Celeste CLI v1.9.3 — Graph-Powered Code Intelligence

Celeste CLI · v1.9.301 / 13

Open Source · MIT · Go 1.26+

CELESTE CLI

Graph-Powered Code Intelligence for AI Development

v1.9.3 Go 1.26+ MIT 54 MB binary Zero runtime deps

Presenterwhykusanagi

Repositorygithub.com/whykusanagi/celeste-cli

Companionceleste-for-claude · celeste-stopwords

Character artwork by ikawasa23 · x.com/ikawasa23

The Product02 / 13

Meet Celeste

celeste · ~/Development/grafana

          ✨ Celeste CLI
          Grok · grok-4-1-fast ✓ · Press Ctrl+C to exit
        
find the auth session handling code
1. loadUser — features/profile/state/actions.ts:37 [14% match, edges=14]
2. revokeUserSession — features/profile/api.ts:31 [13% match, edges=7]
3. loadSessions — features/profile/api.ts:27 [14% match, edges=7]

          ✓ code_search done 0.1s
        

          ◆ tokens: 8.4K / 2.0M ░░░░░░░░░░ 0%  │  compact: 0  │  turn: 1
        

          ❯ Type a message or 'help'...
        
Skills: enabled (45 loaded)
Backend: grok | Model: grok-4-1-fast
Last completed: code_search
Hint: type `skills` to browse available tool functions.

          ● Ready (10.7s · ↑7.9k ↓184)
        

An Agentic AI Coding Assistant

Celeste is a standalone Go binary (54MB, zero deps) that combines a premium corruption-aesthetic TUI with graph-based code intelligence. She indexes your codebase into a persistent code graph, then serves semantic search, structural code review, and dependency analysis — locally, offline, no embeddings needed.

Built for Claude Code & Cursor

Via MCP, Celeste exposes her code graph directly to Claude Code, Cursor, or any MCP client. Five direct tools (celeste_index, celeste_code_search, celeste_code_review, celeste_code_graph, celeste_code_symbols) return verbatim results with no chat-LLM round-trip.

A Companion, Not a Replacement

Celeste doesn't replace your LLM — she augments it. Claude stays in control; Celeste provides the graph intelligence. Three runtime modes: interactive Chat, autonomous Agent, and Orchestrator with built-in debate between two models.

The Problem03 / 13

Code Search Is Broken

What grep gives you

Flat text matches with no call-graph context. "query" matches JQueryStatic as #2 before any real database code.

No way to distinguish tests from implementation. mockHandler ranks the same as handleRequest.

Ubiquitous tokens ("get", "set", "error") appear in >30% of symbols, polluting every result set.

Library name pollution: jQuery decomposes into ["j","query","static"]. Framework nouns become misleading shingles.

Can't answer "which functions call which?" — only character-level pattern matching, no structural awareness.

What Celeste v1.9.3 gives you

MinHash + BM25 fusion — 128-slot semantic fingerprint over camelCase/snake_case tokens, types, body identifiers, and doc keywords.

Structural rerank — edge density, kind boosts, and path filter automatically demote test/mock code.

Stopwords v1.0.0 — statistically derived from 32 OSS repos. Filters noise before MinHash. JQueryStatic is gone.

Call-graph edges — tree-sitter CGo parses 10 languages. Understand which functions actually call which.

No embeddings needed — MinHash runs at parse time. Zero API calls, zero vectors stored, fully offline capable.

grafana · 77,420 symbols · 5 queries · top 10 per query no embeddings · no vector DB · offline

Code Graph04 / 13

The Code Graph

Query

"find the auth session handling code"

◆ loadUser ranks #1 — it's a hub node with 12 edges and 4 callers
◆ Cross-cluster call to QueryBuilder reveals the auth → database relationship
◆ Test nodes (mockHandler, assertEq) automatically demoted in ranking

grafana · 57,497 symbols 13,543 call edges MinHash + BM25 fusion Go · Python · Rust · TS · Java · C · C++ · Ruby graph snapshots · change impact

Search Quality05 / 13

Precision That Ships

22/50 → 35/50

relevant results across 5 benchmark queries (top 10 per query)

+59%

authentication session token validate Find the login code

2/10 → 8/10

✗

#1 result was JQueryStatic — jQuery splits into tokens "j"+"query"+"static", false-matching the search term

✓

loadUser loadSessions revokeUserSession — real session API. BM25 IDF broke the tie.

http request handler middleware Find HTTP handlers

0/10 → 6/10

✗

All 10 results from mocks/server/handlers/ — zero production code in top 10

✓

Real handlers rank first. Mocks auto-demoted via path-based test detection

database connection pool query Find database code

4/10 → 7/10

✗

JQueryStatic at rank #2 — library name decomposition polluted results

✓

All data-query types: ZipkinQuery TempoDataQuery QueriesTab

splitCamelCase fix

Library names like jQuery no longer decompose into misleading search tokens

BM25 + Stopwords

IDF scoring surfaces rare, meaningful tokens. 500 stop words filter noise before MinHash.

Path-based test filter

Files in mock/ test/ *_test.* automatically demoted in ranking

Search Pipeline06 / 13

How Search Works

Your query enters two parallel paths. MinHash compares token fingerprints for structural similarity. BM25 scores how rare and relevant each token is. The results merge via Reciprocal Rank Fusion, then get reranked by call-graph structure.

MinHash Path

Splits identifiers by camelCase/snake_case, removes stop words, then hashes each token set into a 128-slot fingerprint. LSH band tables make lookup 20x faster than brute-force comparison.

BM25 Path

Scores each token by how rare it is across the codebase (IDF). "session" in a query is worth more than "get" because it appears in fewer symbols. Stored in a SQLite inverted index.

Fusion + Rerank

RRF merges both ranked lists without needing to normalize scores. Then structural reranking boosts symbols with more call-graph edges and demotes test/mock paths.

MCP Tools07 / 13

No Middleman

5 Direct MCP Tools

celeste_index celeste_code_search celeste_code_review celeste_code_graph celeste_code_symbols

Streaming Progress

Reindex notifications stream in real-time. Claude Code sees "Indexed 4,231 symbols..." as it happens.

No LLM Round-trip

Direct call returns raw codegraph results. No max_tokens ceiling. No lossy summarization.

8 Total MCP Tools

5 direct codegraph + 3 persona tools. Served by celeste serve on localhost.

.grimoire · Project Memory08 / 13

Project Memory

# .grimoire — auto-initialized on first run, git-stamped

project: celeste-stopwords

version: 0.1.0

git_sha: 39cd8fe

indexed_at: 2026-04-13T09:15:22Z

@include: ./docs/architecture.grimoire

memories:

- "Pipeline uses celeste codegraph for tokenization"

- "Acceptance: <500 stop words, >30% FP reduction"

- "Corpus: 32 OSS repos, Apache-2.0 or MIT only"

hooks:

pre_tool: ["go vet ./..."]

post_write: ["go build ./..."]

⚠ grimoire is 3 commits behind HEAD — run celeste init to refresh

Memories

Persistent project facts injected into every LLM system prompt. Survive session resets. Git-stamped for provenance.

Sessions

JSONL auto-save with resume. Browse and reload any past conversation. Full message history with timestamps.

Checkpoints

File-level undo/revert via snapshot backup on every write. /revert rolls back any agent change.

@include Composition

Compose multiple .grimoire fragments. Monorepo teams split context by subsystem and merge at runtime.

Staleness tracking warns when grimoire is older than recent commits. Auto-reinit on celeste init.

Runtime Modes09 / 13

Four Ways to Work

Chat

celeste chat

Interactive with auto-looping tool calls. 50-turn safety cap.

$ celeste chat

Celeste v1.9.3 · codegraph loaded

You: find database pool code

[code_search] → ConnectionPool...

Agent

celeste agent --goal "..."

Autonomous multi-turn with planning, file I/O, checkpointing.

$ celeste agent --goal "fix retry"

[read_file] retry.go

[patch_file] backoff fix applied

[checkpoint] saved

Subagent DAG

NEW

spawn_agent task_id: depends_on:

Element-named agents (地火水光闇風) with auto-detected DAG dependencies.

✓ 〔火 hi〕 ch1.txt 15.6s

✓ 〔水 mizu〕 ch2.txt 31.7s

✓ 〔地 chi〕 ch3.txt 48.0s

DAG: ch2→ch1, ch3→ch1,ch2

Orchestrator

/orchestrate <goal>

Agent + reviewer model debate loop. Production-grade output.

You: /orchestrate refactor auth

[agent] drafts implementation

[reviewer] finds edge case

consensus: 2 rounds

Subagent DAG auto-detects task dependencies from goal text, sequences execution, and recovers partial results on failure. Orchestrator adds a debate/review loop — no other open-source CLI offers either.

Tools & Providers10 / 13

Built-In Power

45+ Built-in Tools · 7 Categories

⚙

Dev Tools

15+

            bash shell execution · read_file write_file patch_file

            search_files regex · list_dir · git_status git_log

            permissions · spawn_agent · checkpoint

◆

Code Intelligence

6

            code_search MinHash+BM25 semantic

            code_review structural 6-category

            code_graph callers/callees · code_symbols list

            extended_thinking · code_impact

⚡

Direct MCP

5

            celeste_index rebuild graph

            celeste_code_search verbatim results

            celeste_code_review no LLM round-trip

            celeste_code_graph · celeste_code_symbols

🌐

Web

8

            web_search · web_fetch scrape pages

            weather · currency conversion

            twitch_stream · youtube · qr_code · image_gen

🔒

Crypto & Security

7

            ipfs_pin · alchemy blockchain query

            wallet_monitor balance tracking

            hash_gen · uuid · password_gen · base64

🎵

Audio / TTS

NEW

            generate_speech ElevenLabs TTS

            sound_effect · mix ffmpeg multi-track

            audio_render project timeline · batch · Gantt viz

📋

Productivity

4

            reminders scheduled alerts · notes persistent scratch · todo task tracking · all stored locally in ~/.celeste
          

7 LLM Providers

★

Grok / xAI

DEFAULT · 2M context · grok-4-1-fast

✓

OpenAI

gpt-4.1-mini / gpt-4.1

✓

Anthropic

native SDK · prompt caching

✓

Google Gemini

gemini-2.5-flash · free tier

✓

Venice.ai

NSFW · image generation

✓

Vertex AI

enterprise · GCP project

⚠

OpenRouter

multi-provider · parallel fn calls

✓

        Zero runtime deps · 54 MB binary · offline capable
      

Architecture11 / 13

Under the Hood

Competitive Positioning12 / 13

Why Celeste

Feature	Celeste v1.9.3	OpenClaw	gptme	Aider
Language / Runtime	Go (compiled)	Node.js	Python	Python
Deploy size	54 MB binary	~393 MB + node	~150 MB + py	~80 MB + py
Code graph	MinHash + BM25 + tree-sitter (10 langs)	embeddings only	none	none
Semantic search	no embeddings, pure MinHash	vector DB required	none	none
Code review	structural (6 categories)	LLM only	none	none
MCP server	8 tools, 5 direct codegraph	partial	none	none
Runtime modes	Chat · Agent · DAG Subagents · Orchestrator	Chat only	Chat + Agent	Chat only
Multi-agent	DAG dependencies · element-named	none	none	none
Personality	Celeste (lore, grimoire)	none	none	none
Embeddings needed	✓ none required	required	n/a	n/a

The only open-source AI coding tool that combines graph-based code intelligence (10 languages), DAG-orchestrated subagents, and a companion personality — no embeddings, no runtime deps, single binary.

Get Started13 / 13

Open Source · MIT Licensed

GET STARTED

# install

go install github.com/whykusanagi/celeste-cli/cmd/celeste@v1.9.3

# first run

celeste config --set-key $XAI_KEY

celeste chat

↗ github.com/whykusanagi/celeste-cli

↗ github.com/whykusanagi/celeste-for-claude

↗ github.com/whykusanagi/celeste-stopwords · CC BY 4.0

Built bywhykusanagi

LicenseMIT · open source

Characterikawasa23

Versionv1.9.3 · Go 1.26+