reviews it and re-approves it out-of-band (operator-only — not an agent-callable tool). This catches a server that looks benign at install time and turns malicious afterward.
• Poisoning inspection. Tool names and descriptions are scanned for manipulation heuristics (instruction override, secret access, data exfiltration, covert instructions, embedded directives, hidden/zero-width characters). Because heuristics can false-positive, the default is warn (logged and reported, not blocked); set onPoisoning: "block" to enforce.
• Shadowing. When two servers expose a tool with the same name, it's surfaced in the report.

Review everything with the bastion__security tool. Each finding says whether it is heuristic or deterministic; see Security model.

Audit & compliance

Opt-in. Enable audit to record a structured, versioned event for every tool call, including calls refused by the security layer or by authorization:

"audit": {
  "enabled": true,
  "includeArgs": "redacted",            // none | redacted | full (default for every sink)
  "tamperEvident": true,                // hash chain; keyed with integrityKey / MCP_BASTION_AUDIT_KEY
  "onSinkFailure": "continue",          // or "block": refuse calls while no sink can deliver
  "sinks": [
    { "type": "file", "path": "./bastion-audit.jsonl", "includeArgs": "full" },
    { "type": "otlp", "endpoint": "https://collector.example:4318",
      "headersEnv": { "authorization": "OTEL_AUTH_HEADER" }, "includeArgs": "none" }
  ],
  "checkpoint": {                       // signed checkpoints (needs tamperEvident)
    "privateKeyFile": "/etc/bastion/checkpoint-private.pem",
    "stateFile": "/var/lib/bastion/audit-state.json",
    "sinks": [{ "type": "webhook", "url": "https://witness.example/checkpoints" }]
  }
}

Events. Each event names the server, tool, decision (allowed, blocked_rug_pull, blocked_authorization, …), outcome, findings with their kind, the authorization and delegation records when present, a runId, and a mapping to NIST AI RMF functions and OWASP LLM Top 10 categories. bastion__compliance returns aggregate totals.

Each sink is its own security boundary.

• console (stderr), file (JSON lines, created 0600), webhook and otlp (OpenTelemetry logs).
• Remote sinks must use HTTPS unless the host is loopback. Plain HTTP to a remote host is a config error unless the sink sets allowInsecureTransport: true, which logs a warning.
• Credentials come from the environment (headersEnv), not the config file; a missing variable stops startup.
• A failed delivery is retried with exponential backoff instead of being dropped. Events wait in a bounded queue (maxQueue, default 10 000); beyond it the oldest are dropped and counted. Receivers can also spot gaps from each event's seq.
• bastion__status shows each sink's health: healthy, delivered, dropped, queued, lastError.
• With onSinkFailure: "block", calls are refused (blocked_audit_unavailable) while no sink can deliver, for deployments where an unaudited call is not acceptable. Bastion's own control tools stay usable so the sinks can be inspected and recovered.

Redaction is first-class.

• Arguments are not recorded by default. Each sink can record its own level (includeArgs per sink): for example full arguments to a local file and none to a shared SIEM.
• redacted keeps structure and masks values under credential key names and values that look like secrets: cloud and provider keys, GitHub, Slack, Stripe, Google and npm tokens, JWTs, Bearer credentials, passwords in URLs, private-key blocks and NAME=secret assignments. The same patterns redact secrets from tool results before the agent sees them (redactResponseSecrets, on by default). Measured on a 45-case corpus (test/redaction-corpus.test.ts): 29/29 secrets caught, 0/16 benign look-alikes changed.
• Never written to any sink: tool results, delegate-token values, forwarded header values (names only), and the raw user sub when delegation.recordSubject is hash. Arguments are recorded only as includeArgs allows, so a secret passed as an argument is kept under full.
• Redaction is pattern-based. A secret in an unrecognised shape under an unrecognised key can still be recorded, so prefer none for high-sensitivity sinks.

Tamper evidence, and its limits.

• With tamperEvident, events form a hash chain. With an integrityKey the chain is HMAC-SHA256: without the key, a party who rewrites the log cannot recompute it. Unkeyed, it only detects naive edits.
• When sinks record arguments differently, the chain (v2) covers each event without its args view plus an argsDigest (HMAC of the full arguments), so every sink verifies the same chain.
• A hash chain cannot show that the newest records were cut off. Signed checkpoints can: audit.checkpoint writes an Ed25519-signed statement of the chain head every everyEvents (100) events, every everySeconds (60) seconds and at shutdown, and each run starts with a run-start record linked to the previous run through stateFile. Send checkpoints to a witness sink the log's host cannot rewrite.
• mcp-bastion audit verify <log> --checkpoints <witness.jsonl> --public-key <pem> checks the chain, seq continuity, checkpoint signatures, run links, a log cut short and a deleted run. Exit 0 intact, 1 tampered, 2 cannot verify. mcp-bastion audit keygen --out <dir> creates the key pair.
• Limit: a cut is detected only back to the last checkpoint that reached a witness the attacker cannot rewrite; without a witness, cutting a log together with its checkpoints is not detectable. An attacker holding the private key can forge checkpoints, so keep it off the log's host.

Shadow mode. Under security.enforcementProfile: "observe" nothing is blocked, and each event says what enforcement would have done: wouldBlock holds the decision balanced would have recorded, a rug pull appears in findings as tool-definition-changed, and responseFindings / redactedSecrets record what the result scan and inline DLP found (or secret-in-result when redaction is off).

Standards alignment

Every security check maps to recognized frameworks — the NIST AI Risk Management Framework (a U.S. federal standard), the OWASP Top 10 for LLM (2025) and Agentic (2026) Applications, and STRIDE. The full per-check mapping (with reference links to the MCP-security literature the checks are drawn from) is in docs/CHECKS-MAPPING.md.

Coverage is measured independently — bastion is scored against a 22-vector attack surface by the vendor-neutral mcp-defense-bench benchmark, with a public leaderboard and framework mapping.

Client setup

The steps are identical for every client — only the config file location differs:

Client	Where to add the bastion entry
Claude Code	project .mcp.json (or claude mcp add)
Cursor	~/.cursor/mcp.json or project .cursor/mcp.json
Claude Desktop	claude_desktop_config.json
Cline	cline_mcp_settings.json
Windsurf	~/.codeium/windsurf/mcp_config.json

Gradual adoption: you don't have to route every server through Bastion — put only your flaky or untrusted servers behind it and leave the rest connected directly.

Architecture

Bastion is organized into clear layers with a one-directional dependency flow, so each concern is independently testable and easy to evolve:

src/
├── cli.ts              # thin CLI entrypoint (parse → wire → serve)
├── index.ts            # public library API
├── errors.ts           # error hierarchy (BastionError, …)
├── config/             # schema (Zod) + loader
├── core/               # domain: upstream connection lifecycle, aggregation & routing
├── proxy/              # client-facing MCP server + control tools
├── observability/      # logging (audit sinks in v0.3)
└── internal/           # small cross-cutting utilities

Design details — including the client-agnostic rationale, the interceptor pipeline, and the audit-sink strategy — live in the project's design docs.

Development

npm install
npm run check      # format:check + lint + typecheck + test (the full gate)
npm test           # unit + end-to-end (in-memory transport) tests
npm run build      # bundle to dist/ (CLI + library)
npm run dev -- --config bastion.config.json

Script	Does
build	Bundle CLI + library with tsup.
dev	Run the CLI from source with tsx.
typecheck	tsc --noEmit (strict).
lint / lint:fix	ESLint (flat config).
format / format:check	Prettier.
test / test:watch	Vitest.
check	Everything above, as one gate.

Roadmap

Version	Theme	Highlights
v0.1 ✅	Reliability	Aggregating proxy, auto-reconnect, bastion__status / __reconnect.
v0.2 ✅	Runtime security	Tool-definition pinning (rug-pull detection), poisoning inspection, shadowing detection.
v0.3 ✅	Audit & compliance	Pluggable audit sinks (console / file / webhook), NIST AI RMF / OWASP LLM Top 10 mapping.
v1.2 ✅	Delegation	Keyless checks on delegate tokens: audience, subject mixing, expiry, actor chains, signatures.
v1.4	Authorization & evidence	Per-caller runtime authorization; HTTPS-only, retrying, per-sink-redacted audit sinks; signed checkpoints and audit verify.

Both stdio and Streamable HTTP transports are supported (see Transports).

Contributing

Contributions are very welcome — this project is built to be community-owned. Please read CONTRIBUTING.md for the dev setup, project layout, and PR workflow, and our Code of Conduct.

In short: open an issue for non-trivial changes, keep PRs focused with tests, and make sure npm run check passes (CI runs it on Node 18/20/22). Good first areas: additional client setup recipes, more upstream test fixtures, and Streamable HTTP transport support.

Security

mcp-bastion is security-adjacent software, so we hold it to a high bar. Please report vulnerabilities privately — do not open a public issue. See SECURITY.md for the disclosure process.

License

Apache-2.0 © Gowthaman Arumugam and mcp-bastion contributors