cli-bridge-mcp

<!-- mcp-name: io.github.JoaoBerne/cli-bridge-mcp --> <div align="center">

<img src="https://raw.githubusercontent.com/JoaoBerne/cli-bridge-mcp/main/assets/banner.gif" width="860" alt="cli-bridge — your assistant borrows the powers of every AI CLI you already have: huge-context reads, vision, parallel builds, cross-vendor checks">

English · Français · 简体中文 · Español · Português (BR) · 日本語 · Deutsch

</div>

cli-bridge

!CI !PyPI !stars !python !license !MCP !ban--safe

Your assistant is only as good as the one model you opened. cli-bridge is a Model Context Protocol server that lets it borrow the other AI CLIs you already run — a bigger context, vision, a free second opinion from a different vendor, or a delegated build that comes back as a reviewable diff.

No API keys · no token extraction · no Node · no daemon · stdlib + mcp only.

In one sentence

You're talking to one AI assistant. You've also installed and logged into others — Claude Code, Codex, Gemini, opencode, Ollama. cli-bridge connects them: when your assistant needs something it can't do alone, it asks one of the other CLIs and hands you the result.

The problem it solves

Whatever assistant you're using has hard limits. It can't read a 2M-token repo in one pass, can't see a screenshot, can't hand you a generated image, and can't check its own work without bias — but some other CLI on your machine can do each of those. cli-bridge is the bridge between them: it spawns the official CLI as a subprocess (exactly as you'd run it by hand — no keys, no token extraction) and returns the answer to your assistant.

The result: one assistant whose ceiling on every axis is the best tool in your toolbox, not the one you happened to open.

---

The 10-second demo

You're in Claude. Claude can't hand you an image. Codex can — natively with gpt-image-2, or by writing a script that renders one. For a precise layout like a social card, have it script and run it:

ask_build(lane="gpt", task="generate a 1200×630 social card to assets/card.png — write a script that renders it, then run it", zone="assets")
→ Codex writes assets/card.png · you get the path back, never a binary blob (artifact-return)

Your assistant just gained an ability it doesn't have. That's the whole idea — now scale it to giant-context reads, vision, parallel grunt-work, and independent cross-vendor verification.

_(Codex generates the image with gpt-image-2, a real text-to-image model built into the CLI — counted against your ChatGPT plan's usage, no separate API key (image generation needs a paid plan; it's not on the Free tier). It comes back as a path, not a blob, because binaries travel by artifact-return, not the text channel. A build lane can also render charts, diagrams or SVGs by writing code, when that's the better fit.)_

…and it delegates real work, safely

cli-bridge build <lane> "<task>" hands the job to another model running in a throwaway git worktree, then gives you back a diff — your repo is never touched until you apply it yourself.

<p align="center"> <img src="https://raw.githubusercontent.com/JoaoBerne/cli-bridge-mcp/main/assets/demo-borrow.gif" width="860" alt="cli-bridge build: opencode adds a function in a throwaway worktree and returns a reviewable diff; the real repo stays clean"> </p>

---

What you get — the four levers

cli-bridge isn't one feature, it's four abilities your assistant gains. Get these and every tool below slots into place:

1. Borrow — reach a capability your assistant lacks (vision, a 1M-token context window, a file a

coding agent generates, a model that's simply better at this).

1. Spread — when one subscription hits its limit, keep going on another lane you already pay for.
2. Offload — fan laborious, parallel grunt-work across cheap/free lanes while you build elsewhere.
3. Verify — have a different vendor family check the work, because a model can't catch its own

blind spots. This is the one thing a single-vendor tool structurally cannot do.

---

What this unlocks

Each block: one sentence of when you reach for it, the exact call, and what you get back.

Borrow abilities your assistant doesn't have

Every CLI has a different superpower, and each runs non-interactively — so cli-bridge can spawn it. Borrow the one your host lacks (it must be installed + logged in):

| Superpower | Which CLI has it | Borrow it when | |------------|------------------|----------------| | Images | Codex (gpt-image-2, no API key — paid ChatGPT plan, not Free) | your host can't draw | | Huge context | Gemini (1M-token window) | a file/repo won't fit your host's context | | Fresh knowledge | Gemini (Google-Search grounding) · Grok (live web/X) ⚗️ | beat a stale cutoff: "what's the current API of <lib>?" | | Vision | images=[…] on Codex · opencode · Ollama · Apple · Gemini ⚗️ | analyse a screenshot or diagram | | A free second opinion | opencode · Ollama · Apple (local, $0) · Gemini (scarce free tier) | a $0 cross-check | | Generated files | any build lane → artifact-return | get a chart / PDF / diagram back by path | | Video ⚗️ | Grok (Imagine) — if your installed CLI exposes it (Veo isn't exposed by any official Gemini CLI extension) | you need a generated clip |

ask_build(lane="gpt", task="generate a 1200×630 social card to assets/card.png", zone="assets")   # Codex image → file by path, no API key (paid ChatGPT plan)
ask_gemini(task="find the bug across ./src — read the files you need", cwd="path/to/repo")         # 1M-token context
ask_gemini(task="what's the current recommended API for <lib>? check the latest docs")            # fresh knowledge (Search grounding)
ask_apple(task="what's wrong in this UI?", images=["screenshot.png"])                             # vision, on-device, $0, offline

Vision, honestly: images=[…] is no longer one vendor's feature. Five lanes take it and each lane declares its own shape as data, so you always just pass paths — an argv flag (apple --image, gpt -i, opencode -f) or a path folded into the prompt (gemini @path, ollama bare); a sixth lane is one field, not a code path. Caveat: whether the image is actually read depends on the model behind the lane, not the lane — on the same test image one free model transcribed every line and another answered "model lacks vision". Full table in docs/TOOLS.md.

⚗️ = experimental / depends on the installed CLI's current build (e.g. Grok Build is beta) — verify with doctor --deep.

Never stop working when you hit a limit

When your main subscription caps out mid-task. ask_cascade falls through to another lane you already pay for, skipping any lane that's cooled down after a quota/auth/timeout error.

ask_cascade(task="finish wiring this endpoint")   # cheapest→strongest; a cooled-down lane is skipped
ask_best(task="…", mode="deep")                   # let the router pick the most suitable available lane

Offload the grunt work — in parallel, and cheap

When the work is laborious but not hard (refactors, migrations, test coverage). Fan it out, journaled so a server restart resumes instead of restarting; delegate a build and keep working.

batch_run(tasks=[...], dry_run=true)                       # cost envelope first — nothing is spawned
batch_run(tasks=[...], max_calls=20, max_credits=2.0)      # then run under a hard budget (resumable)
ask_build(lane="opencode", task="add the landing page", zone="frontend", mode="direct", async=true)   # delegate, keep building
job_tail(job_id="…")  ·  build_steer(job_id="…", instruction="use Tailwind, not inline CSS")

Break self-confirmation — the 2026 problem one vendor can't solve

When you need to trust a result. A model reviewing its own work (or a sibling's) just confirms its own blind spots. cli-bridge puts a different model family in the reviewer's seat.

workflow(preset="jury", task="is this migration safe?", author_lane="gpt")            # cross-family vote, fail-closed
workflow(preset="verify_repair", task="add retry with backoff",
         builder_lane="gpt", verifier_lane="gemini")                                   # A builds, B reviews, loop to green
security_review(base="origin/main")   ·   review_diff(base="origin/main")              # OWASP, severity-ranked

Get a real second opinion

When you've reached a conclusion and want it pressure-tested, or several models side by side.

challenge(task="I'm dropping the cache layer — here's why: …")                         # one skeptic attacks it
consensus(task="which migration strategy is safest here?")                             # N answer, peer-rank the best
workflow(preset="fanout_compare", task="fix this failing test", lanes=["gpt","gemini","opencode"])

---

The toolbox

~30 tools, grouped by intent — the headline ones:

• Consult (read-only): ask_<lane> (one model), ask_all (every free lane in parallel + a disagreement score), ask_cascade (resilient fall-through), ask_best (router), consensus, challenge.
• Build (opt-in write): ask_build — mode=isolated→diff · mode=direct→zone-guarded · async→steerable, behind an executable Definition-of-Done gate.
• Review & verify: review_diff, security_review (findings tagged by severity and category — security / correctness / scope / ambiguity / performance / ops), debate, and workflow(preset=…) — converge (governance loop: an independent arbiter commits a blind verdict, anonymized cross-vendor peers review, every issue is adjudicated with a reason, then revise-or-converge), jury (cross-family k-of-N vote, fail-closed), verify_repair, fanout_compare, …
• Orchestrate & operate: batch_run (durable, budget-capped fan-out), plus usage_report, rate_lane, lane_stats, set_lane_cost, doctor.

Full reference — every tool, every flag: docs/TOOLS.md (or cli-bridge --help). Run CLI_BRIDGE_LEAN=1 for a curated ~12-tool surface.

There's also a human CLI — cli-bridge doctor|ask|build|review-diff|eval|… — the same engine from your terminal or CI (--json everywhere); cli-bridge build <lane> "<task>" returns a reviewable diff without touching your repo (--apply to land it as unstaged changes).

---

What you actually get when you combine them

One assistant whose ceiling on every axis is the ecosystem's best — not the tool you opened this morning: code with the strongest model, read ~1M tokens when yours is too short, answer with fresh knowledge past a stale cutoff, generate images/video, see screenshots, and fall back to a free/local lane when you're capped — spread across the subscriptions you already pay for.

The emergent property no single CLI has: true cross-vendor control — a different vendor in the reviewer's seat. Same-family subagents (Claude Code's, Grok's) can only self-confirm.

The honest seam: this unites capabilities, not mind — stateless spawns (no shared memory), spawn latency/cost, uneven quality, and the host always drives. It's orchestration, not fusion: you conduct specialists, you don't get one brain with every power.

→ Per-CLI strengths & limits (dated, churns fast): docs/COMPARISON.md.

Why cli-bridge (and not another "call other models" MCP)

• 🛡️ Ban-safe by design. It spawns each model's official CLI, exactly as you'd run it by hand —

no OAuth-token extraction, no API-key reuse. Each CLI handles its own auth and billing.

• 💸 Cost-safe defaults you tune to your plan. Out of the box ask_all / ask_cascade build a

free council and never touch paid quota unless you ask. Each lane ships a tier sourced from the vendor's published plans (dated in docs/COSTS.md, never detected from your account); override per lane with CLI_BRIDGE_<LANE>_COST=free|limited|paid. Two caps are enforced at spawn — CLI_BRIDGE_<LANE>_DAILY_LIMIT (runs/day, any lane) and CLI_BRIDGE_DAILY_CREDIT_CAP — full model in docs/BUDGET.md.

• 🔌 Works from any host. Claude Code, Codex, opencode, Cursor, VS Code (Cline/Continue), Zed —

anything that speaks MCP over stdio. The host's own lane is kept out of fan-out; hide it with CLI_BRIDGE_HIDE_HOST=1. Even a local model can be the host — see examples/local-first-host.md.

• 🧭 The cross-vendor edge is the moat. Independent verification means a different vendor in the

reviewer's seat — the scarce thing as AI writes a larger share of code, and exactly what a single-vendor tool can't offer.

Side by side

How the multi-model orchestrators differ on the axes that bite later — auth model, spend control, and what happens to your repo. (As of June 2026, read from each project's public repo/docs — corrections welcome.)

| | claude-octopus | PAL / zen-mcp | deliberation | cli-bridge | |---|---|---|---|---| | How other models are reached | hybrid: CLI spawn, OAuth-subscription reuse, or API keys | API keys (providers) + CLI spawn (clink) | CLI spawn (Codex/Gemini) + API keys (Grok, OpenRouter) | official CLI subprocess by default — each CLI keeps its own auth; optional opt-in API lanes | | API keys needed | optional fallback | for most providers | for Grok & OpenRouter | never by default — opt-in API lanes stay hidden until you set their key | | Spend control | session-only cost gate (OCTOPUS_MAX_COST_USD; no cross-session history) | none found | none found | enforced: per-lane daily run limit + daily credit cap + per-invocation budget, persisted (docs/BUDGET.md) | | Consensus governance | — | — | blind verdict + mandatory reason for each dismissal | converge: blind arbiter verdict, reasoned adjudication, no-self-approval — over anonymized cross-vendor peers, all 3 guards enforced in code | | Delegated edits | in-place | in-place (bypass/yolo flags) | in-place (workspace-write experts) | throwaway worktree → diff (your repo untouched), or zone-guarded direct mode (+ opt-in read-only-write guard) | | Survives host restart / /compact | session-scoped state | in-memory threads (TTL) | opt-in on-disk; in-memory by default | sqlite: conversations, jobs, fan-out journal | | Runtime deps | Node 18+, npm, bash | Python + pip packages | Node 18+, npm | Python stdlib + mcp | | Hosts | Claude Code-first (plugin; MCP server secondary) | any MCP host | any MCP host (+ Claude Code plugin) | any MCP host (+ a Claude Code plugin) |

Where they're stronger, honestly: claude-octopus ships a much larger workflow surface (49 commands, 32 personas, CI reactions); PAL has the biggest community (~11.6k★) with a polished tool set; and deliberation is a focused, mature single-purpose governance tool that pioneered the blind-verdict / mandatory-dismissal-reason model — cli-bridge has since adopted exactly those guards as workflow preset=converge (and runs them over anonymized cross-vendor peers under enforced budgets), but deliberation remains the more specialized product for that one job. cli-bridge's bet is broader: ban-safe auth, enforced budgets, cross-vendor verification, and delegation that can't wreck your repo — verified by its own shipped eval instead of claimed.

---

How it works

host (Claude/Codex/…) ──MCP──> cli-bridge ──spawn──> official CLI ──> model
                                    │
       keeps the host's own lane out of fan-out · only shows installed, enabled CLIs
       kills the whole process tree on timeout/cancellation · redacts secrets
       classifies errors (auth/limit/failed) · spills huge output to a file

No network calls of its own. No keys stored. It runs the same binaries you already trust, in your working directory, and hands the answer back.

<div align="center">

<img src="https://raw.githubusercontent.com/JoaoBerne/cli-bridge-mcp/main/assets/demo.gif" width="860" alt="cli-bridge security-review demo: a committed auth bypass is caught by a cross-vendor council, merged into one severity-ranked report, $0 on free lanes">

_Real run, real-time: the Verify lever — security-review fans OWASP roles across several models in parallel (claude/gpt/opencode/ollama here); they flag a committed auth bypass blocker, and usage shows the receipts._

</div>

<details> <summary><b>The report it actually prints</b> — real run, $0 free lanes, reproducible in 2 commands</summary>

<br>

Two different vendors independently flag the same bypass — that's the point: uncorrelated blind spots, not one model's opinion. Note what it does not hide: a reviewer came back empty, so the recap says 3/4 answered and the residual-risk line names the category nobody assessed.

# Security review (OWASP-aware)

_Base: `HEAD~1` · reviewers: injection (Mistral (Vibe CLI)), auth & access control (OpenCode
(gateway to many models)), data exposure & SSRF (OpenCode (gateway to many models)) · read-only_

## Reviewers — 3/4 answered

- ✅ **injection (Mistral (Vibe CLI))** _10927ms_ — 1 finding(s)
- ✅ **auth & access control (OpenCode (gateway to many models))** _13875ms_ — 1 finding(s)
- ❌ **secrets & crypto (Mistral (Vibe CLI))** _8465ms_ — empty
- ✅ **data exposure & SSRF (OpenCode (gateway to many models))** _11792ms_ — 0 finding(s)

**2 findings** (1 blocker, 1 high) — _block — blocker-level issues must be fixed before merge_

_By type: 2 security_

## Blocker

- **Unauthenticated user granted admin access** `auth.py:3` — _single_ · OpenCode · _security_
  require_admin returns True when user is None, meaning any unauthenticated/anonymous caller is
  treated as admin. This bypasses the is_admin check entirely.
  **Fix:** If user is None: raise PermissionError('authentication required')

## High

- **Insecure default for anonymous users** `auth.py:3` — _single_ · Mistral (Vibe CLI) · _security_
  The function returns True for user=None, granting admin privileges to unauthenticated sessions.
  This allows anonymous users to bypass admin checks.
  **Fix:** Remove the `if user is None: return True` block and let the function raise
  PermissionError for anonymous users.

## Residual risk

Treat with care — reviewer role(s) failed (secrets & crypto=empty); their categories are
unassessed; this is a static review of the shown diff only — no runtime, dependency, or
deployment/secrets-config analysis was performed.

Reproduce it on the throwaway fixture (a committed if user is None: return True, built in /tmp):

sh docs/demo/setup.sh                 # creates /tmp/demo-authz
cd /tmp/demo-authz && cli-bridge security-review --base HEAD~1

_(Trace footer omitted — CLI_BRIDGE_TRACE_FOOTER=off. Lanes vary with what you have logged in.)_

</details>

---

Writing code safely: two modes

Writes are contained, two ways — you pick review-gated or hands-off:

• isolated (default). Edits in a throwaway git worktree and hands back a diff. Your working

tree is never touched.

• direct. Writes real files, but only inside a zone you declare, behind a per-zone lock

with a post-turn zone-violation check. You in backend/, a delegate in frontend/, concurrently — neither can scribble across your whole repo; undo is zone-scoped, never a global reset.

And a tripwire for the read-only path: set CLI_BRIDGE_VERIFY_PLAN_READONLY=1 and any plan (read-only) delegate that nonetheless writes to a git workspace gets a ⚠️ WORKSPACE MUTATION DETECTED flag on its answer (surfaced, never auto-reverted — you decide).

Delegate re-entry is depth-capped (CLI_BRIDGE_MAX_DEPTH, default 1) so a misconfigured delegate can't fork-bomb the council.

Where a delegate runs. Pass cwd and that wins. Leave it out and cli-bridge asks the host for its MCP workspace root — because hosts launch a user-scoped MCP server from your home directory, so the directory it inherits is an accident, and it decides both where an unscoped build delegate writes and where the delegate's own session file gets filed. Override with CLI_BRIDGE_DEFAULT_CWD if your host declares no roots.

---

Installation (≈5 min)

Prerequisites

• Python 3.10+ and uv (uvx ships with it):

curl -LsSf https://astral.sh/uv/install.sh | sh (macOS/Linux) · winget install astral-sh.uv (Windows).

• At least one AI CLI installed and logged in — that's what cli-bridge borrows. Have any of:

Claude Code, Codex, Gemini CLI, opencode, Ollama (local, $0). You consult only the lanes you already have.

1. Check what cli-bridge can see (no install — uvx fetches, runs, discards):

uvx --from cli-bridge-mcp cli-bridge doctor

doctor lists which CLIs are detected, their resolved paths, and cost tiers. doctor --deep validates each lane against its own --help.

2. Add it to your MCP host. cli-bridge is an MCP server — it runs inside your assistant, not by hand. Point the host at the same command:

• Claude Code — as a plugin (one-time, adds /cli-bridge:council, /cli-bridge:review,

/cli-bridge:security, /cli-bridge:build, /cli-bridge:setup and wires the MCP server): ``bash claude plugin marketplace add JoaoBerne/cli-bridge-mcp claude plugin install cli-bridge@cli-bridge-mcp ``

• Claude Code — MCP only (no slash commands):

  claude mcp add cli-bridge -- uvx cli-bridge-mcp

• Desktop apps — Claude Desktop, Hermes Desktop, … cli-bridge is a plain stdio MCP

server, so any desktop MCP client runs it:

• Claude Desktop: Settings → Developer → Edit Config (claude_desktop_config.json),

add the mcpServers block below, restart the app.

• Hermes Desktop (Nous Research): Settings → MCP servers → Add → command uvx,

args cli-bridge-mcp.

• GUI apps launch servers with a minimal PATH — cli-bridge compensates by also

searching the usual install dirs (/opt/homebrew/bin, /usr/local/bin, ~/.local/bin, ~/.npm-global/bin, …) for your CLIs. If a lane still shows "NOT on PATH" in doctor, point it directly: CLI_BRIDGE_<LANE>_BIN=/full/path/to/cli in the server's env.

• Any other host (Codex, Cursor, VS Code, Zed, …) — add to its MCP config

(~/.claude.json, .mcp.json, or the host's equivalent): ``json { "mcpServers": { "cli-bridge": { "command": "uvx", "args": ["cli-bridge-mcp"] } } } ` Full example with env vars: examples/mcp.example.json. Per-host config paths (Cursor, VS Code, Cline, Windsurf, Continue, Zed, Visual Studio, Neovim, Xcode) and how to make your agent consult it on its own: docs/HOSTS.md`.

Note: cli-bridge-mcp (the MCP server) is the host entry point; cli-bridge (no -mcp) is the human terminal CLI you ran for doctor in step 1.

3. Use it. Restart/reload your host and ask it to consult a lane — e.g. "use cli-bridge to get a second opinion from gpt" or "ask gemini to read ./src and find the bug".

Lanes

Built-in: Claude Code, Codex, Gemini (+ Antigravity agy), Mistral (Vibe), opencode, Ollama (local models, $0, offline), Apple Foundation Models (fm, on-device, $0, offline, unmetered), Qwen Code, Copilot, Cursor (cursor-agent), Grok, and two opt-in lanes that stay hidden until you set their env var (so the ban-safe default surface is unchanged): OpenRouter (400+ models, needs OPENROUTER_API_KEY) and Apple PCC (APPLE_FM_SERVE_URL).

Apple PCC talks to a local fm serve over HTTP. fm refuses PCC to anything cli-bridge spawns — it walks the caller's ancestry to the session boundary and checks that ancestor's code signature — so start the server through Terminal instead, which you can do from anywhere, including from your assistant: osascript -e 'tell application "Terminal" to do script "fm serve --port 1976"'. Terminal becomes the parent, which is the chain fm trusts. Same HTTP pattern serves llama.cpp / vLLM / LM Studio — the bundled cli-bridge-openai bridge now works keyless. See examples/apple-fm-serve.lane.json.

Local runtimes beyond Ollama — LM Studio · MLX · llama.cpp — ship as zero-code recipes: point CLI_BRIDGE_LANES_FILE at examples/lmstudio.lane.json, mlx.lane.json, or llamacpp.lane.json. (Several local runtimes of the same open weights give correlated answers — real council diversity comes from distinct vendors, not a second local runtime.)

Community lanes (examples/community-lanes.json, experimental + limited until you declare their cost): Aider, Goose, Plandex, Amp, Crush, Amazon Q Developer CLI, Droid.

Anything else is ~3 lines of JSON. Add a custom lane, or wrap any OpenAI-compatible endpoint two ways: spawn curl (key kept inside curl, never in argv), or use the bundled cli-bridge-openai stdlib bridge — set availability_env so the lane stays hidden until its key is exported. See examples/openai-compatible.lane.json and examples/ for recipes.

---

The honest part

"More models = better" is fragile — big models share training data, so their errors correlate. We measured our own central claim (cli-bridge eval, no LLM judge): a diverse council did not catch more bugs than one strong model — it cut the false alarms ~2×. Same catch rate, far less noise — which is exactly what keeps a reviewer trustworthy instead of muted. Precision is the product, not recall. The harness ships, so you can confirm it on your CLIs — numbers either way in docs/BENCHMARKS.md.

---

Known limitations

• Ban-safe = no token/key extraction, not a blanket guarantee — non-interactive use of a

provider's CLI isn't formally sanctioned everywhere and can change. Use your own accounts within their terms.

• Async jobs are in-process — a server restart marks running jobs interrupted. batch_run /

workflow are the exception: they journal each task and resume via resume_id.

• The injection guard is heuristic — it catches high-signal patterns, not everything; treat

delegate output as data, not instructions.

• Token/credit figures are estimates (chars/4 + your CREDITS_PER_1K), never exact.
• Cost tiers are sourced defaults, not detection — vendor-plan facts are dated; doctor warns

when the snapshot is stale.

• Experimental (qwen, copilot, grok, community lanes, images=): flags aren't

verified live — doctor --deep checks them against each CLI's --help on your machine.

---

Roadmap

See CHANGELOG.md for shipped history. Currently exploring (not shipped): an independent-oracle verify mode (a cross-family lane writes tests from the spec, blind to the implementation, so the test catches the bug instead of mirroring it) and tighter limit-aware failover. Big inter-agent "bus" ideas (recursive spawn, shared state, wire protocol) are positioned honestly as a direction, never sold as a shipped protocol — see docs/ARCHITECTURE.md.

---

References

The design choices above aren't vibes — each maps to a finding in the literature. Every entry was checked against its source (authors + venue), because a tool that sells "honest cross-vendor verification" should get its own citations right.

| Paper | ID | What it backs here | |-------|----|--------------------| | Du et al. — Improving Factuality and Reasoning via Multiagent Debate | 2305.14325 | debate: models critiquing each other beat one model alone | | ReConcile — Round-Table Conference Improves Reasoning | 2309.13007 | debate convergence + confidence-weighted consensus | | Mixture-of-Agents | 2406.04692 | layered aggregation across diverse models (and its limits) | | Chain-of-Agents | 2406.02818 | role-specialized multi-agent pipelines | | CriticGPT — LLM Critics Help Catch LLM Bugs | 2407.00215 | review_diff / security_review: an LLM critic catches bugs humans miss | | Perez et al. — Discovering Language Model Behaviors (sycophancy) | 2212.09251 | why a same-family judge is weak → cross-vendor jury + peer anonymization | | Wynn, Satija & Hadfield — Talk Isn't Always Cheap | 2509.05396 | debate failure modes → fail-closed verdicts, bounded rounds | | CONSENSAGENT — Consensus via Sycophancy Mitigation (Findings of ACL 2025) | ACL 2025 | sycophancy in consensus → "earn their seat" / anonymized peers | | Maryanskyy — When Agents Disagree: The Selection Bottleneck | 2603.20324 | consensus: selection > synthesis (the deterministic peer-vote default) |

A citation hygiene note. Talk Isn't Always Cheap (2509.05396) is Wynn, Satija & Hadfield — a popular council framework miscites it as "Xiong et al." We double-check attributions before repeating them, and flag it because honesty is the whole pitch.

Development

uv venv && uv pip install -e . pytest pytest-asyncio
pytest -q          # unit + integration (cross-host) tests; no real CLI or network needed

License

Apache 2.0

---

<div align="center">

<img src="https://raw.githubusercontent.com/JoaoBerne/cli-bridge-mcp/main/assets/mark.gif" width="84" alt="cli-bridge">

_{one side · bridged to a council}

</div>