Skip to content

v2.3.4

Choose a tag to compare

@jeffyaw jeffyaw released this 20 Sep 10:19
· 47 commits to main since this release
v2.3.4
6acf810

Fixed

  • aws_iam_simulate answered per action when it promises per resource, and reported resources as denied that AWS allows. On 2026-07-30 IAM changed SimulatePrincipalPolicy server-side, so every CLI version is affected: it now returns ONE EvaluationResult per action, whose top-level EvalDecision is the most restrictive decision across every resource and whose EvalResourceName is an ARN TEMPLATE -- literally arn:${Partition}:s3:::${BucketName}/${KeyName}. The per-resource answers live only in ResourceSpecificResults, which this tool dropped. So "can this role read these three buckets" came back as one row that named none of them and said deny for all three because one was denied: safe and unreadable, and the opposite of the tool's job -- it is meant to stop a model attempting work that will 403, and instead it stopped the model attempting work that would have succeeded, sending the user hunting for permissions they already have. Rows now come from ResourceSpecificResults, one per resource, each with that resource's own decision, matched statement ids and missing context keys, so summary counts (action, resource) pairs again -- which is what the description and the README have promised all along. No field is added or removed. AWS's reference says the older one-result-per-resource shape carried the same aggregate decision on every result, so reading ResourceSpecificResults is right whichever shape AWS serves: one code path, no version sniffing. An action AWS does not break down per resource falls back to its own top-level result, so no action can drop out of results because AWS sent a shape we had not seen. Three corollaries, each a wrong answer on its own: a call that names no resources reports resource * again rather than the ARN template (that is the commonest call shape, "can I do this at all?", and AWS applies ['*'] server-side); missingContextValues survives that call, because the API reports those keys on the TOP-LEVEL result for a * simulation and puts them per resource only when the call named resources -- only a * entry inherits them, since the top-level list is a union across resources and copying it onto a specific ARN would blame that ARN for another's keys; and a decision outside allowed / explicitDeny / implicitDeny now reads unknown, as the README always said, where before only a MISSING decision landed there and every other string, including "" and a miscased "Allowed", passed through verbatim and was counted in denied. organizationsDecision and permissionsBoundaryDecision are reported by AWS per ACTION, so a resource row that was allowed now reads "allowed" instead of the action's aggregate "denied" -- an SCP or boundary deny cannot coexist with an allow -- while a row that is not allowed still carries the action's value, which the description now says rather than leaving the caller to guess. Every iam_simulate fixture the suite had was pre-change shape, which is why nothing caught this; the four new ones were transcribed from captures the real aws-cli 2.34.3 rendered from the documented XML through a loopback stub, and the 2-actions-by-3-resources call now yields 4 rows with no ${ in any resource.
  • aws_iam_simulate rejects an STS session ARN up front, with the lookup that fixes it. aws_whoami reports arn:aws:sts::<account>:assumed-role/<role>/<session> for every SSO session, which makes it the ARN a model reaches for first, but the simulator takes a user, group or role ARN only -- so the first call from an SSO session went to IAM and came back with nothing to act on. The handler now refuses it before anything spawns and names the iam get-role lookup for the role it parsed out. It deliberately does not rebuild the role ARN, because the session ARN drops the role's path and SSO roles live under aws-reserved/sso.amazonaws.com/. The description and the README's rows also state what the simulator cannot see: SCP statements never appear in matchedStatementIds, keys only an SCP references are never reported missing (so a region-locking SCP needs aws:RequestedRegion passed in contextEntries yourself), and "allowed" is necessary but not sufficient -- RCPs, the target resource's own policy, session policies and VPC endpoint policies are not evaluated. The pagination wording now matches what the CLI does: aws iam simulate-principal-policy follows IAM's IsTruncated/Marker itself and prints the merged pages, so hasMore is false and marker null on every first call, and only a call that resumed from marker can carry either. The fake was emitting a truncated FIRST page the real CLI never produces -- the same fake-vs-real gap that hid the aws_logs_tail bug -- and now models the CLI both ways. The response shape is unchanged.
  • aws_docs_read could not open the index.html pages aws_docs_search hands out. docs.aws.amazon.com answers every <path>/index.html with a 301 to <path>/, and the post-redirect check added in 2.0.0 re-used the input pattern, which demands a .html suffix. So the canonical form of a URL the site itself returns came back as "redirected to '/', which is not an 'https://docs.aws.amazon.com/...html' page" -- an error worded like a blocked off-site redirect. In a live census of eight queries, 126 of 800 search hits were index.html URLs, among them the AWS CLI's own per-service command indexes (cli/latest/reference/s3api/index.html), which are the most relevant family of pages there is for a server that wraps the CLI. The landing URL is now judged three ways instead of one, and the 2.0.0 hardening is kept whole: it must still be https, on docs.aws.amazon.com, on the default port, with no credentials, and an off-site redirect is refused exactly as before.
  • A guessed page name now says it does not exist, instead of coming back as another page's content. The site does not 404 a missing page inside a guide that exists: it 302s to the guide's landing page, and a missing CLI command to that service's command index. So the redirect fix on its own would have answered a guessed command name with the whole s3api index as the content of a command nobody can run -- which is what accepting every same-host redirect, the behavior an origin-only check gives you, amounts to. A PAGE request that lands on some other directory is now ok: false with "does not exist" and the landing URL, so a model that guessed learns that it guessed. Every shape here was re-verified against the live site on 2026-09-19, and the fixtures are trimmed captures of those responses rather than invented HTML.
  • A docs fetch blocked by a corporate gateway blamed AWS's backend for it. Node's fetch rejects with a bare TypeError: fetch failed and puts the real verdict in err.cause, which no production file here read -- so behind a TLS-inspecting gateway or a private CA (the ordinary shape of a corporate network, and these are the only two tools in this server that reach the internet without the aws CLI) both docs tools reported "fetch failed" and nothing else, and aws_docs_search went on to say the undocumented backend at proxy.search.docs.aws.com "may have changed or be unreachable" about a request that never left the machine. The unit test meant to cover this threw new Error("ECONNREFUSED"), a shape fetch never produces, so the suite stayed green over it. All four fetch branches now append the cause -- read defensively, a non-empty string code only so a numeric errno cannot print as a bare "0", with any userinfo@ redacted out of both halves -- to the LAST @ of the authority and with no : required, because WHATWG userinfo runs to the final @ (new URL("http://svc:Pa@ss@h:3128") gives password Pa%40ss on 22.22.2, so an unencoded @ in a password is a legitimate value) and because http://<token>@proxy is auth too; ? and # end the authority like / does, so a docs URL whose query or fragment holds an @ is untouched -- and name the remedy where the verdict has one: an untrusted certificate chain asks for NODE_EXTRA_CA_CERTS or NODE_USE_SYSTEM_CA=1 (Node 22.19+) and says both are read at process start, so they belong in the MCP config's env block rather than your shell; an expired certificate or a hostname mismatch deliberately gets no CA advice, because a CA fixes neither; HTTPS_PROXY set without NODE_USE_ENV_PROXY=1 (Node 22.21+) gets the fact that Node's fetch ignores it, which is why the aws CLI -- which reads HTTPS_PROXY itself -- can keep working while these two tools cannot, and that hint also reaches the 30s timeout branch, because on a proxy-only network the likelier symptom is nothing answering at all; and an uncoded transport failure under oam, which reports no code for one, gets the certificate remedy plus a note that an unreachable proxy looks identical from there. Every shape was reproduced locally on 2026-09-19 rather than read out of documentation: a self-signed HTTPS server on a loopback port (Node 22.22.2 gives DEPTH_ZERO_SELF_SIGNED_CERT, oam 0.16.2 gives no code at all, and NODE_EXTRA_CA_CERTS fixes both), a dead port and an unresolvable host, and HTTPS_PROXY pointed at a dead port with the opt-in present, absent, set to 0, and via NODE_OPTIONS.
  • aws_call no longer blames you for the params you did supply. aws_call with service: 's3api', operation: 'get-object', params: {Bucket, Key} exited 252 with the following arguments are required: --bucket, --key -- naming the two values that were in params -- and no suggestion at all, so a model reads it as its own mistake and retries with the same params. Same for bedrock-runtime invoke-model, bedrock-agentcore invoke-agent-runtime, lambda invoke, logs tail, s3 cp/ls/sync and configure get. The cause is not the parameters: the AWS CLI never registers --cli-input-json on an operation whose argument table carries an outfile (awscli/customizations/cliinput.py _add_cli_input_argument: if 'outfile' not in argument_table:, unchanged from 2.0.0 to v2 HEAD), nor on its hand-written BasicCommands, and that flag is the only way aws_call passes params -- so those commands are unreachable through this tool whatever you send, and the required-arguments list is argparse describing flags it wanted, not members it was missing. The error now says so and names the way out: aws_lambda_invoke for Lambda, aws_logs_tail for logs tail, bedrock-runtime converse for text inference, aws s3 cp s3://BUCKET/KEY - in a shell for an object body, otherwise the shell with explicit flags. A call that sent NO params and hit the same message gets the other branch -- pass the members in params by API member name, with the rule rather than a guessed name (--bucket -> Bucket, but --model-id -> modelId on bedrock-runtime), because PascalCase is wrong for ECS, EKS, ECR, Logs, Step Functions, API Gateway, Batch and Bedrock. Detection is exit 252 plus argparse's own wording plus whether params was sent, and that third input is load-bearing: s3api head-object with no params emits BYTE-IDENTICAL stderr to get-object with params. It cannot fire on a call that worked, and it stays silent rather than guessing. No field moves -- errorKind stays nonzero_exit and rawBody is unchanged -- and the hint rides in suggestion and is embedded in error too, the convention runAwsCall already uses, because toMcpResult renders only error. The description also stops sending waiters to the shell: aws ec2 wait DOES register --cli-input-json, and s3api wait object-exists with {Bucket, Key} went on to the endpoint on both CLIs tested. Verified against real aws-cli 2.34.3 and real 2.22.0, whose stderr differs structurally (2.22.0 prints its usage block first, then aws.exe: error:), and in every one of the six error formats 2.34.3 offers -- seventeen rows through dist, each failing at argparse or a dead loopback port.
  • aws_multi_region refused the commercial partition its own description advertises. The regions cap was 32 and the aws partition has 34 regions (botocore's partitions.json, already shipped inside the 2.34.3 CLI on this machine, lists 34 once aws-global is dropped), so an account with every opt-in region enabled that piped its ec2 describe-regions output straight in got Too many regions: 34 requested, max 32 and nothing else -- the exact use the description promises ("describe-instances across all our regions"). All 34 names pass the region pattern, so the refusal was ours alone. The cap is now 64: 64 rather than 34 because an exact count goes stale at the next region launch, and 2x headroom costs nothing here, since the response is bounded separately by the unchanged 5 MB aggregate cap and parallelism by the unchanged concurrency limit of 32. Both numbers are interpolated into the tool description and the regions field description, so what the model is told cannot drift from what the schema and the handler enforce. Nothing on the stable list moves: no input is removed or reshaped, every previously valid call returns an identical result, and the cap was never documented as a blast-radius bound (that is destructiveHint: true, unchanged). Verified through dist against the loopback stub with both real CLIs: all 34 regions return okCount 34 on 2.34.3, and on 2.22.0 too -- it lists only 30 of them but resolves endpoints for the rest by pattern, so the cap was purely server-side.
  • An operation newer than your aws CLI looked like your typo, not an upgrade. The README promises new AWS operations are reachable "the moment your local aws CLI knows them"; when it does not, every tool that shells out returned the CLI's own argparse rejection with no suggestion at all, and nothing said the fix might be an upgrade. Measured on aws-cli 2.34.3 against Batch's September bulk cancel (CLI 2.36.44): aws_call {service: "batch", operation: "cancel-jobs"} came back argument operation: Found invalid choice 'cancel-jobs', exit 252, suggestion absent -- which reads as "you misspelled it". parseAwsError now recognizes that rejection and names both causes, spelling first, because a model's mistyped operation is at least as likely as an old CLI. It covers all three of argparse's dests and the noun follows the caller's vocabulary rather than argparse's: command is the SERVICE they passed, operation is the operation, subcommand is a waiter or an s3 / configure subcommand. The upgrade half names the real constraint -- update.py's _SUPPORTED_SOURCES is ('exe', 'script-exe', 'update-exe'), so aws update (2.36.0+) works for installer and install-script installs and raises UpdateError for anything else, which is why the installer and package managers are named too. The pattern is unanchored, and that is load-bearing: the phrase moves depending on the CLI and the user's error format. Verified on real 2.34.3 in its default enhanced format, in legacy (also what a --profile the CLI cannot find produces), and in the json format where the phrase sits inside Message behind literal \n escapes; and on real 2.22.0, which predates error formats, prints its usage block first and names no choice at all on the service form. It sits after the standard-wrapper branch, so a genuine service error that merely contains the phrase keeps its code-based remedy. The name it quotes is vetted as a command token first, because argparse's "invalid choice" is not always a name the caller typed: when operation gives a subcommand GROUP (ec2 wait), that group's parser registers no --cli-input-json -- aws ec2 wait help lists the flag 0 times on 2.34.3, the leaf ec2 wait instance-running help 3 times -- so argparse reads the params JSON aws_call passes through that flag as the missing positional. Quoting that back told a model the CLI had "no subcommand named {"InstanceIds":["i-1"]}", to check that string's spelling, and to run aws update, which fixes nothing; it also copied params into a field documented as a one-line remedy. aws_call answers that shape itself now, since only the tool side knows params were sent: aws ec2 wait is a subcommand group, name the waiter in operation. On 2.22.0, whose wording names no choice at all, the shape is indistinguishable from a mistyped waiter and keeps the spelling remedy -- with nothing echoed either way. No field is added and errorKind stays nonzero_exit: suggestion is already in the envelope and already documented as best-effort.
  • A TLS-inspecting gateway or an unreachable proxy on the CLI path no longer reads like an AWS outage. These are the CLI half of what the docs tools' fetch failures now explain, and both came back with no remedy. Behind a TLS-inspecting gateway botocore raises SSLError (SSL validation failed for <url> ...) and an unreachable proxy raises ProxyConnectionError (Failed to connect to proxy URL: ...); parseAwsError now says where the fix has to live -- AWS_CA_BUNDLE or the profile's ca_bundle for the first, HTTPS_PROXY / NO_PROXY for the second, and in this server's MCP-config env block, because the aws subprocess inherits this server's environment and botocore takes the proxy URL from the environment only (endpoint.py's _get_proxies is get_environ_proxies). The TLS remedy does not offer to turn verification off. The proxy sentence was reproduced on 2.34.3 against a dead loopback proxy; the TLS sentence is botocore's own fmt string read from the CLI's bundled copy, because two attempts to force it against a self-signed loopback server on this box came back as a read timeout instead. botocore masks the credentials in a proxy URL before raising, so the value quoted back is already ***:***@.
  • An unhandled promise rejection no longer takes the server down. An aws_script whose code calls aws.call(...) without awaiting it returned success to the model and then killed the server: the script returns first, the un-awaited bridge call rejects a beat later, and with no process-level handler Node's default (--unhandled-rejections=throw) exits 1, so all 28 tools disappear from the session right after a call that reported success, until the host restarts the server. Measured over real MCP stdio on Node 22.22.2, both ways the call can fail -- a service name that fails validation locally, and a real aws CLI failure (an unknown --profile against an empty config, a loopback endpoint and a dead proxy). In both, tools/call aws_script answered {"result":"returned"} and the next tools/list got no reply at all. The entry point now installs a handler beside 2.3.3's hardenWindowsExeSearch() call and keeps serving: one stderr line with the reason's message and stack. It is installed in the entry-point block rather than at module load, so importing this module never changes a host process's rejection semantics -- including the test runner's, which must keep failing on one. The reason is classified with node:util's types.isNativeError as well as instanceof Error, because the motivating reason fails the instanceof test: the aws_script bridge re-throws bridge failures as the sandbox realm's own Error, a different constructor from the server's, and an instanceof-only check would print "Error: " with no stack for exactly the case the handler exists for. A non-error reason goes through String(reason) and is never inspected, the same rule errorToMcpResult follows, so a rejected AwsCallResult-shaped value cannot dump rawStdout / rawStderr into operator stderr. This is process-level rather than a fix in the bridge because the bridge can only mark its OWN promises handled -- a script's bare Promise.reject() and sso.ts's void promise.finally(...) login dedupe would still be fatal. uncaughtException is deliberately left alone: a synchronous throw that unwound to the top left the code that threw it part-way through its work, and a server answering from torn state is worse than one the host restarts. Verified on the same two runs, plus under oam 0.16.1 and 0.16.2 and through bin/aws-mcp.mjs -- 28 tools still listed and a following tool call still answered.
  • A Registry-driven install had no way to offer a profile, so it wrote none. server.json declared no environment variables, so an installer reading the MCP Registry entry had nothing to prompt for and wrote a config with no env block. The server then resolves --profile default on the first tool call (session.ts's getProfile falls through AWS_PROFILE, AWS_DEFAULT_PROFILE, then default), and on a machine with no [default] section that is a no_creds error on the user's first question -- for a server whose whole pitch is that it runs on the profiles you already have. packages[0].environmentVariables now declares AWS_PROFILE and AWS_REGION, both optional, each description stating the real fallback chain rather than a generic blurb -- including that a profile's own region setting is never read, because every call carries --region <resolved> on its argv, so us-east-1 is what an unset region means. Neither carries a default, deliberately both ways: a pre-filled default profile recreates the very no_creds failure this is meant to avoid, and a pre-filled region would outrank a user's own AWS_DEFAULT_REGION. An installer that writes an EMPTY string for a variable the user skipped is safe, since getProfile and getRegion use ||. Nothing is marked secret, because neither value is one, and only these two are declared: the launcher knobs (AWS_MCP_RUNTIME, OAM_BIN) and the legacy spellings stay undeclared, because an installer may prompt for every declared variable and a runtime override is not a thing to ask a new user. Verified on the block as committed with mcp-publisher validate 1.7.9 against registry.modelcontextprotocol.io, with ajv against the 2025-12-11 schema, and through release.sh's own version-sync jq line, which leaves the array intact and both version fields correct.
  • The README's only companion-config example could not install. The single example under "Optional companion" was uvx awslabs.lambda-mcp-server@latest, and all ten releases of that name are yanked on PyPI ("superceeded by awslabs.lambda-tool-mcp-server"), so uv refuses to resolve it -- "Because all versions of awslabs-lambda-mcp-server were yanked ... your requirements are unsatisfiable", reproduced on uv 0.11.7 -- and the MCP entry it was pasted into never started. It was also selling a Python server for "typed lambda_invoke", which aws_lambda_invoke has done natively since 2.2.0, and the renamed package is a different product: a bridge that exposes chosen functions as tools. The example is now awslabs.bedrock-kb-retrieval-mcp-server, picked because its agentic retrieval reaches something aws_call never can -- it calls AgenticRetrieveStream, which awscli/customizations/removals.py deletes from the CLI's command table -- with the note that plain bedrock-agent-runtime retrieve is an ordinary aws_call operation, and that the server lists only knowledge bases tagged mcp-multirag-kb=true by default.

Changed

  • A docs page with no content of its own now fails with the reason instead of coming back as an empty success. This is a behavior change, not merely a fix, and it is the one response in this release that turns from ok: true into ok: false. Guide landing URLs are <meta http-equiv="refresh"> stubs, and the JS SDK v3 and Swift API references are application shells assembled in the browser; converted, they are 0 characters and [Skip to main content](#main). A read whose conversion is under 200 characters of text AND whose HTML carries one of those markers is now ok: false naming the reason -- for a stub, with the absolute URL of the page it forwards to. Both gates have to fire, so a genuinely short REAL page is served as-is, and nothing thin is cached, so a retry after the page is fixed is not served the empty conversion for five minutes. In practice this replaces an empty success rather than taking content away from anyone: every thin shape this work found live, on 2026-09-19 and again on 2026-09-20, is an index.html URL that the pre-existing post-redirect check refused anyway, and it is the redirect fix above that newly makes them reachable. No input, envelope field or annotation changes.
  • The README describes the server AWS runs today, and a comparison that is still true. Every claim was re-verified live on 2026-09-20 and the diffs were all in AWS's favour, so the comparison had drifted into flattering ourselves on stale facts. call_aws is gone -- AWS's tool reference lists eight tools without it, and the endpoint answers -32600 "The call_aws tool has been removed, use another one." -- so every AWS API call there is now a Python script in run_script: a difference in shape, not coverage. Since July 2026 it also takes OAuth through AWS Sign-in with nothing installed locally, so "Requires Python + uv" described only its SigV4 path, the one AWS recommends for terminal and IDE coding agents and the only one that switches profiles per call. Its Agent SOPs are now "skills". The bullet is now a parent plus one sub-bullet per connection path, because a single sentence misstates whichever path it does not describe, and it keeps an honest "nothing installed locally -> AWS MCP Server (OAuth)" row; what Node/npm-only buys is stated as the checkable thing it is -- a local server on your own CLI profiles, no uv to install and no proxy -- not as "no Python", which CLI v2 bundles anyway. Data-plane routing was wrong in three places: s3api get-object and bedrock-runtime invoke-model stream their body to a positional outfile and register no --cli-input-json, Lambda invokes have had their own tool since 2.2.0, and AWS Labs' "DynamoDB with type-marshalling" has eight data-modeling, validation, cost and codegen tools and no data-plane tool at all; dynamodb get-item and bedrock-runtime converse do work and are named instead. Resource types are 1,300+ rather than "hundreds" (AWS's supported-resources list holds exactly 1,300, last updated 2025-12-23), the "services AWS adds tomorrow" examples are September's with the CLI release that added each, and awslabs/mcp is described as AWS now describes it -- succeeded by the Agent Toolkit for AWS, still working and taking contributions -- with the Cloud Control API server's deprecation named and its migration guide's "no direct replacement" for get / list / create / update / delete cited. Two limits are stated where they bind: the 5 MB cap is this server's and is not what users hit first, since Claude Code warns at 10,000 tokens and by default saves any result over 25,000 to a file the model has to read back; and keep this server's config key as aws, because aws configure agent-toolkit (CLI 2.35.0+) registers AWS's hosted server under aws-mcp and returns ALREADY_CONFIGURED on that key's presence alone, without looking at what it runs. Requirements now answers the question every tool depends on: what the CLI is for (every tool that talks to AWS shells out to it -- all but aws_docs_*, aws_session_* and aws_list_profiles), the version this server is developed and tested against, what an older CLI does with a newer service, operation or parameter, and how to upgrade. Documentation only; no change to the published package's behavior.
  • Test-only: a tool description that passes the host's 2 KB text cap now fails the suite. Claude Code truncates every tool description it hands the model, silently, and nothing here was checking. Read out of the installed 2.1.272 binary: one helper is applied both to each tool description and to the server instructions; above 2,048 characters it cuts the text, appends a truncation marker, and logs the loss to the MCP debug log alone -- the server is not told and neither is the user. What goes first is whatever was written last, and the widest description in this server is aws_logs_query's at 1,981 bytes, 67 to spare, whose closing sentences are the two a caller most needs: that a timed-out or cancelled Insights query is NEVER stopped, and that aws_logs_tail is cheaper for plain recent lines. HOST_TEXT_CAP_BYTES lands in a new src/server-instructions.ts -- the module the server instructions block and its own tighter ceiling will join -- and two tests hold every description to it: one against allTools from the esbuild bundle that ships, naming every offender with how far over it is, and one in the existing tools/list loop in the integration suite, which is the byte count a host actually reads off the wire. Checked by lowering the cap to 1,900: both fail, naming aws_logs_query and 1,981 bytes. The gate measures UTF-8 bytes where the host compares UTF-16 code units, which is never fewer for the same text, so a description that passes is never cut; today exactly one description carries a non-ASCII character at all -- aws_assume_role's single em-dash -- and there the gate is 2 bytes stricter than the host.
  • Test-only: three new suites check the facts these fixes rest on against the real thing, not against a fixture. Each is opt-in, so a plain npm test reports them skipped. call.realcli.test.ts drives aws_call at the AWS CLI v2 installed on PATH under the AWS_MCP_REAL_CLI_TESTS=1 gate 2.3.3 established, and asserts the CLASSIFICATION of whatever the CLI printed rather than its bytes, so it survives a wording change that would break a byte-for-byte fixture -- all four cases die in argparse or in botocore's own validation, so there is no stub, no endpoint override and nothing sent. docs.live.test.ts runs the four cases the redirect and landing-page fixes turn around against docs.aws.amazon.com itself, under AWS_MCP_LIVE_DOCS=1: it needs outbound HTTPS and nothing else -- no credentials, no AWS API, no subprocess -- which is why it is .live. rather than the .integration. suffix this repo uses for suites that talk to a real account, and without it the day AWS starts 404ing a missing page the suite stays green and the "does not exist" message silently stops appearing. iam-simulate.integration.test.ts is the only one that watches AWS: read-only (SimulatePrincipalPolicy evaluates policies and mutates nothing), gated on BOTH AWS_MCP_LIVE_TESTS=1 and AWS_MCP_LIVE_IAM_PRINCIPAL_ARN, and in its own file because iam-simulate.test.ts has a file-level before() that routes EVERY spawn to the fake, so a live test placed there would quietly pass against it. Each of the two live files says to run that file alone, because AWS_MCP_LIVE_TESTS=1 npm test also fires a real Cloud Control create/delete.
  • Test-only: a schema-validation case could have spawned 40 real AWS calls once the region cap moved. script.test.ts's "region list above .max(32)" case built 40 regions and ran against the DEFAULT handlers with no fake shim, on the premise that validation rejects it before any spawn. At a cap of 64 that premise is false: 40 validates, and npm test would have spawned 40 real aws processes with the developer's own credentials against real region names. The count now derives from the exported cap instead of a literal, and that whole describe pins the test command at a nonexistent binary, so a future regression of the same shape fails with ENOENT instead of reaching AWS.

Security

  • The four local aws commands that print your credentials or your recorded traffic are refused before anything spawns. aws_call accepted service: 'configure', operation: 'export-credentials', and that command answers with the profile's resolved credentials -- measured on aws-cli 2.34.3 against a throwaway credentials file, the full SecretAccessKey, not the masked form configure list prints, and for a static-key profile a long-lived key no STS call would ever hand out. The output went straight into the model's context and the host's transcript. history show and history list replay what the CLI recorded locally, and configure get reads any config value by name. None of the four sends a request to AWS, so no IAM policy can limit them: scoping the credentials, which is this server's answer for every actual API call, does nothing here. They now return errorKind: "bad_input" naming the command, why it is refused, and the escape hatch -- your own terminal, which is where reading your own keys belongs. Everything else in configure still runs, including configure list. The guard sits in runAwsCall, so aws_script's aws.call and the fan-out tools are covered too. (configure get was already unreachable in practice -- it needs a positional varname and configure rejects --cli-input-json, so the call failed with ParamValidation before reading anything -- and is listed so that stays true if a positional path is ever added.)
  • Requirements now carries a dated AWS CLI security floor: 2.35.3 or newer clears every published AWS CLI v2 advisory. Nothing in this package changes -- the floor is not enforced and no version is rejected -- but it is worth stating where it binds, because v2 ships as an installer, so no dependency scanner will ever flag the CLI you share with everything else on the machine. None of the affected paths is reachable through this server: the affected commands (emr ssh/socks/put/get, codeartifact login, deploy register, iam create-virtual-mfa-device) all register no --cli-input-json, which is the only way aws_call passes parameters, and the remaining advisory is about the opt-in cli_history database, which this server never enables. The three advisories are linked from Requirements, with a pointer to the advisory list for newer ones.