The MCP server “io.github.Anandb71/arbor” provides a Graph-Native Intelligence Layer for codebases. Named “Arbor” in the excerpt, it focuses on understanding code using graph-native approaches, with an emphasis on identifying issues early (“Know what breaks before you break it.”).
Code context for people and agents.
Find a function’s callers. Trace the path through your code.
Know what breaks before you break it.
arbor retirements finds direct OpenAI, Anthropic and Google SDK calls in Python code that use a retired or retiring model, and says what has to change.
OpenAI shuts down the dated gpt-4o and gpt-4.1 snapshots on 14 January 2027. Shutdown list
Side-by-side: an agent navigating tokio with grep-and-read (47 tool calls, still searching) vs the same agent with arbor's code graph (4 graph calls + 1 read, done)
Simulated replay — the arbor commands and their output are real (tokio @ 178k LOC). Methodology: BENCHMARKS.md
v3.1.1 · A fix release. The arbor retirements data was read again on 28 October 2026 and now lists the bare gpt-5 and o3, which OpenAI's deprecations page gives 11 December 2026. npx @anandb71/arbor-cli runs on npm 12, and Arbor no longer takes your home folder as the project because ~/.arbor exists. Release notes
Why Arbor
Most AI coding tools treat code as text. Arbor builds a semantic dependency graph — functions, classes, and modules as nodes; calls, imports, and inheritance as edges — then answers execution-aware questions with deterministic precision:
Question
Arbor answer
If I change this symbol, what breaks?
Blast radius with depth, confidence, and risk level
Who calls this — directly and transitively?
Caller/callee traversal on the call graph
What's the shortest path between A and B?
A* path through real dependencies
Is this PR too risky to merge?
CI gate on blast-radius thresholds
No keyword guessing. No embedding hallucinations. One graph, every interface.
Where the graph is unsure, it says so — edges carry a confidence, and ambiguous resolutions are labelled rather than hidden. An honest unknown beats a confident wrong answer.
What's new in v3.1.1
A fix release. No breaking changes; as after every upgrade, a saved graph is
rebuilt once on first use.
Fix
What was wrong
Bare gpt-5 and o3
OpenAI's deprecations page started listing bare IDs next to the dated snapshots after 3.1.0 was built, so model="gpt-5" or model="o3" was not listed at all. Both are listed now (11 December 2026, to GPT-5.6), with the 13 other IDs the page added.
Data read again on 28 October 2026
Every provider page behind the bundled data was read and checked again. OpenAI's deprecations page lists the five dated gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini snapshots for 14 January 2027; for the bare aliases the report now cites OpenAI's API reference ("Deprecated. Announced shutdown date: 2027-01-14") and its models API. A call moving to claude-haiku-4-5-20251001 with temperature now gets a check line: Anthropic's API version 2026-10-21 rejects it on every model.
(no change) next to arguments that need reasoning off
For a call whose client Arbor could not resolve, temperature, tools or max_tokens went unchecked and "let the model reason" read (no change). They are checked now. When Arbor cannot read some of a call's arguments (**kwargs, extra_body) or a check line names one of them, each option says so instead of no change.
npx on npm 12
npm 12 skips install scripts, so the binary was never downloaded. The wrapper now downloads it on first run and checks it against the release's SHA256SUMS.
The home folder as the project
~/.arbor holds the global config (and, on Windows, the installed binary), so a folder under home with no nearer project marker resolved to the whole home folder, and arbor index . indexed all of it. The search now stops at the home folder.
Code graph: aliased imports and bare calls
After from m import f as g, a call to g() now reaches f (also for import { f as g } in TypeScript and JavaScript). In Python, JavaScript, TypeScript and Go, a bare call to a function in another file counts only when the calling file can see it (in Python, when it imports the name); other languages are unchanged.
When the terminal report of arbor retirements lists a call to a retired or
retiring model, it now ends with an offer of a paid readiness review. The link
carries three values from the run: how many calls are listed, how many are
already retired and the next shutdown date. Arbor sends nothing. The offer is not
printed with --json, --markdown, --annotations or --quiet, or when the CI
environment variable is set (GitHub Actions sets it), and the GitHub Action does
not show it. Release notes
v3.1.0 (arbor retirements and arbor usages)
Two commands for Python code that calls AI model APIs. Both run locally, with no
account, API key or telemetry.
New
What it does
arbor retirements
Lists the direct openai, anthropic and Google SDK calls that use a model its provider has retired or is retiring: the date, the replacement the provider lists, the arguments that change with it and the choices your repo makes first. OpenAI's deprecations page lists the dated gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini snapshots for shutdown on 14 January 2027; OpenAI's API reference and models API give the bare aliases the same date. More below
GitHub Action mode: retirements
The same check in CI: a job summary of the affected calls it finds, and warnings on the most urgent lines. It warns by default and fails the build only when you set retirements-fail-on, or when the check cannot run.
arbor usages and find_api_usages
The OpenAI, Anthropic and Gemini SDK calls Arbor finds, with their file and line, client routing (direct, Azure, custom base_url, Bedrock, Vertex) and model IDs, for people and agents. A value is confirmed only inside a small supported grammar; everything else carries a typed gap. What a report confirms
v3.0.3 (Rust calls through paths and macros, per-symbol diff, --base)
A fix release. Every item below was a wrong answer, not a missing feature, and
each ships with a regression test.
Fix
What was wrong
Rust calls resolve
crate::jobs::enqueue(), Type::new(), self.helper() and calls inside assert!/format! produced no edge, so heavily used Rust functions reported no callers. Paths now resolve through the module tree, and macro arguments are read as code.
Per-symbol arbor diff
Every symbol in a touched file counted as changed, so adding a function to a busy file reported that file's whole blast radius. New symbols now carry none; only modified ones do. Tests that call the change are listed as worth running, not counted as impact.
--base and --staged
arbor diff, check and summary saw only uncommitted edits unless ARBOR_DIFF_BASE was set. --base origin/main compares against the merge base, which is what a pull request shows; --staged checks just what is staged.
Languages stay apart
A TypeScript enqueue could be reported as a caller of a Rust enqueue. Resolution stays within a language family, and callers/callees list each same-named definition separately.
No stale graphs
After a branch switch, answers came from the old branch's graph when no remaining file was newer than it. The graph now records the commit it was built from and refreshes when HEAD moves.
Inheritance edges
class Middle(Base) produced no edge, so changing a base class showed zero blast radius. extends/implements edges are emitted and inherited methods stay reachable.
Call cycles rank as one
A closed ring of functions filled the top of every centrality ranking. Cycles are condensed before PageRank.
HTTP bridge hardening
arbor bridge --http accepted cross-origin and DNS-rebinding requests. It now checks Origin/Host and requires JSON bodies.
Quieter bridge
The bridge re-indexed ignored build directories in a busy loop.
Also new: arbor receipt. After each coding-agent turn it
explains, in plain English, what changed and what was touched that you didn't
ask for, with arbor receipt undo to put a turn back.
Cached graphs from 3.0.0 are rebuilt automatically on first use.
v3.0.0 — The Right Node (symbol resolution consults the importing file)
One fix, measured.
Symbol resolution consults the importing file. When a bare name matched
definitions in several modules, resolve_ref fell through to SameDir and
attached the edge to whichever definition sat in the caller's own directory —
not a dropped edge, a confidently misrouted one, stamped at 0.55 confidence.
GraphBuilder already kept a per-file import map, but only
apply_import_validation read it, and that scores an edge after one has been
chosen. It never saw the references going to the wrong node. Consulting it
between the same-file and same-directory checks keeps a local definition
shadowing an import, while letting a written import beat mere adjacency.
Resolution::ViaImport scores 0.93, above SameDir's 0.55.
Measured
A fixture of 260 modules across 10 layers, each layer defining the same 26
function names. Ground truth is derived from the generator's own edge list, so
the expected answer is exact rather than estimated.
True downstream
v2.6.0
v3.0.0
179
0
163
178
0
161
161
0
133
143
22
133
122
22
119
36
22
61
16
22
46
Previously flat at about 22 regardless of the real answer. Now it tracks. Risk
on the largest hub moves from LOW to CRITICAL.
Total edge count barely moves (1335 → 1334). That is the signature of
misrouting rather than loss: the edges were always there, pointing at the wrong
nodes.
Breaking
Resolution gains a ViaImport variant — an exhaustive match will not compile
Edges land on different nodes, so cached graphs, stored node ids, and
centrality baselines from 2.6.0 will differ
Call cycles are condensed before PageRank. Each strongly connected component is ranked once and that mass is shared across its members, so a closed ring does not fill the top of the ranking and a cycle that calls out keeps its members together. The number CentralityScores reports is still the v2.6.0 percentile, i / (n - 1).
Known and still open
Written down rather than left to be discovered:
Small targets now over-report (36 → 61, 16 → 46). Safer direction than
silence, but not yet correct.
Dynamic and reflective imports (importlib, __import__, import(),
eval(require(...))) are unresolvable by construction and are documented as
expected misses in the fixture rather than counted as defects.
Correctness, not speed. Each of these was silently wrong before.
Fix
Why it mattered
Colliding symbols are kept
SymbolTable used HashMap::insert, so a second handler, new, or process replaced the first. The loser had zero callers and was invisible to blast radius.
Resolution is deterministic
Same-directory locality was decided by iterating a HashMap. Rust seeds RandomState per process, so the same binary on the same input could build different edges between runs. Now asserted across eight fresh processes.
Edges carry confidence
A proven same-file call and a same-directory guess were identical evidence. Each edge now scores [0,1] by how it resolved.
Exported TS symbols indexed once
export_statement recursed into its children, then the generic loop recursed again — every exported symbol became two vertices sharing one node id. 133 phantom nodes on a 149-file app, 25% of the graph.
Method calls on untyped receivers resolve
obj.method() was dropped outright, leaving the graph nearly edgeless on TS/JS — and an empty graph reports a blast radius of zero, which reads as "safe" rather than "unknown".
Centrality is a percentile rank
Scores were divided by the graph maximum, so the top node was 1.0 by construction and a 0.6 threshold meant nothing consistent between repos. Adding one hub rescaled every other node.
Resolution is O(1), not O(refs × nodes × files)
Unresolvable references — stdlib and third-party calls, most call sites in real code — paid the worst case. Suffixes are now indexed.
New capability — concept search. Substring matching cannot find get_authenticated from login; they share no substring. Identifiers are now tokenized and expanded through curated concept clusters, and docstrings, signatures, and paths are indexed alongside names. Deterministic, offline, no model. Available on the library as ArborGraph::search_ranked (arbor query remains literal-substring for now).
New capability — hunk-level impact.changed_node_ids_for_ranges keeps only symbols whose lines actually changed, instead of every symbol in a touched file.
Measured on identical node sets, after the duplicate-extraction fix:
Codebase
Before
After
TypeScript (149 files)
172 edges
196 (+14%)
Rust (arbor-graph)
116 edges
167 (+44%)
Graph caches from earlier versions are invalidated — centrality now means something different, so a stale cache would be read wrong.
v2.5.0 — The Last Excuse (PageRank 23x, parallel indexing, warm-start centrality)
Git-diff-aware impact analysis via shared arbor-graph::compute_blast_radius
Pagination
offset / limit / hasMore on search_symbols and get_map
Benchmarks
Criterion suite + CI regression gate — see BENCHMARKS.md
Quickstart
bash
# Install (crates.io can lag the latest release: see docs/INSTALL.md)
cargo install arbor-graph-cli
# Index your project (one command)cd your-project && arbor setup
# Explore before you edit
arbor map . --exclude-test # ranked project skeleton (~1k tokens)
arbor refactor parse_file # blast radius of changing a symbol
arbor diff # impact of uncommitted git changes
arbor diff --base origin/main # impact of this branch, as its PR shows it# Wire up your AI agent
claude mcp add --transport stdio --scope project arbor -- arbor bridge
After every turn:arbor hook claude makes Claude Code show a receipt of what it changed, what it touched that you didn't ask for, and what to test. Receipts →
Agent workflow: call get_map first → search_symbols / get_file_graph to locate code → Read only the target file. Full MCP guide →
Find code that calls retiring AI models
OpenAI's deprecations page lists five dated GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini snapshots, such as gpt-4o-2024-08-06, for shutdown on 14 January 2027. It gives the bare gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini no row, but OpenAI's API reference lists each as "Deprecated. Announced shutdown date: 2027-01-14", and its models API returned that date when checked on 18 October 2026. Anthropic and Google retire models on their own dates. arbor retirements lists the direct openai, anthropic and Google SDK calls that use a retired or retiring model. For each one it shows the replacement the provider lists, the arguments that have to change with it, and the choices your repo makes first (such as whether gpt-6-luna should reason). Calls through LangChain, LiteLLM or other wrappers are not checked yet.
bash
npx @anandb71/arbor-cli retirements .
text
3 calls use models OpenAI shuts down on 14 Jan 2027 (88 days)
app/chat.py:12 gpt-4o (MODEL, line 8) → gpt-6-luna
also change: max_tokens → max_completion_tokens
decide: keep reasoning off (add reasoning_effort="none"; keep temperature) or let the model reason (remove temperature)
app/chat.py:21 gpt-4o-mini (line 22) → gpt-6-luna fallback of an environment or config lookup: confirm the deployed value
decide: keep reasoning off (add reasoning_effort="none"; check stop) or let the model reason (check stop)
check: stop and logit_bias are not supported on gpt-6-luna (400).
Runs locally. No account, no API key, no telemetry; no code leaves your machine.
Honest about confidence. A literal is confirmed only where the usage scan has no doubt about the call. os.getenv("MODEL", "gpt-4o") fallbacks, parameter defaults, class attributes a subclass may override and calls whose routing is not certain are marked confirm first with the reason, and fail the check only with --count-unconfirmed. Calls through Azure, OpenRouter or a proxy are listed apart (--include-routed), because the provider's dates may not apply there; a base_url on the provider's own host is a direct call. Models Arbor cannot read, and code it did not analyze, are counted, not hidden.
Arbor does not decide for you. When the replacement behaves differently by design (gpt-6-luna runs at medium reasoning effort by default, so it can reason, where gpt-4o did not; claude-sonnet-5-5 thinks by default, where the Sonnet models it replaces did not), the report shows both options and what each costs, and changes nothing until you choose.
CI ready.--json, --markdown, --annotations, --markdown-out <file> and --fail-within <days>. Exit code 1 means a confirmed call uses a retired model, or one retiring inside your window; the shutdown day itself counts as retired (UTC).
Scope today: Python with the openai, anthropic and Google GenAI SDKs. TypeScript and JavaScript are coming.
GitHub Action (warns by default; it fails the build only if you ask it to):
yaml
-uses:Anandb71/arbor@v3.1.1with:mode:retirementsretirements-fail-on:never# or: retired, or a number of days such as 90
One-hop graph traversal. Same-named symbols are listed per definition; narrow with jobs::enqueue, Type.method or src/jobs.rs:enqueue
arbor entry-points
HTTP handlers, main, jobs, webhooks
arbor file-graph <path>
Symbols + edges in one file
arbor inspect <sym>
Full symbol detail
arbor path <a> <b>
Shortest call-graph path
arbor refactor <sym>
Blast radius before refactoring
arbor usages
OpenAI, Anthropic and Gemini SDK calls in Python: client routing (direct, Azure, custom base_url, Bedrock, Vertex), model IDs and typed gaps; never prints credentials. --sdk openai|anthropic|google, --json. What a value with no gap confirms: docs/SDK_USAGES.md
arbor retirements
Direct SDK calls to AI models their provider has retired or is retiring (Python): date and days left, listed replacement, the argument changes and per-repo choices it needs. --fail-within <days>, --count-unconfirmed, --include-routed, --json, --markdown, --annotations (guide)
arbor diff
Git-change impact report, per symbol: new symbols carry no blast radius. --base <ref> compares against the merge base (what a PR shows), --staged only staged changes
arbor check
CI safety gate (--max-blast-radius N, --base <ref>)
arbor summary
Auto-generate PR description (--base <ref>)
arbor agent review
Autonomous PR architecture review
arbor agent onboard
Codebase onboarding guide
arbor agent guard
Real-time architectural safety gate
arbor bridge
MCP server (add --http for HTTP transport)
arbor watch
Live re-index on file changes
arbor receipt list / show / undo
Plain-English receipts of what each agent turn changed, and undo for a turn (guide)
arbor hook claude
Wire Arbor into Claude Code: directives, receipts after every turn
arbor gui
Native desktop UI
All query commands support --json. map additionally supports --tokens N, --focus "pattern", --focus-changed.
Visual tour
Installation
bash
# macOS / Linux: prebuilt binary from the latest GitHub release
curl -fsSL https://raw.githubusercontent.com/Anandb71/arbor/main/scripts/install.sh | bash
# Windows (PowerShell)
irm https://raw.githubusercontent.com/Anandb71/arbor/main/scripts/install.ps1 | iex
# Scoop (Windows)
scoop install https://raw.githubusercontent.com/Anandb71/arbor/main/packaging/scoop/arbor.json
# Rust / Cargo (crates.io can lag the latest release)
cargo install arbor-graph-cli
# npm wrapper (cross-platform)
npx @anandb71/arbor-cli
# Docker
docker pull ghcr.io/anandb71/arbor:latest
mode: retirements runs arbor retirements instead. It writes a job summary that lists each call it finds to a retiring model and annotates the 10 most urgent lines, GitHub's limit per step (guide).